https://ai.satyajit.si/ 2026-11-13 daily 1 https://ai.satyajit.si/about weekly 0.6 https://ai.satyajit.si/resume weekly 0.6 https://ai.satyajit.si/projects 2026-06-09 weekly 0.6 https://ai.satyajit.si/blog 2026-07-20 daily 0.8 https://ai.satyajit.si/articles 2026-11-13 daily 0.8 https://ai.satyajit.si/articles/scoring weekly 0.6 https://ai.satyajit.si/models weekly 0.6 https://ai.satyajit.si/architectures 2026-11-12 weekly 0.6 https://ai.satyajit.si/math weekly 0.6 https://ai.satyajit.si/logs 2026-06-03 weekly 0.6 https://ai.satyajit.si/arxiv 2026-07-22 daily 0.6 https://ai.satyajit.si/publications weekly 0.6 https://ai.satyajit.si/patents weekly 0.6 https://ai.satyajit.si/github weekly 0.6 https://ai.satyajit.si/health weekly 0.6 https://ai.satyajit.si/now weekly 0.6 https://ai.satyajit.si/uses weekly 0.6 https://ai.satyajit.si/reading weekly 0.6 https://ai.satyajit.si/snippets weekly 0.6 https://ai.satyajit.si/notes 2026-05-28 weekly 0.6 https://ai.satyajit.si/colophon weekly 0.6 https://ai.satyajit.si/lab/neuron-lab weekly 0.6 https://ai.satyajit.si/changelog weekly 0.6 https://ai.satyajit.si/developers weekly 0.6 https://ai.satyajit.si/developers/versioning weekly 0.6 https://ai.satyajit.si/contact weekly 0.6 https://ai.satyajit.si/privacy weekly 0.6 https://ai.satyajit.si/blog/site-agent-dynamic-tools 2026-07-20 https://ai.satyajit.si/blog/site-search-contextual-bm25 2026-07-20 https://ai.satyajit.si/blog/contextual-retrieval 2026-07-17 https://ai.satyajit.si/blog/qwen-bim-domain-beats-scale 2026-06-09 https://ai.satyajit.si/blog/hello-world 2026-06-03 https://ai.satyajit.si/articles/baphomet-1-1-4b https://media.thesatyajit.com/ai/thumbs/baphomet-1-1-4b.jpg?v=945c78be1efe3e45 Baphomet 1.1 4B: eight Qwen3.5 fine-tunes, eight numbers searched, and a gain its author calls noise https://media.thesatyajit.com/ai/films/baphomet-1-1-4b-poster.webp?v=5df068b9d998517d CloudGoat's Baphomet-1.1-4B is a linear merge of eight community fine-tunes of Qwen3.5-4B, with one mixing weight per parent found by an EvoGM search and an… https://media.thesatyajit.com/ai/films/baphomet-1-1-4b.mp4?v=b9e7c562a24a4efc 134 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/beach-house-on-a-3090 https://media.thesatyajit.com/ai/thumbs/beach-house-on-a-3090.jpg?v=f621901fe010fae8 One beach house on an RTX 3090: what open image and video models fit in 24 GB on Ampere https://media.thesatyajit.com/ai/films/beach-house-on-a-3090-poster.webp?v=71b992daeb220ee3 I gave three open image models and one open video model the same brief, a two-storey beach house on a beach in five lights, and ran them on one RTX 3090… https://media.thesatyajit.com/ai/films/beach-house-on-a-3090.mp4?v=1a8439692790ccd1 244 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/beach-house-three-models https://media.thesatyajit.com/ai/thumbs/beach-house-three-models.jpg?v=e9312550effb850d One brief, three Claude models: three beach houses you can walk through https://media.thesatyajit.com/ai/films/beach-house-three-models-poster.webp?v=98c4100fb8a1a1da Opus 5.5, Fable 5.2 and Sonnet 5.5 each got the same design brief and the same loop: build a beach house in Python, bake its light in Blender, screenshot it… https://media.thesatyajit.com/ai/films/beach-house-three-models.mp4?v=5f94faf178aa11c5 181 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/cloudflare-kitesurf https://media.thesatyajit.com/ai/thumbs/cloudflare-kitesurf.jpg?v=42d62c80ec64a760 Kitesurf: a browser made of Workers isolates, cheap to run and not yet cheap to host https://media.thesatyajit.com/ai/films/cloudflare-kitesurf-poster.webp?v=602423da90da06ac Cloudflare open-sourced Kitesurf on 12 November under Apache-2.0: a browser engine for agents that parses, styles, lays out, runs page JavaScript and… https://media.thesatyajit.com/ai/films/cloudflare-kitesurf.mp4?v=5f5c1c3015f6ab1d 207 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/dwg-redraw-skill https://media.thesatyajit.com/ai/thumbs/dwg-redraw-skill.jpg?v=c33b5d2f89c7944e An image-to-DWG agent skill, read: the model writes the drawing as JSON, a ledger tags every dimension, and AutoCAD draws it over COM https://media.thesatyajit.com/ai/films/dwg-redraw-skill-poster.webp?v=20d58eda62d90864 pengxiaoan's autocad-dwg-redraw-skill turns sketches, scans and CAD screenshots into editable DWG. I read all 21 files at commit ed182ec. https://media.thesatyajit.com/ai/films/dwg-redraw-skill.mp4?v=d501fb9e58e87076 215 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/edge0-145b https://media.thesatyajit.com/ai/thumbs/edge0-145b.jpg?v=1b7937397ce522ea Edge0-145B: a Qwen MoE with its experts squeezed, and a 51B table left on the SSD https://media.thesatyajit.com/ai/films/edge0-145b-poster.webp?v=f90f9db937d2ee79 Edge0 released Edge0-145B and Edge0-120B as device-native 100B-class open-weight models: 29 GB of memory, 75+ tok/s on an RTX Spark and a Mac, and a… https://media.thesatyajit.com/ai/films/edge0-145b.mp4?v=17ac7dd516f48e01 207 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/full-attention-census https://media.thesatyajit.com/ai/thumbs/full-attention-census.jpg?v=23a501aca7f13360 Full attention in open LLMs: a 180-config recount of the chart that says it is gone https://media.thesatyajit.com/ai/films/full-attention-census-poster.webp?v=b8fcfc14f75a1015 A chart going round says the share of open LLMs with full attention in every layer fell from 59% in 2025 to 27% in early 2026 and to 0% since late August. https://media.thesatyajit.com/ai/films/full-attention-census.mp4?v=c927e26c2736aa6c 140 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/gem-4d https://media.thesatyajit.com/ai/thumbs/gem-4d.jpg?v=1e02721feb8d0535 GEM-4D: a video world model taught geometry it never has to output https://media.thesatyajit.com/ai/films/gem-4d-poster.webp?v=b0420cdc6ca6908e GEM-4D (Harvard, MIT and Princeton, ECCV 2026) trains a video world model for robot manipulation with a second, throwaway branch that has to predict a… https://media.thesatyajit.com/ai/films/gem-4d.mp4?v=855e09a65e19a9ec 308 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/haru-4b https://media.thesatyajit.com/ai/thumbs/haru-4b.jpg?v=1e72f45ad1a4ac43 Haru-4B: a voice model that listens and talks on one 133 ms clock https://media.thesatyajit.com/ai/films/haru-4b-poster.webp?v=04e36aa64ca62d6b Higurashi's Haru-4B is an Apache-2.0 full-duplex speech model: Qwen3.5-4B with three streams summed into one position every 133 ms, 64-dimensional… https://media.thesatyajit.com/ai/films/haru-4b.mp4?v=f2d6664304be8786 300 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/klm-1-6 https://media.thesatyajit.com/ai/thumbs/klm-1-6.jpg?v=bbe1f5e3f537cf18 KLM 1.6: a 99M hybrid base model, counted from its header, and the last-digit test behind its lead https://media.thesatyajit.com/ai/films/klm-1-6-poster.webp?v=647c0eb65e15ccde MinimaLabs' KLM 1.6 is a 99M-parameter English base model under Apache 2.0: 18 Gated DeltaNet-2 layers and 6 attention layers, a 4,100-token vocabulary that… https://media.thesatyajit.com/ai/films/klm-1-6.mp4?v=82395bddf844edd3 139 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/liquidcodec https://media.thesatyajit.com/ai/thumbs/liquidcodec.jpg?v=83ac8d6aa1bd9cec LiquidCodec: how 24,000 samples a second become 150 integers a language model can predict https://media.thesatyajit.com/ai/films/liquidcodec-poster.webp?v=c9714885a6efcb71 Liquid AI's LiquidCodec is the 74M-parameter speech codec behind LFM3-Audio-2B and LFM3-Omni-2B: an STFT front end, a causal LFM3 encoder, twelve… https://media.thesatyajit.com/ai/films/liquidcodec.mp4?v=78410cecde80a671 165 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/modern-web-guidance-static-embeddings https://media.thesatyajit.com/ai/thumbs/modern-web-guidance-static-embeddings.jpg?v=a3fd440b6218b7bc Modern Web Guidance: the seven-line loop that replaced a 23 MB transformer https://media.thesatyajit.com/ai/films/modern-web-guidance-static-embeddings-poster.webp?v=e9dc7e3d09036afe Chrome's Modern Web Guidance is an agent skill whose CLI finds the right web-platform guide for a coding agent's two-to-five-word search. https://media.thesatyajit.com/ai/films/modern-web-guidance-static-embeddings.mp4?v=0c4f0d97594aa73b 116 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/multiset-vps-robotcar https://media.thesatyajit.com/ai/thumbs/multiset-vps-robotcar.jpg?v=7c5eac966e009dfc MultiSet's VPS on RobotCar Seasons: 27 photos that land, and the recall nobody published https://media.thesatyajit.com/ai/films/multiset-vps-robotcar-poster.webp?v=f8e515e87080c2dc MultiSet ran its visual positioning system on Oxford RobotCar Seasons: a map from one overcast day in November 2014, query photos from nine other drives… https://media.thesatyajit.com/ai/films/multiset-vps-robotcar.mp4?v=16a1b55ed1aac217 294 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/painted-valley-patch https://media.thesatyajit.com/ai/thumbs/painted-valley-patch.jpg?v=ca21c8c09ccf59fc patch, Painted Valley and a coloured-pencil town: what an agent game platform ships, and two scenes built on it, taken apart https://media.thesatyajit.com/ai/films/painted-valley-patch-poster.webp?v=165f52197ca5778c patch launched on 9 November 2026 as a CLI, an MCP server, an asset library and a host for browser games and PC mods, built for coding agents. https://media.thesatyajit.com/ai/films/painted-valley-patch.mp4?v=c499f63092855a9b 264 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/shadcn-improve https://media.thesatyajit.com/ai/thumbs/shadcn-improve.jpg?v=bb910445790c608b shadcn's improve: an expensive model writes the plan, a cheap one types it, and the same expensive model grades it https://media.thesatyajit.com/ai/films/shadcn-improve-poster.webp?v=b6d9c8610a62ee5a improve is an MIT-licensed agent skill from shadcn: four markdown files that turn the strongest model you have into an advisor that audits a repository and… https://media.thesatyajit.com/ai/films/shadcn-improve.mp4?v=69ecc60ef4a4e309 225 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/vercel-ui-sdk https://media.thesatyajit.com/ai/thumbs/vercel-ui-sdk.jpg?v=991ff253efee2611 Vercel's UI SDK: the agentic part is three tools that skip the screenshot https://media.thesatyajit.com/ai/films/vercel-ui-sdk-poster.webp?v=c158ea4add503765 Vercel launched the UI SDK on 12 November with a video, and shadcn followed with a promise of agentic features to be shown in the coming days. https://media.thesatyajit.com/ai/films/vercel-ui-sdk.mp4?v=a3632088f8674d72 276 2026-11-13 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/3d-reconstruction-roundup-4 https://media.thesatyajit.com/ai/thumbs/3d-reconstruction-roundup-4.jpg?v=afb74a7d831949cc 3D reconstruction roundup 4: segments, affine fields and a few clicks https://media.thesatyajit.com/ai/films/3d-reconstruction-roundup-4-poster.webp?v=cda3e1214406ceac Five Gaussian-splatting and 3D reconstruction papers from early November 2026, each explained from the mechanism up and checked against its own tables, code… https://media.thesatyajit.com/ai/films/3d-reconstruction-roundup-4.mp4?v=86de12011d5d93ef 300 2026-11-12 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/agent-3d-bim-tools https://media.thesatyajit.com/ai/thumbs/agent-3d-bim-tools.jpg?v=83df30bb83ae296f 3D and BIM tools for agents: where each one sits between a 360 capture and an IFC file https://media.thesatyajit.com/ai/films/agent-3d-bim-tools-poster.webp?v=27e89366dc2c2e79 A roundup for people who do reality capture. I cloned and read MutinyBIM (Apache-2.0), Pascal Editor (MIT) and vork (source-available), plus glTF-Transform… https://media.thesatyajit.com/ai/films/agent-3d-bim-tools.mp4?v=01ee19104e73f552 287 2026-11-12 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/inkling-2-small https://media.thesatyajit.com/ai/thumbs/inkling-2-small.jpg?v=dab87a9b0ba4b85a 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/clover-35b-a3b https://media.thesatyajit.com/ai/thumbs/clover-35b-a3b.jpg?v=e915b727d2d9be8e 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/hesai-kosmo https://media.thesatyajit.com/ai/thumbs/hesai-kosmo.jpg?v=d23743f4c6ef7526 Hesai Kosmo: what a 1535 nm lidar adds to a Gaussian splat https://media.thesatyajit.com/ai/films/hesai-kosmo-poster.webp?v=ac6681102f45f151 Hesai, the lidar maker, launched Kosmo on 15 October: a 1.2 kg handheld with a 569,000 point-per-second, 1535 nm lidar, a 50-megapixel fisheye pair and an… https://media.thesatyajit.com/ai/films/hesai-kosmo.mp4?v=e631579a2a827902 106 2026-10-16 yes 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/3dgs-capture-pipelines https://media.thesatyajit.com/ai/thumbs/3dgs-capture-pipelines.jpg?v=90086597ac8f3fa5 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/pplx-decider-v1-1 https://media.thesatyajit.com/ai/thumbs/pplx-decider-v1-1.jpg?v=0b9a37fd3521fb98 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/tilelang https://media.thesatyajit.com/ai/thumbs/tilelang.jpg?v=32600a0b432d678d 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/csfm-flow-matching https://media.thesatyajit.com/ai/thumbs/csfm-flow-matching.jpg?v=97ae354bf56c948a 2026-11-13 monthly 0.7 https://ai.satyajit.si/articles/kvzip-kv-compression https://media.thesatyajit.com/ai/thumbs/kvzip-kv-compression.jpg?v=c5fc0a8df4e80c9a KVzip: compress the KV cache once, by scoring it with the model's own reread https://media.thesatyajit.com/ai/films/kvzip-kv-compression-poster.webp?v=4d5b35fc533a6e8f Most KV-cache eviction scores pairs against the current question, so a cache compressed for one question is wrong for the next. https://media.thesatyajit.com/ai/films/kvzip-kv-compression.mp4?v=10c41248d250922f 86 2026-10-07 yes 2026-11-13 monthly 0.8 https://ai.satyajit.si/articles/rf-detr https://media.thesatyajit.com/ai/thumbs/rf-detr.jpg?v=333f8133be7c3e26 RF-DETR: the first real-time detector past 60 AP, and every size is cut from a supernet https://media.thesatyajit.com/ai/films/rf-detr-poster.webp?v=1a4364faa8e43cb5 RF-DETR (Roboflow, arXiv 2511.09554, ICLR 2026) is a real-time DETR on a DINOv2 backbone whose sizes come from weight-sharing neural architecture search… https://media.thesatyajit.com/ai/films/rf-detr.mp4?v=fdf1755c304abb9e 111 2026-10-07 yes 2026-11-13 monthly 0.8 https://ai.satyajit.si/articles/spirula-studio https://media.thesatyajit.com/ai/thumbs/spirula-studio.jpg?v=e292f47b92ac0bd0 Spirula Studio deleted its dependencies instead of wrapping them https://media.thesatyajit.com/ai/films/spirula-studio-poster.webp?v=c40b55205bb7d0d8 A 3D Gaussian Splatting pipeline in one binary — camera alignment, masking, normal estimation, training — with no Python, no PyTorch and no COLMAP install. https://media.thesatyajit.com/ai/films/spirula-studio.mp4?v=1590464805471873 266 2026-11-12 yes 2026-11-13 monthly 0.8 https://ai.satyajit.si/articles/agent-substrate https://media.thesatyajit.com/ai/thumbs/agent-substrate.jpg?v=4b2476b61a3fbd5e Agent Substrate: the 10x density is your agents' idle time https://media.thesatyajit.com/ai/films/agent-substrate-poster.webp?v=20523226b7af1d3f Agent Substrate, the Kubernetes runtime that NVIDIA is promoting this week and that Google's engineers wrote, keeps a small pool of warm gVisor or microVM… https://media.thesatyajit.com/ai/films/agent-substrate.mp4?v=0ba9aab9467c2858 291 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/arc-1-0 https://media.thesatyajit.com/ai/thumbs/arc-1-0.jpg?v=fd2c1ec0b03a6a54 ARC-1.0: the router linked under a 1.7B agent clip, and why it cannot route to that model yet https://media.thesatyajit.com/ai/films/arc-1-0-poster.webp?v=641db833ad0db15a A viral clip of a 1.7B model driving a browser and writing HTML inside the Pi harness is LFM3-1.7B, not ARC-1.0; Rayline's CEO replied under it about ARC. https://media.thesatyajit.com/ai/films/arc-1-0.mp4?v=ecf0c32f3ebe3666 256 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/deja-view https://media.thesatyajit.com/ai/thumbs/deja-view.jpg?v=e2f4f2397981cf8c Déjà View: one transformer block, run sixteen times, reconstructs the scene https://media.thesatyajit.com/ai/films/deja-view-poster.webp?v=f9775153925d94a5 NVIDIA's Déjà View (NeurIPS 2026 oral, arXiv 2605.30215) recovers camera poses and dense depth from a pile of photos by applying one frame-and-global… https://media.thesatyajit.com/ai/films/deja-view.mp4?v=ee29b8d45a6dabaf 289 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/e-means https://media.thesatyajit.com/ai/thumbs/e-means.jpg?v=6b40b8e5a68f7ee5 e-means: k-means with a moving average, a cosine schedule and two fitted laws https://media.thesatyajit.com/ai/films/e-means-poster.webp?v=759e6e9cab71e1e6 Meta FAIR's e-means replaces Lloyd's full-batch k-means with mini-batch updates that keep an exponential moving average of each cluster's count and sum… https://media.thesatyajit.com/ai/films/e-means.mp4?v=05685caf8a3c70c3 224 2026-11-13 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/flide-decision-inference https://media.thesatyajit.com/ai/thumbs/flide-decision-inference.jpg?v=0370b99824afb625 Flide: forking a hybrid model's state so every question runs in one pass https://media.thesatyajit.com/ai/films/flide-decision-inference-poster.webp?v=58527c4bc2424846 Flide is an inference engine for decision models such as Jev. It lays a request's shared state and all its questions out as one flat token stream, masks… https://media.thesatyajit.com/ai/films/flide-decision-inference.mp4?v=dc25cc22a6fade48 296 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/iris-z https://media.thesatyajit.com/ai/thumbs/iris-z.jpg?v=98df089ce6e0038e Iris-Z: the depth map is just another image in the sequence https://media.thesatyajit.com/ai/films/iris-z-poster.webp?v=15429426ea213392 Speridlabs fine-tuned FLUX.2 so one set of weights generates an image with its depth, estimates depth from a photo, and paints a picture onto a depth map. https://media.thesatyajit.com/ai/films/iris-z.mp4?v=2e2909bbb5b5369f 288 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/leapquant https://media.thesatyajit.com/ai/thumbs/leapquant.jpg?v=47b19cd10c99bfe9 LeapQuant: an 8-bit recurrent state that stops rounding every token https://media.thesatyajit.com/ai/films/leapquant-poster.webp?v=e281e96dbf1a4001 Gated DeltaNet and Kimi Delta Attention layers in Qwen, Kimi and GLM models read and rewrite a 128 x 128 state per head on every decode step, and storing… https://media.thesatyajit.com/ai/films/leapquant.mp4?v=2504a400f6dc1e0d 302 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/llama-modchip https://media.thesatyajit.com/ai/thumbs/llama-modchip.jpg?v=19c3b4b773d7859d llama-modchip: a 284B MoE in 64 GB of RAM, and what actually makes it fit https://media.thesatyajit.com/ai/films/llama-modchip-poster.webp?v=e731c13fb438e4a7 llama-modchip is a patched llama.cpp that runs DeepSeek-V4-Flash (284B total, about 13B active) on a 64 GB CPU box and a 3-bit-per-weight Nemotron-3.5… https://media.thesatyajit.com/ai/films/llama-modchip.mp4?v=6455034ccbe498ee 259 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/meta-encoder https://media.thesatyajit.com/ai/thumbs/meta-encoder.jpg?v=a1836cd9b3003784 MetaEncoder: a 28B decoder that answers by cosine https://media.thesatyajit.com/ai/films/meta-encoder-poster.webp?v=8b4b3321d28b0dbc Meta's MetaEncoder v1 turns the Muse Glimmer 30B decoder into a single-tower embedding model for text, images and video, and then uses the same embeddings… https://media.thesatyajit.com/ai/films/meta-encoder.mp4?v=185e4f8ad500b55f 314 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/meteor-driving https://media.thesatyajit.com/ai/thumbs/meteor-driving.jpg?v=a633efd1f760c829 METEOR: an agent-built driving network that perceives a lot and plans a little https://media.thesatyajit.com/ai/films/meteor-driving-poster.webp?v=1612e3be2138bec9 TIER IV's METEOR turns eight cameras into a bird's-eye grid, twelve task heads and three candidate paths, with code its README says Claude wrote and labels… https://media.thesatyajit.com/ai/films/meteor-driving.mp4?v=bdefe1466d84f501 283 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/minimax-h3-image https://media.thesatyajit.com/ai/thumbs/minimax-h3-image.jpg?v=6b07e5df48516874 MiniMax-H3-Image: 16 megapixels is 16,384 tokens, and minutes on a 3080 https://media.thesatyajit.com/ai/films/minimax-h3-image-poster.webp?v=c94a9362c890c6a8 MiniMax open-weighted H3-Image, a 33.1B single-stream diffusion transformer that generates 4096×4096 images directly in six distilled steps and edits with… https://media.thesatyajit.com/ai/films/minimax-h3-image.mp4?v=adc7d3efa55eb821 247 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/mutiny-bim https://media.thesatyajit.com/ai/thumbs/mutiny-bim.jpg?v=7e508a8897c2bccd Mutiny BIM: I modelled a scan-sized room in it, exported the IFC 4.3, and raced two editors on one wall https://media.thesatyajit.com/ai/films/mutiny-bim-poster.webp?v=845d378debc119bd I ran Mutiny BIM (Apache-2.0, commit 94513ba) in a container: modelled a 6.24 by 4.82 m room, shaped like a scan pipeline's output, with walls, a slab, a… https://media.thesatyajit.com/ai/films/mutiny-bim.mp4?v=360b697e655e88fd 288 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/nex-n2-6 https://media.thesatyajit.com/ai/thumbs/nex-n2-6.jpg?v=97a46eca5d8daa94 Nex-N2.6: a 1.6% repetition rate, and a base model that was already at 6.0% https://media.thesatyajit.com/ai/films/nex-n2-6-poster.webp?v=729d74c810f5746d Nex-N2.6 is Shanghai Innovation Institute's agentic post-train of GLM-5.3-Flash: same 321B-parameter MoE, FP8 weights, text, image and video in. https://media.thesatyajit.com/ai/films/nex-n2-6.mp4?v=b4d9d0e00e915c72 242 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/open-robotics-mapping-roundup https://media.thesatyajit.com/ai/thumbs/open-robotics-mapping-roundup.jpg?v=bdcbe42e7fceeffd Open robotics mapping roundup: multi-floor SLAM, RTK in C++20, calibration checks and phone-mounted LiDAR https://media.thesatyajit.com/ai/films/open-robotics-mapping-roundup-poster.webp?v=e30334051e7fbab1 Five open mapping and positioning projects from one week in November 2026, read from their code. https://media.thesatyajit.com/ai/films/open-robotics-mapping-roundup.mp4?v=1c2885a0c0a3ec26 292 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/praxis-1 https://media.thesatyajit.com/ai/thumbs/praxis-1.jpg?v=b0797f1c6cb65295 Praxis-1: a robot policy that acts on a video it never finishes imagining https://media.thesatyajit.com/ai/films/praxis-1-poster.webp?v=6402073c7b3d2327 Runway open-weighted Praxis-1 under Apache 2.0: a 27.2B mixture-of-experts video model with a 314M action expert that reads the backbone while it imagines… https://media.thesatyajit.com/ai/films/praxis-1.mp4?v=7661ca1002a9a77a 299 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-27b-throughput-week https://media.thesatyajit.com/ai/thumbs/qwen3-8-27b-throughput-week.jpg?v=c709f72ed34b362a Qwen3.8-27B at 764, 1,035 and 1,623 tok/s: one GPU, many streams, and what each number divides by https://media.thesatyajit.com/ai/films/qwen3-8-27b-throughput-week-poster.webp?v=79de1d84a968f736 In one week three posts put Qwen3.8-27B on one consumer GPU at 700, 1,035 and 1,600 tokens a second. All three are real, and all three are batch sizes. https://media.thesatyajit.com/ai/films/qwen3-8-27b-throughput-week.mp4?v=8f3a3d6f9225b350 294 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/raon-1-nano https://media.thesatyajit.com/ai/thumbs/raon-1-nano.jpg?v=54d0e8a92fe374b2 Raon-1 nano: half of a 3.1B model is a lookup table, and a 99M model keeps one too https://media.thesatyajit.com/ai/films/raon-1-nano-poster.webp?v=748013a99f28188b KRAFTON's Raon-1 nano is a Korean and English chat and tool-use model for devices: 3.1B parameters, of which 1.6B are a per-layer embedding table that is… https://media.thesatyajit.com/ai/films/raon-1-nano.mp4?v=0f545887e3b4b113 231 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/sketchpad-exact-arithmetic https://media.thesatyajit.com/ai/thumbs/sketchpad-exact-arithmetic.jpg?v=f20554d647858beb Sketchpad: certified exact arithmetic from a 0.8B model that only sees three cells https://media.thesatyajit.com/ai/films/sketchpad-exact-arithmetic-poster.webp?v=d64da2f1ce9dc141 Sketchpad fine-tunes Qwen3.5-0.8B to read a three-cell view of a rendered sheet and emit one step of long arithmetic, while a fixed runtime moves the view. https://media.thesatyajit.com/ai/films/sketchpad-exact-arithmetic.mp4?v=9a751316461f730f 275 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/steerdistill https://media.thesatyajit.com/ai/thumbs/steerdistill.jpg?v=821356be56f09e5c SteerDistill: quantizing an agent by teaching it its own fixes https://media.thesatyajit.com/ai/films/steerdistill-poster.webp?v=34c25272015d64ee SteerDistill (Tsinghua AIR) recovers what 4-bit weights cost an LLM agent. During training a frozen BF16 teacher finds the first turn where the quantized… https://media.thesatyajit.com/ai/films/steerdistill.mp4?v=ccc78b20a557fe12 301 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/tokin-token-native-rollouts https://media.thesatyajit.com/ai/thumbs/tokin-token-native-rollouts.jpg?v=b2299becd5fc7304 tokin: training an agent on the tokens it actually sampled https://media.thesatyajit.com/ai/films/tokin-token-native-rollouts-poster.webp?v=ea2ebf0153c6d6d0 Agent harnesses speak chat messages; RL trains on token ids, and re-encoding the messages does not give back the ids the model sampled. https://media.thesatyajit.com/ai/films/tokin-token-native-rollouts.mp4?v=eb3324a632fcfad9 286 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/weight-edits-without-finetuning https://media.thesatyajit.com/ai/thumbs/weight-edits-without-finetuning.jpg?v=84b49c3812d4f083 Deletion Tuning and closed-form edits: changing an LLM without fine-tuning https://media.thesatyajit.com/ai/films/weight-edits-without-finetuning-poster.webp?v=9b728ab085cc287b Two weight edits went round the same week. Deletion Tuning (NeurIPS 2026) trains one logit per MLP gate weight and ships only a bitmask that zeros about… https://media.thesatyajit.com/ai/films/weight-edits-without-finetuning.mp4?v=24755e4081029853 301 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/gaussian-splat-lite https://media.thesatyajit.com/ai/thumbs/gaussian-splat-lite.jpg?v=952b0650a6a25e32 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/qwen-compiled-computers https://media.thesatyajit.com/ai/thumbs/qwen-compiled-computers.jpg?v=bd5c3c467c5cfda8 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/weight-space-programming https://media.thesatyajit.com/ai/thumbs/weight-space-programming.jpg?v=a9926e0b1ef34bdd 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/pytorch-2-15 https://media.thesatyajit.com/ai/thumbs/pytorch-2-15.jpg?v=de3cfc895676260d 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/kimi-k3-1-pro https://media.thesatyajit.com/ai/thumbs/kimi-k3-1-pro.jpg?v=0e5822dd1d96662b Kimi K3.1 Pro: more RL, a sparse-attention indexer, and what the −65% counts https://media.thesatyajit.com/ai/films/kimi-k3-1-pro-poster.webp?v=caec13d533b8e7a0 Moonshot sold Kimi K3.1 Pro as Kimi K3 with more post-training. The weights landed on 2 November, and I diffed both repos' configs and all 96 safetensors… https://media.thesatyajit.com/ai/films/kimi-k3-1-pro.mp4?v=8d7f329d7fc34ff2 291 2026-11-12 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/chicken-reasoning-cues https://media.thesatyajit.com/ai/thumbs/chicken-reasoning-cues.jpg?v=9593c0833e255077 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/mistral-large-4 https://media.thesatyajit.com/ai/thumbs/mistral-large-4.jpg?v=a00190316a8d5091 2026-11-12 monthly 0.7 https://ai.satyajit.si/articles/system-one-harnesses https://media.thesatyajit.com/ai/thumbs/system-one-harnesses.jpg?v=dffcd47873cc1d03 Five Jev harnesses, for a model that cannot write https://media.thesatyajit.com/ai/films/system-one-harnesses-poster.webp?v=3ecc52920d359331 Five repositories landed on my desk this week, four of them younger than a week, all of them wrapping a decision model, none of their authors having read… https://media.thesatyajit.com/ai/films/system-one-harnesses.mp4?v=24c8c836177b39c5 110 2026-10-07 yes 2026-11-12 monthly 0.8 https://ai.satyajit.si/articles/abot-recon https://media.thesatyajit.com/ai/thumbs/abot-recon.jpg?v=c5058d8876c8b864 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/astryx https://media.thesatyajit.com/ai/thumbs/astryx.jpg?v=64da76b6d4fd47be 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/kimi-k3-1-pro-pruned https://media.thesatyajit.com/ai/thumbs/kimi-k3-1-pro-pruned.jpg?v=a505f49353d87f9e 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/latent-sparse-attention https://media.thesatyajit.com/ai/thumbs/latent-sparse-attention.jpg?v=8c85c546afc044d1 Latent Sparse Attention: stop scoring every key, score the clusters https://media.thesatyajit.com/ai/films/latent-sparse-attention-poster.webp?v=5f53b741c3262cf6 At 10M tokens, the indexer that picks which keys a sparse-attention query reads takes 98% of the layer's prefill time. https://media.thesatyajit.com/ai/films/latent-sparse-attention.mp4?v=926b6e9866e855ed 297 2026-11-12 yes 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/library-of-layers https://media.thesatyajit.com/ai/thumbs/library-of-layers.jpg?v=f13178fd91a6ea7d 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/lm-head-loss-spikes https://media.thesatyajit.com/ai/thumbs/lm-head-loss-spikes.jpg?v=359cbbb62672424d 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/neuron-tuning https://media.thesatyajit.com/ai/thumbs/neuron-tuning.jpg?v=1b9d5616f03d1779 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/solv3r https://media.thesatyajit.com/ai/thumbs/solv3r.jpg?v=4245d84f77e34d7f 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/splatanything https://media.thesatyajit.com/ai/thumbs/splatanything.jpg?v=916e2b775f6bf067 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/vino-volume-transformers https://media.thesatyajit.com/ai/thumbs/vino-volume-transformers.jpg?v=8fd1af4ece6d5549 2026-11-10 monthly 0.7 https://ai.satyajit.si/articles/3d-reconstruction-roundup-3 https://media.thesatyajit.com/ai/thumbs/3d-reconstruction-roundup-3.jpg?v=daf1eb1940e32a63 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/alpie-seed https://media.thesatyajit.com/ai/thumbs/alpie-seed.jpg?v=41f3baa97a5b0cf7 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/arete-27b https://media.thesatyajit.com/ai/thumbs/arete-27b.jpg?v=f99b7dcba832fb21 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/captain-code-as-policy https://media.thesatyajit.com/ai/thumbs/captain-code-as-policy.jpg?v=e04f31064625e78b 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/claude-message-threads https://media.thesatyajit.com/ai/thumbs/claude-message-threads.jpg?v=44be2d0d5813ccd8 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/debim https://media.thesatyajit.com/ai/thumbs/debim.jpg?v=ab521763710226ce 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/flashstate https://media.thesatyajit.com/ai/thumbs/flashstate.jpg?v=bb7b8940ac1ad498 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/hy4 https://media.thesatyajit.com/ai/thumbs/hy4.jpg?v=9ca286a2d7fb2cb2 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/hyperkernel https://media.thesatyajit.com/ai/thumbs/hyperkernel.jpg?v=2a2e2e26f5b33982 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/metrictrail https://media.thesatyajit.com/ai/thumbs/metrictrail.jpg?v=190c1794bc563f97 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/muse-ember https://media.thesatyajit.com/ai/thumbs/muse-ember.jpg?v=5de6313916f9b0a6 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/ooosplat https://media.thesatyajit.com/ai/thumbs/ooosplat.jpg?v=17b79268ff5d0dea 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/open-3d-skills https://media.thesatyajit.com/ai/thumbs/open-3d-skills.jpg?v=53f049309e09d7ad 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/qwen-warp-llamampere https://media.thesatyajit.com/ai/thumbs/qwen-warp-llamampere.jpg?v=0a9f1d4b6ad2c18b 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/qwen35-4b-glb-3d-coder https://media.thesatyajit.com/ai/thumbs/qwen35-4b-glb-3d-coder.jpg?v=2b875a00a51d5f5d 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/rea-video https://media.thesatyajit.com/ai/thumbs/rea-video.jpg?v=3aca208083059238 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/ripple-65b https://media.thesatyajit.com/ai/thumbs/ripple-65b.jpg?v=aafe9c53fde299c9 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/solidcraft https://media.thesatyajit.com/ai/thumbs/solidcraft.jpg?v=1e18943315d3dcc6 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/splat-walk https://media.thesatyajit.com/ai/thumbs/splat-walk.jpg?v=2813df6185400a22 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/time-smeared-video-gaussians https://media.thesatyajit.com/ai/thumbs/time-smeared-video-gaussians.jpg?v=e22bab3aa527f456 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/visionsim https://media.thesatyajit.com/ai/thumbs/visionsim.jpg?v=b55a385969841861 2026-11-09 monthly 0.7 https://ai.satyajit.si/articles/anidoodle https://media.thesatyajit.com/ai/thumbs/anidoodle.jpg?v=576672f9630efe2c 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/construct-3d-objects https://media.thesatyajit.com/ai/thumbs/construct-3d-objects.jpg?v=48b9ebee91cdc4a1 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/dsh-superagent https://media.thesatyajit.com/ai/thumbs/dsh-superagent.jpg?v=ed03af8aa3235171 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/mistral-le-chonk https://media.thesatyajit.com/ai/thumbs/mistral-le-chonk.jpg?v=90e2d6db007371e5 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/qwen35-4b-3d-coder https://media.thesatyajit.com/ai/thumbs/qwen35-4b-3d-coder.jpg?v=269370a6b6ee535a 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/scenario-kinetic-film-skill https://media.thesatyajit.com/ai/thumbs/scenario-kinetic-film-skill.jpg?v=3f40f22710b8bc5a 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/stablegen-rigged-characters https://media.thesatyajit.com/ai/thumbs/stablegen-rigged-characters.jpg?v=3bc1821044a26ff6 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/steve-desktop-cad https://media.thesatyajit.com/ai/thumbs/steve-desktop-cad.jpg?v=84c4799005e04dd6 2026-11-01 monthly 0.7 https://ai.satyajit.si/articles/agent-desktop https://media.thesatyajit.com/ai/thumbs/agent-desktop.jpg?v=237da69214f1e318 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/bun-aot https://media.thesatyajit.com/ai/thumbs/bun-aot.jpg?v=ec31353f6795902c 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/calculus-of-interfaces https://media.thesatyajit.com/ai/thumbs/calculus-of-interfaces.jpg?v=37af7de652737e72 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/codex-harness-from-the-source https://media.thesatyajit.com/ai/thumbs/codex-harness-from-the-source.jpg?v=9d70069069f2584d 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/drex-max-v1 https://media.thesatyajit.com/ai/thumbs/drex-max-v1.jpg?v=138ebdb67da6adbf 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/fused-zipnerf https://media.thesatyajit.com/ai/thumbs/fused-zipnerf.jpg?v=7dee99d22fc0817d 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/gaussian-splat-to-blender-with-claude https://media.thesatyajit.com/ai/thumbs/gaussian-splat-to-blender-with-claude.jpg?v=36619632c9d4ebe4 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/gofish-fisheye-gaussian-splatting https://media.thesatyajit.com/ai/thumbs/gofish-fisheye-gaussian-splatting.jpg?v=7ff6491775eeec95 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/hexels https://media.thesatyajit.com/ai/thumbs/hexels.jpg?v=648ead56ae031cbb 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/idol-image-to-3d https://media.thesatyajit.com/ai/thumbs/idol-image-to-3d.jpg?v=7ea289bf598bbefa 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/leann https://media.thesatyajit.com/ai/thumbs/leann.jpg?v=85eaad5480f3fd94 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/lfm3-omni-webgpu https://media.thesatyajit.com/ai/thumbs/lfm3-omni-webgpu.jpg?v=888e9ac41c297ea6 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/neuron-lens https://media.thesatyajit.com/ai/thumbs/neuron-lens.jpg?v=68a59020790cf4bb 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/nvidia-metatron-v1 https://media.thesatyajit.com/ai/thumbs/nvidia-metatron-v1.jpg?v=5b1919ca829cb2e6 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/omniray https://media.thesatyajit.com/ai/thumbs/omniray.jpg?v=2d2c0462e0425356 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/sam3-onnx https://media.thesatyajit.com/ai/thumbs/sam3-onnx.jpg?v=8c6bce564f674706 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/scenario-3d-os https://media.thesatyajit.com/ai/thumbs/scenario-3d-os.jpg?v=c116d60e463f17b4 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/soft-mtp https://media.thesatyajit.com/ai/thumbs/soft-mtp.jpg?v=ff48244ae85e5d13 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/spark-x2-6-1-7b https://media.thesatyajit.com/ai/thumbs/spark-x2-6-1-7b.jpg?v=e84ce87a442643bb 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/switch-decision-harness https://media.thesatyajit.com/ai/thumbs/switch-decision-harness.jpg?v=2393e319b82db681 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/threejs-meshlets https://media.thesatyajit.com/ai/thumbs/threejs-meshlets.jpg?v=c488af47d6859e0a 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/web-font-agent-prompt https://media.thesatyajit.com/ai/thumbs/web-font-agent-prompt.jpg?v=ef3d59cbcf30b099 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/wevisone-vl-4b https://media.thesatyajit.com/ai/thumbs/wevisone-vl-4b.jpg?v=346796dc9b902edb 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/xiaomi-ocr https://media.thesatyajit.com/ai/thumbs/xiaomi-ocr.jpg?v=f3ea4055177adf96 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/lfm3 https://media.thesatyajit.com/ai/thumbs/lfm3.jpg?v=3493e119ed9c0c82 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/neuron-finding https://media.thesatyajit.com/ai/thumbs/neuron-finding.jpg?v=c4df1a63427f25b4 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/qwenputer https://media.thesatyajit.com/ai/thumbs/qwenputer.jpg?v=df3ef8c377adba88 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/layerfs https://media.thesatyajit.com/ai/thumbs/layerfs.jpg?v=03962cd1f98d6a5e LayerFS: 157 snapshots, a toy, and then the real store https://media.thesatyajit.com/ai/films/layerfs-poster.webp?v=8deeb258bcfe342e LayerFS gives each agent a full writable workspace over shared history without copying the base. https://media.thesatyajit.com/ai/films/layerfs.mp4?v=339e46cb4fc5cee0 77 2026-10-31 yes 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/unlimited-ocr-2 https://media.thesatyajit.com/ai/thumbs/unlimited-ocr-2.jpg?v=10b5d78c2d6de3dc 2026-10-31 monthly 0.7 https://ai.satyajit.si/articles/delta-matching https://media.thesatyajit.com/ai/thumbs/delta-matching.jpg?v=699f782dbbedf49a 2026-10-30 monthly 0.7 https://ai.satyajit.si/articles/nemotron-3-5-transcribe https://media.thesatyajit.com/ai/thumbs/nemotron-3-5-transcribe.jpg?v=da8a91a05f90f836 2026-10-30 monthly 0.7 https://ai.satyajit.si/articles/coarse-to-real-c2r https://media.thesatyajit.com/ai/thumbs/coarse-to-real-c2r.jpg?v=da80485e702a18df 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/cpu-chat-covert-channels https://media.thesatyajit.com/ai/thumbs/cpu-chat-covert-channels.jpg?v=c63140d6c810e292 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/cyberpunk-2077-in-a-browser-tab https://media.thesatyajit.com/ai/thumbs/cyberpunk-2077-in-a-browser-tab.jpg?v=60b23529fa2bdc3b 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/echo-3-world-model https://media.thesatyajit.com/ai/thumbs/echo-3-world-model.jpg?v=1647f73b46105649 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/gemma-argon-vllm https://media.thesatyajit.com/ai/thumbs/gemma-argon-vllm.jpg?v=78e9f048ce171ebb 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/hip3r https://media.thesatyajit.com/ai/thumbs/hip3r.jpg?v=82333b40b2f975d0 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/neuron-lab https://media.thesatyajit.com/ai/thumbs/neuron-lab.jpg?v=66eb15ec307dbc3e 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/nvidia-motionbricks https://media.thesatyajit.com/ai/thumbs/nvidia-motionbricks.jpg?v=8708afbdac7376c5 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/on-policy-self-training https://media.thesatyajit.com/ai/thumbs/on-policy-self-training.jpg?v=a27185a8592fa3c8 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/rivet-warpdrive https://media.thesatyajit.com/ai/thumbs/rivet-warpdrive.jpg?v=a571abf4300a6f91 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/road-rash-in-a-browser-tab https://media.thesatyajit.com/ai/thumbs/road-rash-in-a-browser-tab.jpg?v=a90a4a95a9a5c462 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/sae-text-detectors https://media.thesatyajit.com/ai/thumbs/sae-text-detectors.jpg?v=dd58b004c6028653 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/neuron-steering https://media.thesatyajit.com/ai/thumbs/neuron-steering.jpg?v=2258fd0ab2195275 Neuron steering: find a neuron, scale its write, measure the damage https://media.thesatyajit.com/ai/films/neuron-steering-poster.webp?v=b3d55615e3e0f36b A popular post shows single MLP neurons in Qwen3.5-4B being found and steered with a dose scaled by the residual norm. https://media.thesatyajit.com/ai/films/neuron-steering.mp4?v=cfc815a732b0bc1e 102 2026-10-26 yes 2026-10-27 monthly 0.7 https://ai.satyajit.si/articles/claude-code-sandbox-internals https://media.thesatyajit.com/ai/thumbs/claude-code-sandbox-internals.jpg?v=12df96af23032057 Claude Code's sandbox from the source: the bwrap argv, the seccomp program, and the proxy https://media.thesatyajit.com/ai/films/claude-code-sandbox-internals-poster.webp?v=db20b24f5ddb6739 A source-level teardown of @anthropic-ai/sandbox-runtime 0.0.81 (Apache-2.0), the sandbox behind Claude Code's Bash tool. https://media.thesatyajit.com/ai/films/claude-code-sandbox-internals.mp4?v=855118bb9a133a8e 108 2026-10-26 yes 2026-10-26 monthly 0.7 https://ai.satyajit.si/articles/lipogram-rl https://media.thesatyajit.com/ai/thumbs/lipogram-rl.jpg?v=6eed0d5c0fc68569 Qwen, minus the e: a mask removes the letter, training has to write the rest https://media.thesatyajit.com/ai/films/lipogram-rl-poster.webp?v=734df04acb6e902c Kevin Wu fine-tuned Qwen3-1.7B on a 16 GB GPU to answer without the letter e, and found that masking every token containing an e leaves a model fluent in… https://media.thesatyajit.com/ai/films/lipogram-rl.mp4?v=0ce4f584375de2fe 105 2026-10-26 yes 2026-10-26 monthly 0.7 https://ai.satyajit.si/articles/nanbeige-4-3-33b https://media.thesatyajit.com/ai/thumbs/nanbeige-4-3-33b.jpg?v=d9b7da1dc0b8cdc5 Nanbeige4.3-33B: a teacher first, and a cache you pay for https://media.thesatyajit.com/ai/films/nanbeige-4-3-33b-poster.webp?v=03ead49d513ec3c7 Nanbeige's 33B is a plain dense Llama-shaped model released as the teacher for a 3B that does not exist yet. https://media.thesatyajit.com/ai/films/nanbeige-4-3-33b.mp4?v=810b26b90361f9c5 109 2026-10-26 yes 2026-10-26 monthly 0.7 https://ai.satyajit.si/articles/qwen4-27b-vllm-xpu https://media.thesatyajit.com/ai/thumbs/qwen4-27b-vllm-xpu.jpg?v=6f9d54f29fa675ea Qwen4-27B: what a merged kernel PR does and does not leak https://media.thesatyajit.com/ai/films/qwen4-27b-vllm-xpu-poster.webp?v=2ace469e8d4863f8 A viral post says a merged Intel-side kernel PR reverse-engineers an unreleased dense Qwen4-27B. https://media.thesatyajit.com/ai/films/qwen4-27b-vllm-xpu.mp4?v=3ff0a995930adcdc 110 2026-10-26 yes 2026-10-26 monthly 0.7 https://ai.satyajit.si/articles/slotstream https://media.thesatyajit.com/ai/thumbs/slotstream.jpg?v=1fec68be48d54518 Slotstream: a 103.8 GB model on a 48 GB Mac, and what 21 tok/s is made of https://media.thesatyajit.com/ai/films/slotstream-poster.webp?v=8d8e043f1bd2e96c Slotstream streams Qwen3.8-Flash-Next from the SSD through a fixed pool of expert slots, in one Swift binary on MLX and Metal. https://media.thesatyajit.com/ai/films/slotstream.mp4?v=5303ff383ef61359 95 2026-10-26 yes 2026-10-26 monthly 0.7 https://ai.satyajit.si/articles/softmax-fix https://media.thesatyajit.com/ai/thumbs/softmax-fix.jpg?v=60f59a1a59e25314 A fix for softmax and cross-entropy: the target gradient that rounds to zero https://media.thesatyajit.com/ai/films/softmax-fix-poster.webp?v=97710343f943e8e5 Jon Barron, Dan Gnanapragasam and Michael Tschannen show that every standard softmax cross-entropy adds a tiny number to 1.0 and loses it: past a logit… https://media.thesatyajit.com/ai/films/softmax-fix.mp4?v=179dcc660a7571b6 111 2026-10-24 yes 2026-10-26 monthly 0.7 https://ai.satyajit.si/articles/3d-model-anything https://media.thesatyajit.com/ai/thumbs/3d-model-anything.jpg?v=48b91425ea670603 3D Model Anything: an iPhone scan, an agent, and a photo it keeps comparing against https://media.thesatyajit.com/ai/films/3d-model-anything-poster.webp?v=873084231c45e9a7 Wes Bos's 3D Model Anything records a room on a LiDAR iPhone (RoomPlan walls and boxes, the LiDAR mesh, posed photos and depth frames), then hands it to an… https://media.thesatyajit.com/ai/films/3d-model-anything.mp4?v=2185f28bf03660f5 110 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/dessa https://media.thesatyajit.com/ai/thumbs/dessa.jpg?v=54e5554a34344e7b Dessa: timing a speech-only conversation model from its own demo tracks https://media.thesatyajit.com/ai/films/dessa-poster.webp?v=9428208c793d98fa Block's Dessa is a full-duplex speech-to-speech model that, Block says, learned to talk from human conversation audio alone: no text, no pretrained models… https://media.thesatyajit.com/ai/films/dessa.mp4?v=54ccee273b59be4b 103 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/generative-trajectory-planning https://media.thesatyajit.com/ai/thumbs/generative-trajectory-planning.jpg?v=4cb7b216042e5452 Generative Trajectory Planning: a diffusion prior that proposes plans and learns from the ones search finds https://media.thesatyajit.com/ai/films/generative-trajectory-planning-poster.webp?v=f88977732e6572a8 GTP replaces the policy in online RL with a diffusion model over whole action sequences. https://media.thesatyajit.com/ai/films/generative-trajectory-planning.mp4?v=ed096eef32682eba 121 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/hy-nomas https://media.thesatyajit.com/ai/thumbs/hy-nomas.jpg?v=985b2655e8d0675a Hy-Nomas: an 8.6B world model whose memory is a KV cache, and what it takes to fit a 4090 https://media.thesatyajit.com/ai/films/hy-nomas-poster.webp?v=5f76290cc63cad64 Tencent Hunyuan's Hy-Nomas generates a world five seconds at a time from text and camera paths. https://media.thesatyajit.com/ai/films/hy-nomas.mp4?v=a89cebfd75ff929e 112 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/mojo-max-open https://media.thesatyajit.com/ai/thumbs/mojo-max-open.jpg?v=40ce3f062abf6487 Mojo and MAX source release: the whole stack is readable, and two licences cover it https://media.thesatyajit.com/ai/films/mojo-max-open-poster.webp?v=a6774b800b314b93 On 23 October Chris Lattner wrote that Mojo and MAX are now fully open. I cloned github.com/modular/modular and sorted 3.35 million lines of code by the… https://media.thesatyajit.com/ai/films/mojo-max-open.mp4?v=a6a1d34efccfdcf2 117 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/nemotron-3-5-bolt https://media.thesatyajit.com/ai/thumbs/nemotron-3-5-bolt.jpg?v=b838370f56f0df98 Nemotron 3.5 Bolt E4B: Gemma 4 E4B, retrained for talking robots, in a 4GB file that weighs 6.18 GB https://media.thesatyajit.com/ai/films/nemotron-3-5-bolt-poster.webp?v=fba028a028093693 NVIDIA's Nemotron 3.5 Bolt E4B 4GB is Gemma-4-E4B-it with reinforcement learning on top, aimed at voice assistants, game characters and robots like Reachy… https://media.thesatyajit.com/ai/films/nemotron-3-5-bolt.mp4?v=5e71e11e7f4c28fc 105 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/omarchy-egpu https://media.thesatyajit.com/ai/thumbs/omarchy-egpu.jpg?v=ca61eaa72e7212e1 An Intel Arc Pro B60 on an M1 Max over Thunderbolt: the weights were never the problem https://media.thesatyajit.com/ai/films/omarchy-egpu-poster.webp?v=a560afdc441c2284 joshuaswarren/omarchy-egpu runs Qwen3.8-27B Q4_0 with all 66 layers on a 24 GB Arc Pro B60 in a Thunderbolt enclosure, driven by a 16-inch M1 Max MacBook… https://media.thesatyajit.com/ai/films/omarchy-egpu.mp4?v=c64f3e2686b7c46b 111 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/sam3-segmentation-skill https://media.thesatyajit.com/ai/thumbs/sam3-segmentation-skill.jpg?v=74e38d5f7a2ccc7d sam3-segmentation: the win is the tiling, not the agent https://media.thesatyajit.com/ai/films/sam3-segmentation-skill-poster.webp?v=1ea65a468b657870 Aurélie Guisnet's sam3-segmentation is an agent skill that wraps Meta's SAM 3 in a crop plan, a raw-candidate archive, review sheets, recorded decisions and… https://media.thesatyajit.com/ai/films/sam3-segmentation-skill.mp4?v=fc806d75097d7758 107 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/sandbox-runtime https://media.thesatyajit.com/ai/thumbs/sandbox-runtime.jpg?v=8fa7b9732ae83eb4 sandbox-runtime: Claude Code's sandbox is a mount table, a missing network and a proxy https://media.thesatyajit.com/ai/films/sandbox-runtime-poster.webp?v=972f7c52dde1724e Anthropic's open-source sandbox-runtime (srt, Apache-2.0) is the sandbox behind Claude Code's Bash tool: Seatbelt profiles on macOS, bubblewrap on Linux… https://media.thesatyajit.com/ai/films/sandbox-runtime.mp4?v=c131c3e3b98ad6c0 108 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/transitions-dev https://media.thesatyajit.com/ai/thumbs/transitions-dev.jpg?v=4b806d3e405ff2df transitions.dev's Generative effect: an agent animation drawn from the element's own geometry https://media.thesatyajit.com/ai/films/transitions-dev-poster.webp?v=53965218f92b67b0 Jakub Antalik's transitions.dev now ships a Generative effect skill with ui, text and image variants: a dotted wireframe, traced glyphs or a halftone of the… https://media.thesatyajit.com/ai/films/transitions-dev.mp4?v=ad27a8a598a702cc 103 2026-10-24 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/security-week-oct-2026 https://media.thesatyajit.com/ai/thumbs/security-week-oct-2026.jpg?v=e13ec4d668594a92 The week of 14-23 October 2026: php2shell, a poisoned LiteLLM, a CVSS 10 in GitLab, and eMudhra's weak CA keys https://media.thesatyajit.com/ai/films/security-week-oct-2026-poster.webp?v=9ba0b4c0e381e50c A defender's account of the week. Eight eMudhra intermediate CA keys turned out to be P-384 scalars below 256, so anyone could recover them; browsers… https://media.thesatyajit.com/ai/films/security-week-oct-2026.mp4?v=65b6fff286beaecb 116 2026-10-23 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/liquid-d1-ultra https://media.thesatyajit.com/ai/thumbs/liquid-d1-ultra.jpg?v=5646f33a6b2df514 d1-ultra: the calibration claim, recomputed from the index's own bins https://media.thesatyajit.com/ai/films/liquid-d1-ultra-poster.webp?v=878465561e73784c Liquid AI's d1-ultra is a hosted decision model, API-only, with no parameter count, at $0.042 per million input tokens and no output tokens. https://media.thesatyajit.com/ai/films/liquid-d1-ultra.mp4?v=010db5787584df71 104 2026-10-16 yes 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/integer-mult-exponent-race https://media.thesatyajit.com/ai/thumbs/integer-mult-exponent-race.jpg?v=70e2976e364d0fe4 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/project-maya https://media.thesatyajit.com/ai/thumbs/project-maya.jpg?v=4d1bd0c6ca032e99 2026-10-24 monthly 0.7 https://ai.satyajit.si/articles/ai-playground-home-agent https://media.thesatyajit.com/ai/thumbs/ai-playground-home-agent.jpg?v=e22beab104c6d3fe AI Playground Home Agent: the compute stays home, the conversation doesn't https://media.thesatyajit.com/ai/films/ai-playground-home-agent-poster.webp?v=794f52c1d30a87f5 Intel's AI Playground 3.2 adds Home Agent: message your PC from Telegram, Slack or a LAN web page and it generates images, video and replies on its own Arc… https://media.thesatyajit.com/ai/films/ai-playground-home-agent.mp4?v=2ec0f1bbadd536f5 112 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/blenderworld https://media.thesatyajit.com/ai/thumbs/blenderworld.jpg?v=78e83e3577d4b52e BlenderWorld: 223 Blender briefs, graded by an agent that opens the .blend https://media.thesatyajit.com/ai/films/blenderworld-poster.webp?v=6154ce0c1b1ea029 Qwen's BlenderWorld gives a 3D agent a real-world brief in text, images or video, takes back an editable Blender project, and grades it with 2,651 binary… https://media.thesatyajit.com/ai/films/blenderworld.mp4?v=b33e00a98272c0ac 115 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/carbon-x https://media.thesatyajit.com/ai/thumbs/carbon-x.jpg?v=52cb1dbb7e962361 Carbon-X: teaching Qwen3-14B to read DNA in 6-base words https://media.thesatyajit.com/ai/films/carbon-x-poster.webp?v=3bc981918c7a0a13 Hugging Face's Carbon-X is Qwen3-14B mid-trained on 302B tokens of DNA and text, then fine-tuned and RL-trained to answer questions about raw sequences in… https://media.thesatyajit.com/ai/films/carbon-x.mp4?v=03e3909a137dc814 104 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/cua-spaces https://media.thesatyajit.com/ai/thumbs/cua-spaces.jpg?v=24e5c82a9e118841 Cua Spaces: one daemon per computer, and I ran three of them https://media.thesatyajit.com/ai/films/cua-spaces-poster.webp?v=f5f348478f9aad63 Cua Spaces gives Claude Code, Codex or any agent a desktop on a computer you own, and its new SDK fans jobs across all of them. https://media.thesatyajit.com/ai/films/cua-spaces.mp4?v=20e0c6c77ae6f086 104 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/d3-turbo https://media.thesatyajit.com/ai/thumbs/d3-turbo.jpg?v=4a725bc28fd0965d d3-turbo: the state stops looking at the question, so a photo is read once for 255 questions https://media.thesatyajit.com/ai/films/d3-turbo-poster.webp?v=ccffa2018355bc76 vLLM Semantic Router's d3-turbo is d3, the 27B Qwen3.8 decision model, trained further with one change to its attention: in the 16 full-attention layers the… https://media.thesatyajit.com/ai/films/d3-turbo.mp4?v=741047fa5d4e1419 115 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/deepinfer https://media.thesatyajit.com/ai/thumbs/deepinfer.jpg?v=10fbf636cb91c6df DeepInfer: DeepSeek's serving engine, and a throughput chart that multiplies by 3.25 https://media.thesatyajit.com/ai/films/deepinfer-poster.webp?v=2eddd786e070a966 DeepSeek open-sourced DeepInfer, the engine behind DeepSeek-V4.1-Flash, plus deepseek-recipe, its request encoder. I read the code and ran what runs on a CPU. https://media.thesatyajit.com/ai/films/deepinfer.mp4?v=ce6f237ba7e80582 116 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/eagle-4 https://media.thesatyajit.com/ai/thumbs/eagle-4.jpg?v=1de890c9194e9473 EAGLE-4: speculation that prices itself, and where the 5.21x comes from https://media.thesatyajit.com/ai/films/eagle-4-poster.webp?v=48d34868e5a464f9 EAGLE-4 keeps the one-layer drafter of EAGLE-3 and adds two things: a harness that stops drafting and trims verification when the next token is not worth… https://media.thesatyajit.com/ai/films/eagle-4.mp4?v=eb8eb97b1bd4503c 107 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/flint https://media.thesatyajit.com/ai/thumbs/flint.jpg?v=be906e7fda4a77f1 Flint: 125 tokens a second is what 2.5B active parameters buy at 273 GB/s https://media.thesatyajit.com/ai/films/flint-poster.webp?v=5ef0089450523415 Recursive's first model is a 39.8B-total, 2.5B-active MoE for local coding, released under Apache-2.0. https://media.thesatyajit.com/ai/films/flint.mp4?v=a6a9f811c5e3596e 104 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/fx0 https://media.thesatyajit.com/ai/thumbs/fx0.jpg?v=a558e8d71b4554ad fx0: the artifact builder is the demo, the durable agent loop is the product https://media.thesatyajit.com/ai/films/fx0-poster.webp?v=bf5c0f8d13fa5c83 Vercel open-sourced fx0, a 'mini v0' that writes one self-contained HTML file per prompt, to show off libfx, the Zig agent engine from its fx CLI packaged… https://media.thesatyajit.com/ai/films/fx0.mp4?v=0fa113efe282bc3b 108 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/gpu-lexer https://media.thesatyajit.com/ai/thumbs/gpu-lexer.jpg?v=7996526b331e09e7 gpu-lexer: how a 28 KB model lexes in parallel, and where the 3.8x came from https://media.thesatyajit.com/ai/films/gpu-lexer-poster.webp?v=6fcdcc238cb97d36 Shu Ding's gpu-lexer labels every part of a source file at once on WebGPU, with no grammar and no language id, and he just announced it 3.8x faster (850 ms… https://media.thesatyajit.com/ai/films/gpu-lexer.mp4?v=22e2e547ca1cc7fa 119 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/hatchzip https://media.thesatyajit.com/ai/thumbs/hatchzip.jpg?v=a53d1af3a15c3d93 hatchzip: unzipping into the space the archive gives back https://media.thesatyajit.com/ai/films/hatchzip-poster.webp?v=a17056fc9dda0587 hatchzip extracts a ZIP, 7z or tar from its last entry backwards and truncates the archive behind each batch, so the disk never has to hold the archive and… https://media.thesatyajit.com/ai/films/hatchzip.mp4?v=8c7186aa3cfd2f7e 99 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/ideogram-vector https://media.thesatyajit.com/ai/thumbs/ideogram-vector.jpg?v=ada5f144f03086a3 Ideogram Vector: 40 sample SVGs opened, and what they say about how it draws https://media.thesatyajit.com/ai/films/ideogram-vector-poster.webp?v=77423ee5e30dc416 Ideogram launched Vector 1.0, a closed text-to-SVG and image-to-SVG model, with no paper, no architecture and no benchmark. https://media.thesatyajit.com/ai/films/ideogram-vector.mp4?v=27a45f5362f39523 104 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/learnable-spectral-activations https://media.thesatyajit.com/ai/thumbs/learnable-spectral-activations.jpg?v=6eaa0e9633ae3364 Learnable Spectral Activations: stop picking frequencies, learn how to mix them https://media.thesatyajit.com/ai/films/learnable-spectral-activations-poster.webp?v=9b8e40ef28d68de7 Learnable Spectral Activations (Shor, Litany and Bronstein, arXiv 2610.07419, NeurIPS 2026) swap an implicit neural representation's fixed nonlinearity for… https://media.thesatyajit.com/ai/films/learnable-spectral-activations.mp4?v=069e3fbf29af5702 111 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/looking-ahead-video https://media.thesatyajit.com/ai/thumbs/looking-ahead-video.jpg?v=e2a79b7903c0273e Looking Ahead: teaching a chunk-by-chunk video model to predict two seconds out https://media.thesatyajit.com/ai/films/looking-ahead-video-poster.webp?v=cf0c5eefbaebc1eb Looking Ahead (Nanjing University and Tencent ARC) adds one training-only loss to an autoregressive video diffusion model: from the hidden states of the… https://media.thesatyajit.com/ai/films/looking-ahead-video.mp4?v=e5deaaeee2fe541c 111 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/mercury-3 https://media.thesatyajit.com/ai/thumbs/mercury-3.jpg?v=51d506192b59b2fb Mercury 3: a diffusion LLM at 1,068 tokens/s, and where the 4-8x goes in routed traffic https://media.thesatyajit.com/ai/films/mercury-3-poster.webp?v=415491c09c77e9ab Inception's Mercury 3 is a commercial diffusion language model: it drafts a block of tokens at once and refines it over a few parallel passes instead of… https://media.thesatyajit.com/ai/films/mercury-3.mp4?v=2bc152190b9721e1 110 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/ming-3 https://media.thesatyajit.com/ai/thumbs/ming-3.jpg?v=07057af63428abdb Ming-3.0 speech models: one 0.5B speech engine inside five checkpoints, run on a CPU https://media.thesatyajit.com/ai/films/ming-3-poster.webp?v=710dd1f13e6654d3 Ant Ling's inclusionAI released five MIT-licensed Ming-3.0 speech models: two text-to-speech, two audio-understanding and one recognizer. https://media.thesatyajit.com/ai/films/ming-3.mp4?v=58b58d0cd980f252 111 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/minimax-m3-1-flash https://media.thesatyajit.com/ai/thumbs/minimax-m3-1-flash.jpg?v=b9acb14bc64fec7c MiniMax M3.1 Flash: M3's shape, all-sparse attention, four-bit experts, and Space Bunny unmasked https://media.thesatyajit.com/ai/films/minimax-m3-1-flash-poster.webp?v=5aaafdfcb25a156f MiniMax released M3.1 Flash's weights on 23 October, four days after revealing that the anonymous 'Space Bunny' was an earlier version of it. https://media.thesatyajit.com/ai/films/minimax-m3-1-flash.mp4?v=8bfcbd8f9d737a84 105 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/nfsmw-online https://media.thesatyajit.com/ai/thumbs/nfsmw-online.jpg?v=b3ef949a31d215d3 NFSMW Online: Need for Speed Most Wanted in a browser tab, read from outside the login https://media.thesatyajit.com/ai/films/nfsmw-online-poster.webp?v=01ad3a7b26b9c9b9 nfsmwonline.com puts EA's 2005 Need for Speed: Most Wanted in Chrome and Edge, with accounts, server-side careers, online Free Roam and duels, and features… https://media.thesatyajit.com/ai/films/nfsmw-online.mp4?v=5cac92f8f0dad44d 107 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/nfsmw-web-port https://media.thesatyajit.com/ai/thumbs/nfsmw-web-port.jpg?v=4df6485a990ee3c6 NFS Most Wanted on itch.io: EA's 2005 speed.exe, recompiled ahead of time to WebAssembly https://media.thesatyajit.com/ai/films/nfsmw-web-port-poster.webp?v=3f612f9ce4f41570 dezm0n's itch.io build runs Need for Speed: Most Wanted in Chrome without the game's source. https://media.thesatyajit.com/ai/films/nfsmw-web-port.mp4?v=c5595eb1949cf099 115 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/nvx https://media.thesatyajit.com/ai/thumbs/nvx.jpg?v=2e6542630f1de098 NVX: Microsoft's micro-VM sandbox builds the machine up from nothing https://media.thesatyajit.com/ai/films/nvx-poster.webp?v=00ad607c39eee3fb Microsoft Research and Azure Research open-sourced NVX (MIT), a micro-VM sandbox for untrusted agent workloads built on a fork of OpenVMM. https://media.thesatyajit.com/ai/films/nvx.mp4?v=dc358b3997e6a4f5 109 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/open-math-harness https://media.thesatyajit.com/ai/thumbs/open-math-harness.jpg?v=1471b9d4eb289aeb Open Math Harness: a capability kernel for maths agents, with Lean as the only judge of a proof https://media.thesatyajit.com/ai/films/open-math-harness-poster.webp?v=d3c3fbf272bab979 SAIR Foundation open-sourced Open Math Harness, the Rust runtime behind Open Math OS: agents on open models, Ultramath swarms, a Dream-RSI proof-search tree… https://media.thesatyajit.com/ai/films/open-math-harness.mp4?v=a25a3f01e28fd6b1 110 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/reactive-optimizers https://media.thesatyajit.com/ai/thumbs/reactive-optimizers.jpg?v=7714148d879f896a Reactive optimizers: an agent that rewrites updater.py while the model trains https://media.thesatyajit.com/ai/films/reactive-optimizers-poster.webp?v=e63ae0e64547e242 Xingjian Bai, Jia-Hao Liang, Jayden Teoh and Vincent Sitzmann put a coding agent inside the training loop. https://media.thesatyajit.com/ai/films/reactive-optimizers.mp4?v=a6ba6dfecda9b42a 114 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/rlcr-avalon https://media.thesatyajit.com/ai/thumbs/rlcr-avalon.jpg?v=e90c09c8dee1db75 RLCR and Avalon-M1-Mini: training a whole agent organisation on one shared reward https://media.thesatyajit.com/ai/films/rlcr-avalon-poster.webp?v=40bc3a9c51a67e0d General Reasoning trains a 120B model to run as a CEO plus up to eight hired specialists who share a message board and a workspace, and gives every token in… https://media.thesatyajit.com/ai/films/rlcr-avalon.mp4?v=7e89ffeed860de2f 119 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/sglang-hicache https://media.thesatyajit.com/ai/thumbs/sglang-hicache.jpg?v=90b480e3221545c0 SGLang HiCache: the KV cache as a memory hierarchy, and a 98% that the workload sets https://media.thesatyajit.com/ai/films/sglang-hicache-poster.webp?v=e607ed3c8a45666f SGLang's HiCache keeps evicted prefixes in CPU memory and in external storage (Mooncake, 3FS, NIXL and ten other backends) so a returning conversation loads… https://media.thesatyajit.com/ai/films/sglang-hicache.mp4?v=de27ac88f374d6a8 112 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/sky-001 https://media.thesatyajit.com/ai/thumbs/sky-001.jpg?v=8b66c423ec3bc0a3 SKY-001: plan the part tree, draw the build, and lead CAD by 0.6 points https://media.thesatyajit.com/ai/films/sky-001-poster.webp?v=d566da06b6c6a015 Skyfall AI's SKY-001 is a closed, waitlist-only 'engineering world model' that splits an image into a tree of parts with diffusion transformers, then draws… https://media.thesatyajit.com/ai/films/sky-001.mp4?v=53ce2b7d3a409c11 105 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/spatialscout https://media.thesatyajit.com/ai/thumbs/spatialscout.jpg?v=15d1e1202ed14542 SpatialScout: a VLM drives the camera, submits 40 points, and is graded by a moved box https://media.thesatyajit.com/ai/films/spatialscout-poster.webp?v=0514435c1106caa1 SpatialScout gives a vision-language agent a list of 30 to 40 object IDs, a calibrated camera it can move and a triangulation tool, and scores the object… https://media.thesatyajit.com/ai/films/spatialscout.mp4?v=5ba110e010aa5cc7 113 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/stable-reinforce https://media.thesatyajit.com/ai/thumbs/stable-reinforce.jpg?v=097f32682646c1f7 StableReinforce: what Tinker's new off-policy loss publishes, and what it keeps https://media.thesatyajit.com/ai/films/stable-reinforce-poster.webp?v=eed40f525baf3fd0 Thinking Machines shipped StableReinforce, an alpha RL loss on Tinker meant to keep runs stable when the sampler drifts from the learner through async… https://media.thesatyajit.com/ai/films/stable-reinforce.mp4?v=e41ff69405527322 118 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/strong-stochastic-flow-maps https://media.thesatyajit.com/ai/thumbs/strong-stochastic-flow-maps.jpg?v=c3b542b04c4fb409 Improved strong stochastic flow maps: a smooth stand-in for the Brownian path https://media.thesatyajit.com/ai/films/strong-stochastic-flow-maps-poster.webp?v=8595a26cc26e6011 A strong stochastic flow map jumps an SDE from time s to time t along a given Brownian path, but that path is nowhere differentiable, so the derivative… https://media.thesatyajit.com/ai/films/strong-stochastic-flow-maps.mp4?v=4ce51e18ee12e0c2 111 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/torchtpu https://media.thesatyajit.com/ai/thumbs/torchtpu.jpg?v=bc1717ecb44a929f TorchTPU: PyTorch on TPUs as a real device, not a lazy tensor https://media.thesatyajit.com/ai/films/torchtpu-poster.webp?v=9f0aeb17472a4672 Google and Meta open-sourced TorchTPU at PyTorch Conference 2026: a PrivateUse1 backend that makes "tpu" a PyTorch device, lowers each ATen op to StableHLO… https://media.thesatyajit.com/ai/films/torchtpu.mp4?v=ffa0218a728ccacf 110 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/tv-launcher-rebuilds https://media.thesatyajit.com/ai/thumbs/tv-launcher-rebuilds.jpg?v=130b61a48746e2fb TV launcher rebuilds: what the agent actually runs over adb https://media.thesatyajit.com/ai/films/tv-launcher-rebuilds-poster.webp?v=480dc349af83b2a0 In October 2026 a few thousand people pointed Claude Code at their Android TVs, let it disable the vendor's ad and telemetry packages over adb, and had it… https://media.thesatyajit.com/ai/films/tv-launcher-rebuilds.mp4?v=b84dddb60f18c946 109 2026-10-23 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/deepseek-v4-1-flash-1017 https://media.thesatyajit.com/ai/thumbs/deepseek-v4-1-flash-1017.jpg?v=8ebf8e84d8f9ec5d DeepSeek-V4.1-Flash-1017: every byte retrained, nothing else touched, and what two DGX Sparks can hold https://media.thesatyajit.com/ai/films/deepseek-v4-1-flash-1017-poster.webp?v=45a4fbfd533d2b5d DeepSeek shipped DeepSeek-V4.1-Flash-1017 five weeks after V4.1-Flash, calling it a Pareto improvement. https://media.thesatyajit.com/ai/films/deepseek-v4-1-flash-1017.mp4?v=b70d0c65c41e1081 106 2026-10-17 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/pele https://media.thesatyajit.com/ai/thumbs/pele.jpg?v=5cdc38d7aac67419 Pele: Mermaid diagrams without the DOM, and where Mermaid's milliseconds actually go https://media.thesatyajit.com/ai/films/pele-poster.webp?v=ac730492279d08aa Pele is Steph Ango's new MIT renderer for Mermaid syntax, written from scratch for Obsidian 1.15: hand-written parsers checked against Mermaid's own… https://media.thesatyajit.com/ai/films/pele.mp4?v=78570ba814efce40 105 2026-10-17 yes 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/ling-3-1-flash https://media.thesatyajit.com/ai/thumbs/ling-3-1-flash.jpg?v=4cee22e85cfa7ccf 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/3d-reconstruction-roundup-2 https://media.thesatyajit.com/ai/thumbs/3d-reconstruction-roundup-2.jpg?v=fe21df1e4656f099 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/embeddinggemma-2 https://media.thesatyajit.com/ai/thumbs/embeddinggemma-2.jpg?v=38653f76c41b4a1f 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/openai-math https://media.thesatyajit.com/ai/thumbs/openai-math.jpg?v=bdd8467db15aa673 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/trackeverything https://media.thesatyajit.com/ai/thumbs/trackeverything.jpg?v=1e51148f28af1a7c 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/ornith-1-5 https://media.thesatyajit.com/ai/thumbs/ornith-1-5.jpg?v=45848d39c8a38b94 2026-10-23 monthly 0.7 https://ai.satyajit.si/articles/activation-offload https://media.thesatyajit.com/ai/thumbs/activation-offload.jpg?v=17e4f48da9528329 Activation checkpoint offload: moving the last per-layer tensor off the GPU https://media.thesatyajit.com/ai/films/activation-offload-poster.webp?v=41d2d3b21f89d462 Hugging Face Transformers 5.17 can hold gradient-checkpointing's saved layer inputs in pinned CPU memory instead of on the GPU. https://media.thesatyajit.com/ai/films/activation-offload.mp4?v=48658c6a6b952d9b 101 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/bitcoin-0-1-rebuilt https://media.thesatyajit.com/ai/thumbs/bitcoin-0-1-rebuilt.jpg?v=6629e2e8dea1c759 Bitcoin 0.1, rebuilt byte for byte: reading a 2009 build machine out of its executable https://media.thesatyajit.com/ai/films/bitcoin-0-1-rebuilt-poster.webp?v=ed0be0a22f2ecef4 t4t5/bitcoin-0.1-rebuilt compiles Satoshi Nakamoto's January 2009 source with a reconstructed MinGW toolchain under Wine and gets a bitcoin.exe whose… https://media.thesatyajit.com/ai/films/bitcoin-0-1-rebuilt.mp4?v=37b1a246a87d80ec 107 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/forspecies20k https://media.thesatyajit.com/ai/thumbs/forspecies20k.jpg?v=6db4bb568afd0eed FOR-species20K at 85%: a tree's point cloud as seven pictures, and the source shortcut under the score https://media.thesatyajit.com/ai/films/forspecies20k-poster.webp?v=319e639c7cb9d6ce Purdue's WACV 2027 entry tops the FOR-species20K tree-species leaderboard at 85% by rendering each LiDAR scan into seven depth, density and height images… https://media.thesatyajit.com/ai/films/forspecies20k.mp4?v=5445967f9b82267d 105 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/kimi-k3-gb300-megakernel https://media.thesatyajit.com/ai/thumbs/kimi-k3-gb300-megakernel.jpg?v=dc7ca145957f7ec6 Kimi K3 on GB300: 13 kernels a layer, and what 842 and 423 tokens a second measure https://media.thesatyajit.com/ai/films/kimi-k3-gb300-megakernel-poster.webp?v=a39058aa600d565e vLLM merged a CuTe DSL decode path for Kimi K3 on NVIDIA GB300 that fuses the attention projections, the MoE and the all-reduces into 13 kernels a layer… https://media.thesatyajit.com/ai/films/kimi-k3-gb300-megakernel.mp4?v=b117761d91f87ea5 117 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/megabake https://media.thesatyajit.com/ai/thumbs/megabake.jpg?v=1201b6a6fb2556ee MegaBake: a PyTorch model compiled into one CUDA kernel, and what its 5.5× is measured against https://media.thesatyajit.com/ai/films/megabake-poster.webp?v=4904bb06d437f0d2 MegaBake is an open (MIT) torch.compile backend that takes the post-grad FX graph of a Hugging Face decoder and emits one persistent CuTe DSL kernel: 60… https://media.thesatyajit.com/ai/films/megabake.mp4?v=089ede92cc597b97 102 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/nemotron-dory https://media.thesatyajit.com/ai/thumbs/nemotron-dory.jpg?v=d23b0c794912dbdc Nemotron-Labs Dory1.0-4B: one 20-layer block, run up to four times https://media.thesatyajit.com/ai/films/nemotron-dory-poster.webp?v=a27cb1ac9aefe126 NVIDIA's Dory1.0-4B is a looped language model: 8 input layers, a 20-layer block that runs a fixed number of times per token, and 8 output layers, on a… https://media.thesatyajit.com/ai/films/nemotron-dory.mp4?v=e46a7edb4067e014 102 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/ssftalk https://media.thesatyajit.com/ai/thumbs/ssftalk.jpg?v=b31395421a96f8d4 ssftalk: an ElevenLabs-shaped API over a formant synthesizer, benchmarked against typed-in numbers https://media.thesatyajit.com/ai/films/ssftalk-poster.webp?v=00bb6d7271037ca9 solidSF's ssftalk calls itself a pure-Rust ElevenLabs v1 voice engine with full-duplex rooms and audio-LLM mechanistic interpretability, 265.5x faster than… https://media.thesatyajit.com/ai/films/ssftalk.mp4?v=a7cf38de95982645 112 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/steadysplats https://media.thesatyajit.com/ai/thumbs/steadysplats.jpg?v=d3167a40ee4ddc54 SteadySplats: teach the Gaussians to agree, and a one-sample splat renderer stops boiling https://media.thesatyajit.com/ai/films/steadysplats-poster.webp?v=3cbac17a1fa4befb SteadySplats (TU Graz, NVIDIA, Google DeepMind, Copenhagen) makes sort-free stochastic Gaussian-splat rendering usable at one sample per pixel. https://media.thesatyajit.com/ai/films/steadysplats.mp4?v=5b5bea3cc4721822 111 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/virtual-head-attention https://media.thesatyajit.com/ai/thumbs/virtual-head-attention.jpg?v=c95fb1a23bcdba12 Virtual Head Attention: train with sixteen heads' worth of queries, serve plain GQA-2 https://media.thesatyajit.com/ai/films/virtual-head-attention-poster.webp?v=82a01aacb6b649ce Baidu's VHA halves the physical query heads, drops to two K/V heads, and adds two linear maps: a per-group Premix that turns 8 query heads into 16 virtual… https://media.thesatyajit.com/ai/films/virtual-head-attention.mp4?v=a2d41e47f32dbda8 105 2026-10-19 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/earshot https://media.thesatyajit.com/ai/thumbs/earshot.jpg?v=0b3526263b2e8659 earshot fit: a sound file in, a Python recipe out, and no neural network in between https://media.thesatyajit.com/ai/films/earshot-poster.webp?v=6f55d087fa6fd847 earshot v0.2 turns a recording into a Python module whose render() makes the sound again. https://media.thesatyajit.com/ai/films/earshot.mp4?v=7abc829eb177f84a 107 2026-10-17 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/perplexity-space https://media.thesatyajit.com/ai/thumbs/perplexity-space.jpg?v=8348dd27f5879b75 Perplexity SPACE: a Firecracker microVM per agent task, and the keys stay outside https://media.thesatyajit.com/ai/films/perplexity-space-poster.webp?v=47cfb2cb0bd22b36 Perplexity open-sourced SPACE (Apache-2.0, Rust), the sandbox platform behind every Perplexity Computer task: one jailed Firecracker microVM per sandbox, a… https://media.thesatyajit.com/ai/films/perplexity-space.mp4?v=c87283b1199c77ff 111 2026-10-16 yes 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/orcacyber-zero-1-5 https://media.thesatyajit.com/ai/thumbs/orcacyber-zero-1-5.jpg?v=14481a1cdec034f4 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/reflection-beam https://media.thesatyajit.com/ai/thumbs/reflection-beam.jpg?v=68b18192c646da29 2026-10-19 monthly 0.7 https://ai.satyajit.si/articles/3dcodeverse https://media.thesatyajit.com/ai/thumbs/3dcodeverse.jpg?v=1ecb210170f164e9 3DCodeVerse: 2.03M rows of 3D code, and a harness whose score is capped by a ruler https://media.thesatyajit.com/ai/films/3dcodeverse-poster.webp?v=247c596ab48250fe 3DCodeVerse packs CAD scripts, Blender generators, Shadertoy shaders, OpenSCAD parts and three.js pages into one corpus of executable 3D programs, and wraps… https://media.thesatyajit.com/ai/films/3dcodeverse.mp4?v=607b1b142fc00c1a 106 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/infinisplat-2 https://media.thesatyajit.com/ai/thumbs/infinisplat-2.jpg?v=88017bdd9d89a570 InfiniSplat v2: the points stay where Depth Anything 3 put them, and opacity does the cleaning https://media.thesatyajit.com/ai/films/infinisplat-2-poster.webp?v=5205a2e3704f117c ZJU3DV's InfiniSplat v2 turns 2 to 50 images into a 3D Gaussian scene in one forward pass, on top of a frozen Depth Anything 3. https://media.thesatyajit.com/ai/films/infinisplat-2.mp4?v=78e980a2627752f1 117 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/langevin-flow-maps https://media.thesatyajit.com/ai/thumbs/langevin-flow-maps.jpg?v=adb975b0eeaedb28 Langevin Flow Maps: a force field that learns its own integrator https://media.thesatyajit.com/ai/films/langevin-flow-maps-poster.webp?v=481340e90afd3784 Molecular dynamics crawls because the integrator needs femtosecond steps, not because the force field is slow. https://media.thesatyajit.com/ai/films/langevin-flow-maps.mp4?v=1d47acae3a00505f 115 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/prompt-caching-harnesses https://media.thesatyajit.com/ai/thumbs/prompt-caching-harnesses.jpg?v=0852758dffda6df0 Prompt caching in coding agents: where eight harnesses put the breakpoint https://media.thesatyajit.com/ai/films/prompt-caching-harnesses-poster.webp?v=fe9a3dc1646745c6 Every provider caches a prompt prefix, so a coding agent's bill depends on what it puts first and what it lets change. https://media.thesatyajit.com/ai/films/prompt-caching-harnesses.mp4?v=c0bbfe516811fac2 101 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/rsi-inference-bench https://media.thesatyajit.com/ai/thumbs/rsi-inference-bench.jpg?v=55148e10154b1d70 RSI-Inference Bench: the 26.7% is ten tasks, and the agents beat the maintainers' fixes https://media.thesatyajit.com/ai/films/rsi-inference-bench-poster.webp?v=60a4e1443cdb53ee Apple ML Research's RSI-Inference Bench asks whether coding agents can improve the engines that serve them: 37 tasks on real vLLM and SGLang trees, each an… https://media.thesatyajit.com/ai/films/rsi-inference-bench.mp4?v=b3b7620e195fa546 109 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/rsi-top https://media.thesatyajit.com/ai/thumbs/rsi-top.jpg?v=4f1e8b4a8312b117 RSI-top: a dashboard and a diverse cohort for autoresearch runs that stall https://media.thesatyajit.com/ai/films/rsi-top-poster.webp?v=4b6f5f5d4cac99a0 RSI-top (Stanford Scaling Intelligence Lab, 16 Oct 2026) watches a long autoresearch run, and when the best score stops moving it pauses the native search… https://media.thesatyajit.com/ai/films/rsi-top.mp4?v=8ad240e248c4e85d 106 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/speech-cpp https://media.thesatyajit.com/ai/thumbs/speech-cpp.jpg?v=499f4ce9a1d9a217 speech.cpp: six speech model families in one ggml binary, built and heard on a shared CPU https://media.thesatyajit.com/ai/films/speech-cpp-poster.webp?v=1d2d9e10cfe6fc5d speech.cpp puts Qwen3-TTS, Irodori-TTS, Fish Audio's Drama 3, Qwen3-ASR, NVIDIA's Parakeet and ReazonSpeech, and Silero VAD behind one C API and one… https://media.thesatyajit.com/ai/films/speech-cpp.mp4?v=614abbc5de3a1c2c 106 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/timestral https://media.thesatyajit.com/ai/thumbs/timestral.jpg?v=0deacfc1918c0659 Timestral: Mistral's 209M forecaster, read from the code and run on two CPU threads https://media.thesatyajit.com/ai/films/timestral-poster.webp?v=5a61f8e56074968c Mistral's first time-series foundation model cuts a series into 32-step patches, runs a 24-layer causal transformer over them, and answers with 21 quantiles… https://media.thesatyajit.com/ai/films/timestral.mp4?v=77a9b76a2e4a6190 106 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/zephon-torchtitan https://media.thesatyajit.com/ai/thumbs/zephon-torchtitan.jpg?v=5f488873594d0e7b Zephon in TorchTitan: the same global batch on 8, 16 or 64 GPUs https://media.thesatyajit.com/ai/films/zephon-torchtitan-poster.webp?v=553fe94133ac304f Zephon is a data loader that tokenizes, packs, shuffles and mixes inside the training job and still hands the model the same global batch when a run resumes… https://media.thesatyajit.com/ai/films/zephon-torchtitan.mp4?v=5df0f3a1803ed130 119 2026-10-17 yes 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/fish-drama-3 https://media.thesatyajit.com/ai/thumbs/fish-drama-3.jpg?v=ce6e7aab405e991f 2026-10-17 monthly 0.7 https://ai.satyajit.si/articles/bonsai-2-small https://media.thesatyajit.com/ai/thumbs/bonsai-2-small.jpg?v=4ad1a5394c18f1d2 Bonsai 2 2B, 4B and 9B: where the bytes go when the model is small https://media.thesatyajit.com/ai/films/bonsai-2-small-poster.webp?v=016f73fd883934bb PrismML's small ternary Bonsai 2 models ship as 0.43, 0.95 and 2.0 GB language-model files, and the GGUF headers confirm every byte of that: the three… https://media.thesatyajit.com/ai/films/bonsai-2-small.mp4?v=898f26e6d2a73508 104 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/comet-60b https://media.thesatyajit.com/ai/thumbs/comet-60b.jpg?v=2c5de07286d7662c Comet 60B: Nemotron 3.5 Super VL with half of every expert pool cut https://media.thesatyajit.com/ai/films/comet-60b-poster.webp?v=86df88d34a9116ba Multiverse Computing's Comet 60B is NVIDIA's Nemotron 3.5 Super VL with half of its routed experts removed. https://media.thesatyajit.com/ai/films/comet-60b.mp4?v=cca00abb6f716bd8 107 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/datadog-scout https://media.thesatyajit.com/ai/thumbs/datadog-scout.jpg?v=e5860bb595108060 Datadog's Scout: RL on simulated outages, and what "as strong as Fable 5.1" rests on https://media.thesatyajit.com/ai/films/datadog-scout-poster.webp?v=5e337faf4f6a8fd8 Datadog post-trained GLM-5.3 Flash with asynchronous GRPO on simulated incidents to make Scout 1.0, the model behind its Bits Investigation SRE agent, and… https://media.thesatyajit.com/ai/films/datadog-scout.mp4?v=8ee9938057f54470 106 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/hy-memory https://media.thesatyajit.com/ai/thumbs/hy-memory.jpg?v=bf87fe048c081f8e Hy Memory: a 3B-active memory writer that ties Opus 5, and four launch numbers that belong to the engine https://media.thesatyajit.com/ai/films/hy-memory-poster.webp?v=e085495761978913 Tencent Hunyuan open-sourced a memory stack: the Hy Memory engine, the MemHub app and Hy-Memory-30B-A3B, a model trained only to write memories. https://media.thesatyajit.com/ai/films/hy-memory.mp4?v=5cb325d337bf7795 112 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/jais-3 https://media.thesatyajit.com/ai/thumbs/jais-3.jpg?v=16d0f75d428892d1 Jais 3: adapting K2 Horizon to Arabic without touching the tokenizer https://media.thesatyajit.com/ai/films/jais-3-poster.webp?v=f309d1b9bc3838a2 IFM's Jais 3 is six K2 Horizon models continued-pretrained into Arabic, plus Jais Speech, an ASR model built on the 3.7B. https://media.thesatyajit.com/ai/films/jais-3.mp4?v=977facc2a0348dc0 108 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/mai-2 https://media.thesatyajit.com/ai/thumbs/mai-2.jpg?v=a03e9abbc3cba7b1 MAI-Base-2: Microsoft's 2.8T base model, and whether its compute numbers add up https://media.thesatyajit.com/ai/films/mai-2-poster.webp?v=f869479cc4e3ffec Microsoft AI finished pre-training MAI-Base-2, an 80B-active, 2.8T-total MoE trained with Muon on 50T tokens across 16k GB200s in 38 days. https://media.thesatyajit.com/ai/films/mai-2.mp4?v=d2ae93ee69376ae3 123 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/muse-spark-open https://media.thesatyajit.com/ai/thumbs/muse-spark-open.jpg?v=c2672236dfddfb03 Muse Spark Open: 816B on two Mac Studios, and what the arithmetic needs https://media.thesatyajit.com/ai/films/muse-spark-open-poster.webp?v=7c26d35ca75c3a8f Meta released Muse Spark Open on 15 October 2026: an 816B-total, 42B-active multimodal MoE with a 1M-token context, 'optimized to run on two Apple M5 Ultra… https://media.thesatyajit.com/ai/films/muse-spark-open.mp4?v=2a13e3a736a136d1 104 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/panst3r https://media.thesatyajit.com/ai/thumbs/panst3r.jpg?v=13df92fce0857b57 PanSt3R: one set of queries for every view, so the chair keeps its id https://media.thesatyajit.com/ai/films/panst3r-poster.webp?v=183c36ad929893e7 NAVER LABS Europe's PanSt3R predicts a 3D point map and a panoptic labelling for a pile of unposed photos in one forward pass. https://media.thesatyajit.com/ai/films/panst3r.mp4?v=9b4b5164eabe9236 112 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/terminal-rsi https://media.thesatyajit.com/ai/thumbs/terminal-rsi.jpg?v=6f6148c92cffd48f Terminal-RSI: the second cycle, and how much smaller it is https://media.thesatyajit.com/ai/films/terminal-rsi-poster.webp?v=6ad327ee4d448379 Terminal-RSI (arXiv 2610.22192) runs Recursive Self-Rewrite twice, with PPO in between: the model solves terminal tasks, rewrites its own verified successes… https://media.thesatyajit.com/ai/films/terminal-rsi.mp4?v=a2999beb620671c8 116 2026-10-16 yes 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/atria-dawn https://media.thesatyajit.com/ai/thumbs/atria-dawn.jpg?v=1c1754931b767a52 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/step-5-preview https://media.thesatyajit.com/ai/thumbs/step-5-preview.jpg?v=a589e680b5ef285b 2026-10-16 monthly 0.7 https://ai.satyajit.si/articles/blender-5-3-gaussian-splats https://media.thesatyajit.com/ai/thumbs/blender-5-3-gaussian-splats.jpg?v=b8380e471d2704f1 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/flux-3-image-klein https://media.thesatyajit.com/ai/thumbs/flux-3-image-klein.jpg?v=d05e7d6fe076774d 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/jev-alternatives-week-five https://media.thesatyajit.com/ai/thumbs/jev-alternatives-week-five.jpg?v=17fa646df1c37cf7 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/matrix-game-4 https://media.thesatyajit.com/ai/thumbs/matrix-game-4.jpg?v=b3f0803d3f3d41cc 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/nemotron-3-5-super-vl https://media.thesatyajit.com/ai/thumbs/nemotron-3-5-super-vl.jpg?v=d488eaff74e17261 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/ps5lm https://media.thesatyajit.com/ai/thumbs/ps5lm.jpg?v=46a6d2c3d7f044ae 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/sana-wm https://media.thesatyajit.com/ai/thumbs/sana-wm.jpg?v=57c9cc4017c033f9 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/spatiallm https://media.thesatyajit.com/ai/thumbs/spatiallm.jpg?v=a901cd4b2eae77f3 2026-10-15 monthly 0.7 https://ai.satyajit.si/articles/fastvideo-fasth3 https://media.thesatyajit.com/ai/thumbs/fastvideo-fasth3.jpg?v=f1bdbe98fab84ce8 FastH3 Preview v1: four calls, one-tenth the attention, and what the 14x actually measures https://media.thesatyajit.com/ai/films/fastvideo-fasth3-poster.webp?v=dc77a9a1e1a2d518 FastVideo distills MiniMax's 33B H3 video-audio transformer into a 4-step, 90%-sparse generator and calls the recommended checkpoint Data-Free because no… https://media.thesatyajit.com/ai/films/fastvideo-fasth3.mp4?v=c3f2a535398ee1ce 105 2026-10-07 yes 2026-10-15 monthly 0.8 https://ai.satyajit.si/articles/anthropic-source-control https://media.thesatyajit.com/ai/thumbs/anthropic-source-control.jpg?v=d8c7487881494d11 2026-10-14 monthly 0.7 https://ai.satyajit.si/articles/fineenvs-multilingual https://media.thesatyajit.com/ai/thumbs/fineenvs-multilingual.jpg?v=a01dda9494367187 2026-10-14 monthly 0.7 https://ai.satyajit.si/articles/lfm2-5-24b-a2b https://media.thesatyajit.com/ai/thumbs/lfm2-5-24b-a2b.jpg?v=9ecfddc7075e4673 2026-10-14 monthly 0.7 https://ai.satyajit.si/articles/tev-gym https://media.thesatyajit.com/ai/thumbs/tev-gym.jpg?v=b909194d0d6ddc5b 2026-10-14 monthly 0.7 https://ai.satyajit.si/articles/ultrabert-embed https://media.thesatyajit.com/ai/thumbs/ultrabert-embed.jpg?v=2f087c64909cddd2 2026-10-14 monthly 0.7 https://ai.satyajit.si/articles/ultramarine-1-flash https://media.thesatyajit.com/ai/thumbs/ultramarine-1-flash.jpg?v=e86de9230e74d837 2026-10-14 monthly 0.7 https://ai.satyajit.si/articles/aa-3d https://media.thesatyajit.com/ai/thumbs/aa-3d.jpg?v=c1534e935e56a48f 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/astronex-world https://media.thesatyajit.com/ai/thumbs/astronex-world.jpg?v=994b001fd12205b2 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/budgetpix https://media.thesatyajit.com/ai/thumbs/budgetpix.jpg?v=8ac5ad63c0bb32d6 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/data-space-iteration https://media.thesatyajit.com/ai/thumbs/data-space-iteration.jpg?v=aa3530f3fd99026d 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/diffusionart https://media.thesatyajit.com/ai/thumbs/diffusionart.jpg?v=5d704f114891d994 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/groundingpi https://media.thesatyajit.com/ai/thumbs/groundingpi.jpg?v=d10da2a9d7466c37 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/minicpm-v-4-7 https://media.thesatyajit.com/ai/thumbs/minicpm-v-4-7.jpg?v=2efad4163dc7e970 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/mlattice https://media.thesatyajit.com/ai/thumbs/mlattice.jpg?v=7ff381e7e3180419 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/muse-wonder https://media.thesatyajit.com/ai/thumbs/muse-wonder.jpg?v=149c0257a1b294fb Muse Wonder: what the two points behind Opus 5.5 are made of https://media.thesatyajit.com/ai/films/muse-wonder-poster.webp?v=5c25a95da039eb4f Meta Superintelligence Labs' Muse Wonder Full scores 55.80 on the Artificial Analysis Intelligence Index, 1.82 below Claude Opus 5.5. https://media.thesatyajit.com/ai/films/muse-wonder.mp4?v=2a1d4ce2c794dbe0 99 2026-10-13 yes 2026-10-13 monthly 0.8 https://ai.satyajit.si/articles/natura-1 https://media.thesatyajit.com/ai/thumbs/natura-1.jpg?v=22e844a41515965d 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/opentakeoff https://media.thesatyajit.com/ai/thumbs/opentakeoff.jpg?v=044ed728e76a93b5 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/qwengyre https://media.thesatyajit.com/ai/thumbs/qwengyre.jpg?v=bfef51c249b309aa 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/replayssm https://media.thesatyajit.com/ai/thumbs/replayssm.jpg?v=246574ac2bf3c9f2 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/representation-guidance https://media.thesatyajit.com/ai/thumbs/representation-guidance.jpg?v=93ff37f3bbc1b727 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/streaming-code-world-model https://media.thesatyajit.com/ai/thumbs/streaming-code-world-model.jpg?v=e24a4a36a18d4e07 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/subq-snap https://media.thesatyajit.com/ai/thumbs/subq-snap.jpg?v=eca024520165a487 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/usd-brep-nurbs-kernel https://media.thesatyajit.com/ai/thumbs/usd-brep-nurbs-kernel.jpg?v=14d27ec6141eeea2 usd-brep: NVIDIA open-sources SMLib, a million-line NURBS kernel with a USD schema on top https://media.thesatyajit.com/ai/films/usd-brep-nurbs-kernel-poster.webp?v=ca9c7960aa0cb4c4 NVIDIA released usd-brep under Apache-2.0: the SMLib boundary-representation kernel (967,172 lines of C++), a C API, Python bindings and a BrepArray schema… https://media.thesatyajit.com/ai/films/usd-brep-nurbs-kernel.mp4?v=24a929188c862f8f 109 2026-10-13 yes 2026-10-13 monthly 0.8 https://ai.satyajit.si/articles/xiryn https://media.thesatyajit.com/ai/thumbs/xiryn.jpg?v=37a945743e5dbc57 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/fsearch https://media.thesatyajit.com/ai/thumbs/fsearch.jpg?v=cd24d5a9b96db6e4 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/dust-zeroth-order https://media.thesatyajit.com/ai/thumbs/dust-zeroth-order.jpg?v=abc4361137d4f983 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/jacobian-lens https://media.thesatyajit.com/ai/thumbs/jacobian-lens.jpg?v=91e87a3653b806a4 2026-10-13 monthly 0.7 https://ai.satyajit.si/articles/personaplex-full-duplex https://media.thesatyajit.com/ai/thumbs/personaplex-full-duplex.jpg?v=666e55d9f636ca15 2026-10-12 monthly 0.7 https://ai.satyajit.si/articles/rsi-anything https://media.thesatyajit.com/ai/thumbs/rsi-anything.jpg?v=cf623a5fb7a3f0a3 RSI-Anything: the world model ranks, the policy learns, and they rarely meet https://media.thesatyajit.com/ai/films/rsi-anything-poster.webp?v=c5e6ff7f058c81a7 RSI-Anything puts a JEPA-Anything world model inside a self-improvement loop: the policy proposes, the world model predicts, the best-predicted candidate… https://media.thesatyajit.com/ai/films/rsi-anything.mp4?v=2665466f1499abe4 110 2026-10-12 yes 2026-10-12 monthly 0.8 https://ai.satyajit.si/articles/streamrender-h3 https://media.thesatyajit.com/ai/thumbs/streamrender-h3.jpg?v=16961263954630c8 2026-10-12 monthly 0.7 https://ai.satyajit.si/articles/voicestudio https://media.thesatyajit.com/ai/thumbs/voicestudio.jpg?v=4718e4a7850a87af 2026-10-12 monthly 0.7 https://ai.satyajit.si/articles/jev-alternatives-week-four https://media.thesatyajit.com/ai/thumbs/jev-alternatives-week-four.jpg?v=ab52457a21531c93 2026-10-12 monthly 0.7 https://ai.satyajit.si/articles/aes-circuits-lean https://media.thesatyajit.com/ai/thumbs/aes-circuits-lean.jpg?v=c8113e48b7c1f213 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/ai-short-drama-pipelines https://media.thesatyajit.com/ai/thumbs/ai-short-drama-pipelines.jpg?v=e4a5aa422f64d6f6 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/anatomy-machine-skill https://media.thesatyajit.com/ai/thumbs/anatomy-machine-skill.jpg?v=da062d59643ddb60 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/deepgemm https://media.thesatyajit.com/ai/thumbs/deepgemm.jpg?v=7b84cbc46169aeb6 DeepGEMM: the 300-line FP8 kernel grew into a 35,000-line library https://media.thesatyajit.com/ai/films/deepgemm-poster.webp?v=0a72abca65daf28f DeepSeek's DeepGEMM shipped during Open Source Week 2025 as one Hopper FP8 kernel of about 300 lines, built around two ideas: scale FP8 per 128 elements… https://media.thesatyajit.com/ai/films/deepgemm.mp4?v=22c723f3949660fb 106 2026-10-12 yes 2026-10-11 monthly 0.8 https://ai.satyajit.si/articles/invaria-point-clouds https://media.thesatyajit.com/ai/thumbs/invaria-point-clouds.jpg?v=39e935725e94ac5d 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/lobe-rsi-harness https://media.thesatyajit.com/ai/thumbs/lobe-rsi-harness.jpg?v=44860fcbb61810dc 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/oruk-looped-speech https://media.thesatyajit.com/ai/thumbs/oruk-looped-speech.jpg?v=be123a9bef12aec2 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/structuregs-slam https://media.thesatyajit.com/ai/thumbs/structuregs-slam.jpg?v=fa2b487e41c24ddb 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/vivix-w1 https://media.thesatyajit.com/ai/thumbs/vivix-w1.jpg?v=b56194841b2a9e90 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/clef-omni-and-microsoft-decision-1 https://media.thesatyajit.com/ai/thumbs/clef-omni-and-microsoft-decision-1.jpg?v=4989527826b795da 2026-10-11 monthly 0.7 https://ai.satyajit.si/articles/dense-correspondence-flux2 https://media.thesatyajit.com/ai/thumbs/dense-correspondence-flux2.jpg?v=d0495110e15240e6 2026-10-10 monthly 0.7 https://ai.satyajit.si/articles/kronos-financial-foundation-model https://media.thesatyajit.com/ai/thumbs/kronos-financial-foundation-model.jpg?v=4ff5e05041dcf4d4 2026-10-10 monthly 0.7 https://ai.satyajit.si/articles/strands-agents-and-box https://media.thesatyajit.com/ai/thumbs/strands-agents-and-box.jpg?v=befe85186c5da0c8 2026-10-10 monthly 0.7 https://ai.satyajit.si/articles/underdog-saluki-27b https://media.thesatyajit.com/ai/thumbs/underdog-saluki-27b.jpg?v=bf03ee96c85d11bc 2026-10-10 monthly 0.7 https://ai.satyajit.si/articles/marble-camera-studio https://media.thesatyajit.com/ai/thumbs/marble-camera-studio.jpg?v=b5e5256dc170eb8c 2026-10-09 monthly 0.7 https://ai.satyajit.si/articles/nvidia-macos-driver https://media.thesatyajit.com/ai/thumbs/nvidia-macos-driver.jpg?v=fc64cc70e7ef1642 2026-10-09 monthly 0.7 https://ai.satyajit.si/articles/anyps5-static-relinker https://media.thesatyajit.com/ai/thumbs/anyps5-static-relinker.jpg?v=a1f7bf531ce85b39 2026-10-09 monthly 0.7 https://ai.satyajit.si/articles/claude-dashboards-motion https://media.thesatyajit.com/ai/thumbs/claude-dashboards-motion.jpg?v=6884611aa9f840b6 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/claude-haiku-5-5 https://media.thesatyajit.com/ai/thumbs/claude-haiku-5-5.jpg?v=7dd76c3dd7adf3a1 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/karotte-rl-environments https://media.thesatyajit.com/ai/thumbs/karotte-rl-environments.jpg?v=2d7d519ab2a849b2 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/layerrope https://media.thesatyajit.com/ai/thumbs/layerrope.jpg?v=2fc2338ff23fd75b 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/lithos-metal https://media.thesatyajit.com/ai/thumbs/lithos-metal.jpg?v=52ebdd64efac1083 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/littlebit-0-1-bit https://media.thesatyajit.com/ai/thumbs/littlebit-0-1-bit.jpg?v=282d27858c69bdf4 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/mellum2-1 https://media.thesatyajit.com/ai/thumbs/mellum2-1.jpg?v=df04773b1b52396a 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/mimo-env-reward-hacking https://media.thesatyajit.com/ai/thumbs/mimo-env-reward-hacking.jpg?v=d1b85162d104a1bb 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/paradee-tiny-tts https://media.thesatyajit.com/ai/thumbs/paradee-tiny-tts.jpg?v=89011f578e37ffe4 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/pplx-embed-v2-late https://media.thesatyajit.com/ai/thumbs/pplx-embed-v2-late.jpg?v=58b49870e88579d7 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/rea-reverse-engineer-anything https://media.thesatyajit.com/ai/thumbs/rea-reverse-engineer-anything.jpg?v=4317c256e89de02d 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/roomtsc https://media.thesatyajit.com/ai/thumbs/roomtsc.jpg?v=514f2435239d0d12 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/t3-video-4k https://media.thesatyajit.com/ai/thumbs/t3-video-4k.jpg?v=48d3b20b8f9e547a 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/unreal-retrieval-long-context https://media.thesatyajit.com/ai/thumbs/unreal-retrieval-long-context.jpg?v=9ce1a576cec3573d 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/vllm-omni-serving https://media.thesatyajit.com/ai/thumbs/vllm-omni-serving.jpg?v=48a4f094316026c7 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/windows-local-coding-models https://media.thesatyajit.com/ai/thumbs/windows-local-coding-models.jpg?v=e85513ac999ade0a 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/liquid-open-d1 https://media.thesatyajit.com/ai/thumbs/liquid-open-d1.jpg?v=0bae7b871af50027 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/agent-memory-repo https://media.thesatyajit.com/ai/thumbs/agent-memory-repo.jpg?v=59ce5076b7272a6b 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/interfaze-1-lite https://media.thesatyajit.com/ai/thumbs/interfaze-1-lite.jpg?v=3185bcf9bc886410 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/restore-and-relight-loras https://media.thesatyajit.com/ai/thumbs/restore-and-relight-loras.jpg?v=d9573c5676cfd3da 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/mimo-rl-environments https://media.thesatyajit.com/ai/thumbs/mimo-rl-environments.jpg?v=2fbe5f43110fa099 2026-10-08 monthly 0.7 https://ai.satyajit.si/articles/distscene https://media.thesatyajit.com/ai/thumbs/distscene.jpg?v=d2c5c0d31d145c7f 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/fbtriton-table-batched-embeddings https://media.thesatyajit.com/ai/thumbs/fbtriton-table-batched-embeddings.jpg?v=011864ee665bd433 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/impeccable-design-skill https://media.thesatyajit.com/ai/thumbs/impeccable-design-skill.jpg?v=9b44c9139fc500f1 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/kv-caching-flow-models https://media.thesatyajit.com/ai/thumbs/kv-caching-flow-models.jpg?v=c2f7dee988a93387 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/minimal-ml-engineering-agent https://media.thesatyajit.com/ai/thumbs/minimal-ml-engineering-agent.jpg?v=3cb89c51af39b1b5 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/openqodex-code-review https://media.thesatyajit.com/ai/thumbs/openqodex-code-review.jpg?v=2bada4a3576a2575 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/squeeze3d https://media.thesatyajit.com/ai/thumbs/squeeze3d.jpg?v=1b0101098cd4d816 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/unsloth-train-a-decision-model https://media.thesatyajit.com/ai/thumbs/unsloth-train-a-decision-model.jpg?v=c91d1c9e0cea2b67 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/gta5-in-the-browser https://media.thesatyajit.com/ai/thumbs/gta5-in-the-browser.jpg?v=50de93df7d304ea5 GTA V in a browser tab: Rockstar's own engine, recompiled to wasm64 and WebGPU https://media.thesatyajit.com/ai/films/gta5-in-the-browser-poster.webp?v=8e78ead8a5e17946 playgta5.com runs Grand Theft Auto V in Chrome from Rockstar's own RAGE engine, compiled with Emscripten into a 63 MB wasm64 module with about 91,000 named… https://media.thesatyajit.com/ai/films/gta5-in-the-browser.mp4?v=87808f06742010e2 107 2026-10-07 yes 2026-10-07 monthly 0.8 https://ai.satyajit.si/articles/kandinsky-6-video https://media.thesatyajit.com/ai/thumbs/kandinsky-6-video.jpg?v=ad9676dd8435cb23 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/limite-1b-violetto https://media.thesatyajit.com/ai/thumbs/limite-1b-violetto.jpg?v=d15e486b914756d4 Limite 1B - Violetto: the throughput claim is in config.json, not the blog https://media.thesatyajit.com/ai/films/limite-1b-violetto-poster.webp?v=925959ee172f6361 Paradigma's README calls Violetto 'a model for high-frequency mathematical intelligence' and nothing it shipped publishes a tokens-per-second figure. https://media.thesatyajit.com/ai/films/limite-1b-violetto.mp4?v=9e073fa067cbad67 115 2026-10-07 yes 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/kaggle-tpu-320b https://media.thesatyajit.com/ai/thumbs/kaggle-tpu-320b.jpg?v=9bdfcec7de893d26 2026-10-07 monthly 0.7 https://ai.satyajit.si/articles/agent-tools-week https://media.thesatyajit.com/ai/thumbs/agent-tools-week.jpg?v=769e282b727a1103 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/agr-decision-model https://media.thesatyajit.com/ai/thumbs/agr-decision-model.jpg?v=96c17f951243a43a 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/alodlm-looped-diffusion https://media.thesatyajit.com/ai/thumbs/alodlm-looped-diffusion.jpg?v=4bfed0101a3269aa 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/anyapi-scrape-router https://media.thesatyajit.com/ai/thumbs/anyapi-scrape-router.jpg?v=a1c37a7111e059eb 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/arc-cua https://media.thesatyajit.com/ai/thumbs/arc-cua.jpg?v=5b9bedebe68ea638 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/big-moe-one-3090 https://media.thesatyajit.com/ai/thumbs/big-moe-one-3090.jpg?v=750f370743e190cf 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/claude-demo-videos https://media.thesatyajit.com/ai/thumbs/claude-demo-videos.jpg?v=853e991c97da7f39 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/cloudflare-os https://media.thesatyajit.com/ai/thumbs/cloudflare-os.jpg?v=1b5c29c07887c11e 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/cpu-performance-engineering https://media.thesatyajit.com/ai/thumbs/cpu-performance-engineering.jpg?v=c43cfa568e9f39d6 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/efficient-gaussian-appearance https://media.thesatyajit.com/ai/thumbs/efficient-gaussian-appearance.jpg?v=e3d4686194ef2fa0 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/egocentric-video-benchmarks https://media.thesatyajit.com/ai/thumbs/egocentric-video-benchmarks.jpg?v=7e0d56ee0b17dd41 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/eva-empirical-vae https://media.thesatyajit.com/ai/thumbs/eva-empirical-vae.jpg?v=33f8069dc035ba81 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/freevideo-minimax-h3 https://media.thesatyajit.com/ai/thumbs/freevideo-minimax-h3.jpg?v=a358e3845ac516c6 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/gdp-ts https://media.thesatyajit.com/ai/thumbs/gdp-ts.jpg?v=673aa95bc49a264b 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/h-jepa https://media.thesatyajit.com/ai/thumbs/h-jepa.jpg?v=d4a15151ed40060b 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/hilti-slam-dataset https://media.thesatyajit.com/ai/thumbs/hilti-slam-dataset.jpg?v=baff76ebcde5d4f7 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/jepa-anything https://media.thesatyajit.com/ai/thumbs/jepa-anything.jpg?v=24a8466d7aa3ee3e 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/kolc-3dgs-bim https://media.thesatyajit.com/ai/thumbs/kolc-3dgs-bim.jpg?v=e2decc38f5d8eae2 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/kolibri-1 https://media.thesatyajit.com/ai/thumbs/kolibri-1.jpg?v=ac7d61d5b50cbeb2 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/leviathan-agent-index https://media.thesatyajit.com/ai/thumbs/leviathan-agent-index.jpg?v=31afc28401159728 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/lift-loop-flow-transformer https://media.thesatyajit.com/ai/thumbs/lift-loop-flow-transformer.jpg?v=077923ce3790f03a 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/linear-attention-state-roundup https://media.thesatyajit.com/ai/thumbs/linear-attention-state-roundup.jpg?v=f3d9fde5dad70dd7 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/llamacpp-decision-models https://media.thesatyajit.com/ai/thumbs/llamacpp-decision-models.jpg?v=8424d20328aa7770 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/llm-land-or-water https://media.thesatyajit.com/ai/thumbs/llm-land-or-water.jpg?v=8d1e3d44006d30c9 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/loom-looped-moe https://media.thesatyajit.com/ai/thumbs/loom-looped-moe.jpg?v=a825024f305cacfb 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/looped-contrastive-decoding https://media.thesatyajit.com/ai/thumbs/looped-contrastive-decoding.jpg?v=3ba383fa428ea39b 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/lot-diffusion https://media.thesatyajit.com/ai/thumbs/lot-diffusion.jpg?v=d10f5dcac0ea51b3 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/mach-2-additive https://media.thesatyajit.com/ai/thumbs/mach-2-additive.jpg?v=32a6aea4a4b7e56e 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/marin-535b https://media.thesatyajit.com/ai/thumbs/marin-535b.jpg?v=761ec52fb69c645c 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/meta-rebalancer https://media.thesatyajit.com/ai/thumbs/meta-rebalancer.jpg?v=b3b96feb97eb2960 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/multi-harness-rl https://media.thesatyajit.com/ai/thumbs/multi-harness-rl.jpg?v=4c118c95024ec4ac 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/nanogpt-speedrun-record https://media.thesatyajit.com/ai/thumbs/nanogpt-speedrun-record.jpg?v=f4d86d745e2ee010 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/octop https://media.thesatyajit.com/ai/thumbs/octop.jpg?v=be0ca70855968550 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/open-seo https://media.thesatyajit.com/ai/thumbs/open-seo.jpg?v=0c0c9226227cde8b 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/opencut https://media.thesatyajit.com/ai/thumbs/opencut.jpg?v=5ddea7e276e9e648 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/openhuman https://media.thesatyajit.com/ai/thumbs/openhuman.jpg?v=fd2180b8fd9e31f8 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/overmind https://media.thesatyajit.com/ai/thumbs/overmind.jpg?v=33caf8e48629a6dc 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/paper-rewriter-finetune https://media.thesatyajit.com/ai/thumbs/paper-rewriter-finetune.jpg?v=a780066fe8904804 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/pixelumm-pixeldense https://media.thesatyajit.com/ai/thumbs/pixelumm-pixeldense.jpg?v=0374441271ad747c 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/pixijs-3d-benchy https://media.thesatyajit.com/ai/thumbs/pixijs-3d-benchy.jpg?v=eefb4a97f045bcd9 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/pocket-tts-drifting https://media.thesatyajit.com/ai/thumbs/pocket-tts-drifting.jpg?v=f5709f47168241a0 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/portsimenv https://media.thesatyajit.com/ai/thumbs/portsimenv.jpg?v=f0590caddd98a61b 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/prime-inference https://media.thesatyajit.com/ai/thumbs/prime-inference.jpg?v=e9f3e40361d29c69 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/quantui-rs https://media.thesatyajit.com/ai/thumbs/quantui-rs.jpg?v=b62ecd231a8e91ad 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/qwen3guard-stream https://media.thesatyajit.com/ai/thumbs/qwen3guard-stream.jpg?v=72993882a5f23981 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/recursive-self-rewrite https://media.thesatyajit.com/ai/thumbs/recursive-self-rewrite.jpg?v=45a899e536a1be22 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/rwkv-7-g1k https://media.thesatyajit.com/ai/thumbs/rwkv-7-g1k.jpg?v=c5e8bd13997a39f1 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/sail-robot-trajectories https://media.thesatyajit.com/ai/thumbs/sail-robot-trajectories.jpg?v=613b4f1ebce84a5f 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/sentry-agent-recovery https://media.thesatyajit.com/ai/thumbs/sentry-agent-recovery.jpg?v=5be4c31cf9980060 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/shaders-webgpu https://media.thesatyajit.com/ai/thumbs/shaders-webgpu.jpg?v=3fc52f583686b849 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/spark-x2-5-4b https://media.thesatyajit.com/ai/thumbs/spark-x2-5-4b.jpg?v=6a9a432984a9728b 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/strata-engine https://media.thesatyajit.com/ai/thumbs/strata-engine.jpg?v=8445c54b33ee4f90 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/subquadratic-3sum-apsp https://media.thesatyajit.com/ai/thumbs/subquadratic-3sum-apsp.jpg?v=619e5b0b30b87a41 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/tiktok-5-6b-videos https://media.thesatyajit.com/ai/thumbs/tiktok-5-6b-videos.jpg?v=9b1165ed2890f20e 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/toks-tokenizer https://media.thesatyajit.com/ai/thumbs/toks-tokenizer.jpg?v=38a9fd5dd84752d6 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/treg https://media.thesatyajit.com/ai/thumbs/treg.jpg?v=78c2f3025a9f9511 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/whistle-stt https://media.thesatyajit.com/ai/thumbs/whistle-stt.jpg?v=eb0751acee597593 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-flash-next https://media.thesatyajit.com/ai/thumbs/qwen3-8-flash-next.jpg?v=6d3f588223acc10b Qwen3.8-Flash-Next: four changes and an honest report https://media.thesatyajit.com/ai/films/qwen3-8-flash-next-poster.webp?v=79f279564e1586e2 125B parameters plus 51B of N-gram embeddings, 6B active, trained for about a ninth of what its predecessor cost. https://media.thesatyajit.com/ai/films/qwen3-8-flash-next.mp4?v=72e570d450b1b586 115 2026-10-07 yes 2026-10-06 monthly 0.8 https://ai.satyajit.si/articles/dflash2 https://media.thesatyajit.com/ai/thumbs/dflash2.jpg?v=e10b53d7bcee87d6 2026-10-06 monthly 0.7 https://ai.satyajit.si/articles/context-language-models https://media.thesatyajit.com/ai/thumbs/context-language-models.jpg?v=8df4ea51e65b5139 Context Language Models: the bitter lesson comes for context management https://media.thesatyajit.com/ai/films/context-language-models-poster.webp?v=41ee7df98d7198ea A UW and Meta FAIR paper mirrors an agent's live context as an editable file and lets the model rewrite it with Bash. https://media.thesatyajit.com/ai/films/context-language-models.mp4?v=5aab466cab5280b6 102 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/da3-streaming-reconstruction https://media.thesatyajit.com/ai/thumbs/da3-streaming-reconstruction.jpg?v=8627fd3ecb488f30 Streaming 3D from unposed video: R³, AMB3R-SLAM, and the move to bounded memory https://media.thesatyajit.com/ai/films/da3-streaming-reconstruction-poster.webp?v=a26e52a390bfdf34 Two systems from this year build streaming 3D reconstruction on top of Depth-Anything-3 as a frozen backbone, and both turn per-frame pointmaps into an… https://media.thesatyajit.com/ai/films/da3-streaming-reconstruction.mp4?v=2946db8b52585e79 105 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/deepgemm-ascend https://media.thesatyajit.com/ai/thumbs/deepgemm-ascend.jpg?v=6c57e4d79cfc592e 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/depthbench https://media.thesatyajit.com/ai/thumbs/depthbench.jpg?v=4a9034dd4fa23c36 DepthBench: depth is a scaling axis only for the right residual stream https://media.thesatyajit.com/ai/films/depthbench-poster.webp?v=8b5e607900634c54 DepthBench holds the parameter count, the data and the recipe fixed and changes only the width-depth aspect ratio, across ten residual-stream designs on a… https://media.thesatyajit.com/ai/films/depthbench.mp4?v=079b710d022af34c 109 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/dsec-agent-sandbox https://media.thesatyajit.com/ai/thumbs/dsec-agent-sandbox.jpg?v=f0e82c30dd0f4364 DeepSeek DSec: three million agent sandboxes a day, 50× overcommitted https://media.thesatyajit.com/ai/films/dsec-agent-sandbox-poster.webp?v=af766d8517547d9b DeepSeek's report on DSec, the sandbox platform behind its agentic RL training from V3.2 to V4.1: four backends behind one Python SDK, environments composed… https://media.thesatyajit.com/ai/films/dsec-agent-sandbox.mp4?v=bb56778292b26ee7 112 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/far-lio https://media.thesatyajit.com/ai/thumbs/far-lio.jpg?v=3ca6046ce05409ab FAR-LIO: LiDAR-inertial odometry that holds together at 250 km/h https://media.thesatyajit.com/ai/films/far-lio-poster.webp?v=80bfd1c01b6a62ce FAR-LIO (TU Munich, arXiv 2606.26010, IROS 2026) is a CUDA LiDAR-inertial odometry stack built for autonomous racing: a GPU voxel hashmap with adaptive… https://media.thesatyajit.com/ai/films/far-lio.mp4?v=34809aaede8d1c6e 102 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/fframes https://media.thesatyajit.com/ai/thumbs/fframes.jpg?v=a0956dc9b36c9852 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/glm-5-3-cmp170hx https://media.thesatyajit.com/ai/thumbs/glm-5-3-cmp170hx.jpg?v=d68d77f546ef05a4 GLM-5.3-Flash on four mining cards: the hard part is an Ampere kernel https://media.thesatyajit.com/ai/films/glm-5-3-cmp170hx-poster.webp?v=ffd9ee14916c7369 A 320B hybrid-attention MoE with a 262K context, served on four NVIDIA CMP 170HX mining cards. https://media.thesatyajit.com/ai/films/glm-5-3-cmp170hx.mp4?v=4b46a51ef0f5d7a0 109 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/hindsight-agent-memory https://media.thesatyajit.com/ai/thumbs/hindsight-agent-memory.jpg?v=bc36268a5e111e85 Hindsight: +44 points on the same 20B backbone, from the memory layer https://media.thesatyajit.com/ai/films/hindsight-agent-memory-poster.webp?v=506e40a3ca5616ad Hindsight is an open-source (MIT) agent-memory system — about 44k GitHub stars — that organizes memory into four epistemically distinct networks (world… https://media.thesatyajit.com/ai/films/hindsight-agent-memory.mp4?v=42fe26a4981b525e 104 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/holo4 https://media.thesatyajit.com/ai/thumbs/holo4.jpg?v=4680bdf292db650e 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/index-translate https://media.thesatyajit.com/ai/thumbs/index-translate.jpg?v=823653e1e6d26113 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/iquest-q1 https://media.thesatyajit.com/ai/thumbs/iquest-q1.jpg?v=d78470965e4ce9b6 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/jev-alternatives-week-three https://media.thesatyajit.com/ai/thumbs/jev-alternatives-week-three.jpg?v=f713de171619e979 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/lattice-deduction-transformers https://media.thesatyajit.com/ai/thumbs/lattice-deduction-transformers.jpg?v=bb7b68d2f68ead12 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/longlive-plug https://media.thesatyajit.com/ai/thumbs/longlive-plug.jpg?v=482a944552efd881 LongLive-Plug: distill the speedup once, plug it into 54 video models https://media.thesatyajit.com/ai/films/longlive-plug-poster.webp?v=c9282ad35324be75 Specialized video-diffusion models each re-run their own distillation stage to get few-step sampling, single-pass classifier-free guidance, or long-context… https://media.thesatyajit.com/ai/films/longlive-plug.mp4?v=37a59b7b7bb00341 94 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/model-casting-sparse-ffn https://media.thesatyajit.com/ai/thumbs/model-casting-sparse-ffn.jpg?v=551dac4d3737b73c Model casting: compute the gate, then skip most of the FFN https://media.thesatyajit.com/ai/films/model-casting-sparse-ffn-poster.webp?v=99234f63ce2b08b5 A Meta FAIR and ENS mid-training recipe swaps a transformer's FFN activation to a ReLU/SiLU hybrid and adds an adaptive L1 penalty so ~90% of hidden units… https://media.thesatyajit.com/ai/films/model-casting-sparse-ffn.mp4?v=8250900ba2f5dad7 91 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/naive-n05-flash https://media.thesatyajit.com/ai/thumbs/naive-n05-flash.jpg?v=d37d0c7e44b31bbb Naive-N0.5-Flash: no full-attention layers, and a runtime built by AI https://media.thesatyajit.com/ai/films/naive-n05-flash-poster.webp?v=2f7c5afde1276ce8 A 309B-total, 15.5B-active open (MIT) MoE with native 1M context whose every attention layer is either a 128-token sliding window or a top-2,048 DeepSeek… https://media.thesatyajit.com/ai/films/naive-n05-flash.mp4?v=5ba0eb780817c2f8 102 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/olmo-core-3 https://media.thesatyajit.com/ai/thumbs/olmo-core-3.jpg?v=299a36408473c649 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/openshell https://media.thesatyajit.com/ai/thumbs/openshell.jpg?v=b365cda370eb0ce5 OpenShell: give the agent a real shell, keep the boundary in the kernel https://media.thesatyajit.com/ai/films/openshell-poster.webp?v=4d88bbb433a2a567 NVIDIA open-sourced OpenShell (Apache-2.0, Rust): a sandbox that gives an AI agent real files, real network, and real credentials in play, but moves the… https://media.thesatyajit.com/ai/films/openshell.mp4?v=6eb1af89e2c71f83 106 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/oren-sdf https://media.thesatyajit.com/ai/thumbs/oren-sdf.jpg?v=44b3c53a49bea141 OREN: real-time Euclidean SDF, because the octree carries the prior https://media.thesatyajit.com/ai/films/oren-sdf-poster.webp?v=f2e9f92cf00a715b OREN (UCSD Existential Robotics Lab, arXiv 2510.18999, IROS 2026) reconstructs a continuous, non-truncated signed distance field online from streaming point… https://media.thesatyajit.com/ai/films/oren-sdf.mp4?v=e12d76a6c672298a 110 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/pruna-superpoint https://media.thesatyajit.com/ai/thumbs/pruna-superpoint.jpg?v=5cfe44b0abb4859d 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/ptxbench-kda-kernels https://media.thesatyajit.com/ai/thumbs/ptxbench-kda-kernels.jpg?v=704e3bab04a5b758 AI-written GPU kernels: PTXBench grades them, and an agent gamed the grade with the constant 0.1778209953 https://media.thesatyajit.com/ai/films/ptxbench-kda-kernels-poster.webp?v=248dff179e6abe4a Two 2026 results ask whether a language model can write a frontier GPU kernel. PTXBench makes models write CUDA with inline PTX from scratch, vendor… https://media.thesatyajit.com/ai/films/ptxbench-kda-kernels.mp4?v=dbbccb816e913d37 100 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/recursivemas https://media.thesatyajit.com/ai/thumbs/recursivemas.jpg?v=b4dba7289122ff6c RecursiveMAS: a multi-agent system folded into one looped transformer https://media.thesatyajit.com/ai/films/recursivemas-poster.webp?v=7363b8fb6af6a4b5 A looped transformer scales reasoning by reusing its own layers in latent space. https://media.thesatyajit.com/ai/films/recursivemas.mp4?v=2f8850464fc1f86d 109 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/rrsi-harness-evolution https://media.thesatyajit.com/ai/thumbs/rrsi-harness-evolution.jpg?v=bee8690212a9e1ba 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/siamjepa https://media.thesatyajit.com/ai/thumbs/siamjepa.jpg?v=482fd359cc00cb1d SiamJEPA: shuffle the teacher, and a patch has to say what it is https://media.thesatyajit.com/ai/films/siamjepa-poster.webp?v=ef43e3fcbb28bbdb SiamJEPA runs two masked Siamese student encoders against an EMA teacher, aligned by a KL term, in the I-JEPA lineage of predicting latent embeddings of… https://media.thesatyajit.com/ai/films/siamjepa.mp4?v=09a90d9677dcf701 110 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/sog-splat-format https://media.thesatyajit.com/ai/thumbs/sog-splat-format.jpg?v=60745316fbc4d0b1 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/surflo https://media.thesatyajit.com/ai/thumbs/surflo.jpg?v=3c873f778f336d85 Surflo: one latent for a scene, a surface at any resolution https://media.thesatyajit.com/ai/films/surflo-poster.webp?v=43376b2af43f89a5 Surflo (Guédon et al., arXiv 2606.13644, listed NeurIPS 2026 Oral on the project page) compresses a variable number of unposed RGB views into K=128 latent… https://media.thesatyajit.com/ai/films/surflo.mp4?v=e04ab9590e99fd8c 100 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/thomson-n7-lean https://media.thesatyajit.com/ai/thumbs/thomson-n7-lean.jpg?v=cfd74063b55e804c Thomson N=7: ten agents, 17,895 lines of Lean, one trusted kernel https://media.thesatyajit.com/ai/films/thomson-n7-lean-poster.webp?v=600e12ef097fe217 Seven electrons on a sphere settle into a pentagonal bipyramid — the one small Thomson case that had no proof at all. https://media.thesatyajit.com/ai/films/thomson-n7-lean.mp4?v=1a612beb3fb2b78a 93 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/jevgrep https://media.thesatyajit.com/ai/thumbs/jevgrep.jpg?v=d39cd92955ed0f7e 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/runtime-dynamic-compression https://media.thesatyajit.com/ai/thumbs/runtime-dynamic-compression.jpg?v=9800926ad2dd68a8 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/tensorfold-strata-engines https://media.thesatyajit.com/ai/thumbs/tensorfold-strata-engines.jpg?v=ded48dcd3ac2a918 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/k2-horizon-mova https://media.thesatyajit.com/ai/thumbs/k2-horizon-mova.jpg?v=46e4270faf6237f1 K2-Horizon-MoVA-36B: routing the value vectors, and why the KV cache does not notice https://media.thesatyajit.com/ai/films/k2-horizon-mova-poster.webp?v=7a36706d94b00a82 IFM's K2-Horizon-MoVA-36B-A4B replaces attention's v_proj with a router over 64 value experts, top-4 per token. https://media.thesatyajit.com/ai/films/k2-horizon-mova.mp4?v=77ce6f562ebf6364 110 2026-10-07 yes 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/qwen-image-2-1 https://media.thesatyajit.com/ai/thumbs/qwen-image-2-1.jpg?v=72cd1073c9f685b9 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/tabfm https://media.thesatyajit.com/ai/thumbs/tabfm.jpg?v=5fa9084be302557c Google's TabFM: 98.8% of the weights never see a cell https://media.thesatyajit.com/ai/films/tabfm-poster.webp?v=a26c9dba876e68eb Google's TabFM is 1,639,444,522 parameters, and 1,619,925,002 of them are an ordinary 24-block transformer reading one 2048-d vector per row. https://media.thesatyajit.com/ai/films/tabfm.mp4?v=dc4ba2a6fe8f80bc 99 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/minicpm5-2b https://media.thesatyajit.com/ai/thumbs/minicpm5-2b.jpg?v=94eacd9dfc34c8e0 MiniCPM5-2B: how 2 KV heads pay for a 131K window https://media.thesatyajit.com/ai/films/minicpm5-2b-poster.webp?v=7354cfcb185ffdcb MiniCPM5-2B ships with no tech report of its own — the model repo's only arXiv tags are MiniCPM4's paper and a data-tiering paper, and the closing BibTeX… https://media.thesatyajit.com/ai/films/minicpm5-2b.mp4?v=8be59ea452868ef5 107 2026-10-07 yes 2026-10-02 monthly 0.8 https://ai.satyajit.si/articles/nanogpt-speedrun-frontier https://media.thesatyajit.com/ai/thumbs/nanogpt-speedrun-frontier.jpg?v=01cba3ca84c3572e 2026-10-02 monthly 0.7 https://ai.satyajit.si/articles/functional-gradient-descent https://media.thesatyajit.com/ai/thumbs/functional-gradient-descent.jpg?v=013a4c6d5f41c5d5 2026-09-27 monthly 0.7 https://ai.satyajit.si/articles/julia-1-decision-model https://media.thesatyajit.com/ai/thumbs/julia-1-decision-model.jpg?v=9847f22b8d9da825 2026-09-27 monthly 0.7 https://ai.satyajit.si/articles/qwen-image-2-1-outpaint-lora https://media.thesatyajit.com/ai/thumbs/qwen-image-2-1-outpaint-lora.jpg?v=2116fcc00a4b669d 2026-09-27 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-27b-on-12gb https://media.thesatyajit.com/ai/thumbs/qwen3-8-27b-on-12gb.jpg?v=2857b703cbc608d9 2026-09-27 monthly 0.7 https://ai.satyajit.si/articles/stable-diffusion-cpp https://media.thesatyajit.com/ai/thumbs/stable-diffusion-cpp.jpg?v=6d90efa38e493dcc 2026-09-27 monthly 0.7 https://ai.satyajit.si/articles/3d-reconstruction-roundup https://media.thesatyajit.com/ai/thumbs/3d-reconstruction-roundup.jpg?v=0eb6c2923777cbdf 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/btl-interference-search https://media.thesatyajit.com/ai/thumbs/btl-interference-search.jpg?v=61f43bfa4181a93e 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/code2skill https://media.thesatyajit.com/ai/thumbs/code2skill.jpg?v=30dc8dac87da9869 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/contrastive-language-model https://media.thesatyajit.com/ai/thumbs/contrastive-language-model.jpg?v=eb5372c95b192503 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/differential-transformer https://media.thesatyajit.com/ai/thumbs/differential-transformer.jpg?v=98e221969abd4e55 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/fish-audio-s2 https://media.thesatyajit.com/ai/thumbs/fish-audio-s2.jpg?v=8b024d3a1d9c3e87 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/flash-dllm https://media.thesatyajit.com/ai/thumbs/flash-dllm.jpg?v=f03e58a2fa1e3583 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/flux-3-action https://media.thesatyajit.com/ai/thumbs/flux-3-action.jpg?v=81b52e92b9379bf3 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/gliner-2-5-decide https://media.thesatyajit.com/ai/thumbs/gliner-2-5-decide.jpg?v=9bff433202315fe0 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/gtsam-4-3 https://media.thesatyajit.com/ai/thumbs/gtsam-4-3.jpg?v=05c6a08e884a0b24 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/hunyuan-a13b https://media.thesatyajit.com/ai/thumbs/hunyuan-a13b.jpg?v=a27fcea0c7636e1a 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/jev-alternatives-week-two https://media.thesatyajit.com/ai/thumbs/jev-alternatives-week-two.jpg?v=8ba482ce8d09fd69 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/kimi-k3-tpu-megakernel https://media.thesatyajit.com/ai/thumbs/kimi-k3-tpu-megakernel.jpg?v=da5a4b6c9faad0af 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/lensvlm-9b https://media.thesatyajit.com/ai/thumbs/lensvlm-9b.jpg?v=e46cbb56d92dc0c8 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/llm-compressor-0-14 https://media.thesatyajit.com/ai/thumbs/llm-compressor-0-14.jpg?v=1b49f1c2581165d8 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/male-fly-connectome https://media.thesatyajit.com/ai/thumbs/male-fly-connectome.jpg?v=1e5c1072a9da1007 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/matryoshka-attribution https://media.thesatyajit.com/ai/thumbs/matryoshka-attribution.jpg?v=21352f381a605f8d 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/nemotron-3-diarization https://media.thesatyajit.com/ai/thumbs/nemotron-3-diarization.jpg?v=6c7dcebd83752543 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/nvidia-model-optimizer https://media.thesatyajit.com/ai/thumbs/nvidia-model-optimizer.jpg?v=6d307868ec7f5262 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/one-shot-launch-videos https://media.thesatyajit.com/ai/thumbs/one-shot-launch-videos.jpg?v=1851742ca1d76615 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/opendlss-nr https://media.thesatyajit.com/ai/thumbs/opendlss-nr.jpg?v=a58d61d219eda5b1 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/ovis-embedding https://media.thesatyajit.com/ai/thumbs/ovis-embedding.jpg?v=6d62ad56005a5952 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/pmndrs-math-typegpu https://media.thesatyajit.com/ai/thumbs/pmndrs-math-typegpu.jpg?v=18881c107a39fb2a 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/qwen-audio-3-1 https://media.thesatyajit.com/ai/thumbs/qwen-audio-3-1.jpg?v=b83a0507da0ef280 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/qwen-image-2-1-few-step https://media.thesatyajit.com/ai/thumbs/qwen-image-2-1-few-step.jpg?v=c4db7c6954a430ba 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/qwen-intelligence-planner-agent https://media.thesatyajit.com/ai/thumbs/qwen-intelligence-planner-agent.jpg?v=3414f3d75c32f299 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-27b-variants https://media.thesatyajit.com/ai/thumbs/qwen3-8-27b-variants.jpg?v=fbd81e103a2a442b 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/rigel-hybrid-mamba https://media.thesatyajit.com/ai/thumbs/rigel-hybrid-mamba.jpg?v=e5890ffd26cebcbe 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/siglip-2-coreml https://media.thesatyajit.com/ai/thumbs/siglip-2-coreml.jpg?v=4e55280800f1fd7e 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/surfslam https://media.thesatyajit.com/ai/thumbs/surfslam.jpg?v=1042015cb010a0eb 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/teleocr https://media.thesatyajit.com/ai/thumbs/teleocr.jpg?v=da3c65e651527a0e 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/tokenizers-v1-combinatorial-bpe https://media.thesatyajit.com/ai/thumbs/tokenizers-v1-combinatorial-bpe.jpg?v=7707c0b6176dfa91 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/worldcrafter https://media.thesatyajit.com/ai/thumbs/worldcrafter.jpg?v=71e224d64af8e040 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/xhc-residual-connections https://media.thesatyajit.com/ai/thumbs/xhc-residual-connections.jpg?v=2601aec0cb6cf23c 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/qwen-image-2-1-pocket-rewriter https://media.thesatyajit.com/ai/thumbs/qwen-image-2-1-pocket-rewriter.jpg?v=bb8d3fa5b9b67e5b Qwen-Image-2.1 Pocket rewriter: 12x smaller than the 9B, and it made all four test images less faithful to the prompt https://media.thesatyajit.com/ai/films/qwen-image-2-1-pocket-rewriter-poster.webp?v=f15c26da11ce5706 ML-Intern-lab distilled Qwen-Image-2.1's recommended 9B prompt rewriter into a 0.8B and a 2B that answer with no system prompt and no thinking. https://media.thesatyajit.com/ai/films/qwen-image-2-1-pocket-rewriter.mp4?v=ef5d28fc8bf8e742 119 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/mimo-v2-6 https://media.thesatyajit.com/ai/thumbs/mimo-v2-6.jpg?v=7038508786f6e5a9 MiMo-V2.6: a trillion parameters, thirty RL steps, and a 20× that is a decoder https://media.thesatyajit.com/ai/films/mimo-v2-6-poster.webp?v=203eb5affe5c5e94 Xiaomi open-weighted MiMo-V2.6-Pro and Flash. I counted the safetensors: Pro is 1,024.2B stored parameters with 41.9B active, released natively in MXFP4 at… https://media.thesatyajit.com/ai/films/mimo-v2-6.mp4?v=bfdff6ae01efd0e5 117 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/articles/reproducing-jev https://media.thesatyajit.com/ai/thumbs/reproducing-jev.jpg?v=0a29d18a54e41274 Reproducing Jev, from a source I could not find https://media.thesatyajit.com/ai/films/reproducing-jev-poster.webp?v=0b1d9377390662f1 Someone handed me five findings from an independent team's multi-day attempt to reproduce, break and improve Jev — a 27-point regression, a 0.769-to-0.541… https://media.thesatyajit.com/ai/films/reproducing-jev.mp4?v=7ae0dd6125b1f098 114 2026-10-07 yes 2026-09-26 monthly 0.8 https://ai.satyajit.si/articles/attention-mechanisms https://media.thesatyajit.com/ai/thumbs/attention-mechanisms.jpg?v=4480f368513d3a12 A field guide to attention mechanisms https://media.thesatyajit.com/ai/films/attention-mechanisms-poster.webp?v=4f72c0054c3dbc76 One operation, two bills — O(N²) compute and the O(N·layers) KV cache — and a map of how every variant pays them down: MHA, MQA, GQA and MLA on the memory… https://media.thesatyajit.com/ai/films/attention-mechanisms.mp4?v=a35f9b990bbd2545 107 2026-10-07 yes 2026-09-26 monthly 0.8 https://ai.satyajit.si/articles/voxel-musou https://media.thesatyajit.com/ai/thumbs/voxel-musou.jpg?v=2397c98eb8b4c333 Voxel Musou: the army is a 300-slot pool, and the combo is a frame table https://media.thesatyajit.com/ai/films/voxel-musou-poster.webp?v=bc0374fa88904326 A browser-playable Dynasty Warriors homage — Zhao Yun against a field of voxel Wei soldiers — in plain ES modules on a vendored three.js r186, created on 24… https://media.thesatyajit.com/ai/films/voxel-musou.mp4?v=9909cacde7d99b0a 112 2026-10-07 yes 2026-09-24 monthly 0.7 https://ai.satyajit.si/articles/ming-image-0-1-design https://media.thesatyajit.com/ai/thumbs/ming-image-0-1-design.jpg?v=e0e18b9ccd9355b9 Ming-Image-0.1-Design: a Z-Image fine-tune that is #1 of 42 open models and 16th of 129 https://media.thesatyajit.com/ai/films/ming-image-0-1-design-poster.webp?v=87c3a1668e779390 inclusionAI open-sourced two '6B' design models and two agent skills, and announced the first as #1 among open-weight models on Artificial Analysis's UI/UX… https://media.thesatyajit.com/ai/films/ming-image-0-1-design.mp4?v=a71941c1b5a07c39 114 2026-10-07 yes 2026-09-23 monthly 0.7 https://ai.satyajit.si/articles/adapt-1-machina https://media.thesatyajit.com/ai/thumbs/adapt-1-machina.jpg?v=cf7f04f7307d44fa Adapt-1 Machina: no critic, because the simulator is the critic https://media.thesatyajit.com/ai/films/adapt-1-machina-poster.webp?v=84aa8aae5a2bddb1 Rei Labs claims continuous control learned from coarse outcome feedback with no demonstrations and no critic. https://media.thesatyajit.com/ai/films/adapt-1-machina.mp4?v=5dce6e6468f6f594 118 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/agent-beacon https://media.thesatyajit.com/ai/thumbs/agent-beacon.jpg?v=c58837695c396e04 Beacon: Jev grades your agent runs, and the lesson is not in the skill https://media.thesatyajit.com/ai/films/agent-beacon-poster.webp?v=d01a1ad573d6ce33 Beacon captures agent sessions across 29 harnesses and asks a decision model which runs are worth learning from. https://media.thesatyajit.com/ai/films/agent-beacon.mp4?v=11f3bc2115b58aaa 114 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/agent-jev-06b https://media.thesatyajit.com/ai/thumbs/agent-jev-06b.jpg?v=18def465c55a7312 AgentJev-0.6B: exact order invariance, and options that still interact https://media.thesatyajit.com/ai/films/agent-jev-06b-poster.webp?v=99dcdb3f63e28f3f Four articles on this site have described a fork: isolate a decision model's options and you get exact order-invariance but no option can refer to another… https://media.thesatyajit.com/ai/films/agent-jev-06b.mp4?v=2261e979f0311fe4 113 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/altar-1 https://media.thesatyajit.com/ai/thumbs/altar-1.jpg?v=1971f633a521c2b2 Altar-1: a security model measured on recall alone https://media.thesatyajit.com/ai/films/altar-1-poster.webp?v=b877c6ab7f177a2c Aikido's first open-weight security model is GLM-5.3 with 88 of its 256 routed experts deleted and the rest at INT4 — 500.8B parameters in 327.9 GB… https://media.thesatyajit.com/ai/films/altar-1.mp4?v=60bc8954cd8b8f12 118 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/carveout-3dgs-object-labels https://media.thesatyajit.com/ai/thumbs/carveout-3dgs-object-labels.jpg?v=4b98ab4f195e3a8a Carveout: SAM 3 labels for a Gaussian-splat scene, without the multiplex https://media.thesatyajit.com/ai/films/carveout-3dgs-object-labels-poster.webp?v=33e004c48b0e707f Carveout renders synthetic views of a trained 3D Gaussian Splatting scene, segments them with SAM 3 Promptable Concept Segmentation, and lifts the 2D masks… https://media.thesatyajit.com/ai/films/carveout-3dgs-object-labels.mp4?v=5edbfaabe0b48bb2 115 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/cinference https://media.thesatyajit.com/ai/thumbs/cinference.jpg?v=a1d7b2c01b751ddd Cinference runs Qwen3.8-27B at 450 tokens a second: what the number counts https://media.thesatyajit.com/ai/films/cinference-poster.webp?v=790088fccf92e70b A one-command installer claims 450 TPS on a custom engine and above 300 at 256K context. https://media.thesatyajit.com/ai/films/cinference.mp4?v=4b27a59a62f0e0ce 112 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/figures4papers https://media.thesatyajit.com/ai/thumbs/figures4papers.jpg?v=aa38f7f27cc76274 figures4papers: a house style, audited against its own figures https://media.thesatyajit.com/ai/films/figures4papers-poster.webp?v=b5eb6dbe46545ab9 A Yale PhD student open-sourced every plotting script from eight papers, with the data and the generated figures committed beside them, and packaged the… https://media.thesatyajit.com/ai/films/figures4papers.mp4?v=2921daf1260497a8 111 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/gae-geometry-native https://media.thesatyajit.com/ai/thumbs/gae-geometry-native.jpg?v=406499abce49d17d TencentARC's GAE: the geometry decoder never sees the picture https://media.thesatyajit.com/ai/films/gae-geometry-native-poster.webp?v=4e2dc34fa49ea2f4 TencentARC's GAE compresses Depth Anything 3's four-level feature hierarchy into a 64-channel latent, runs one conditional flow model in that space, and… https://media.thesatyajit.com/ai/films/gae-geometry-native.mp4?v=61a5dee5c010f259 119 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/google-ax https://media.thesatyajit.com/ai/thumbs/google-ax.jpg?v=89f8f07d18758795 Google's ax agent runtime: the single writer is a Go map https://media.thesatyajit.com/ai/films/google-ax-poster.webp?v=9a908d7bf105b19f Google open-sourced ax, a distributed agent runtime, on the strength of two design claims: a single-writer architecture for consistent session state, and a… https://media.thesatyajit.com/ai/films/google-ax.mp4?v=b02f1759a752ee36 116 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/halo https://media.thesatyajit.com/ai/thumbs/halo.jpg?v=2bf929ce36b79b9c Halo's 2.8× over TRL is not a kernel, and its own benchmark proves it https://media.thesatyajit.com/ai/films/halo-poster.webp?v=adf5700e1eefbf01 A post-training framework claiming up to 2.8× stock TRL's throughput at less peak memory, on native HuggingFace checkpoints. https://media.thesatyajit.com/ai/films/halo.mp4?v=5fc5f95445db69b3 115 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/husky-msi https://media.thesatyajit.com/ai/thumbs/husky-msi.jpg?v=0df0ced555bf7331 Husky's 4.5× over MLX: the method is published, and it is speculative decoding https://media.thesatyajit.com/ai/films/husky-msi-poster.webp?v=c4ca41e4efa2ca4c Two days after a 50× that turned out to be 1.39×, a second Apple-silicon engine claims 4.5× over MLX and 730 tokens a second — and this one ships the table. https://media.thesatyajit.com/ai/films/husky-msi.mp4?v=fb8f2f1bb7766656 108 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/jev-engineering-swarms https://media.thesatyajit.com/ai/thumbs/jev-engineering-swarms.jpg?v=b4330418c491635a Jev swarms: one call, ten agents, and the answer moves https://media.thesatyajit.com/ai/films/jev-engineering-swarms-poster.webp?v=185ea9a98cab22f6 The Jev Engineering pitch has grown a swarm variant, and the swarm variant contains the first genuinely new argument in the family: batch every agent's… https://media.thesatyajit.com/ai/films/jev-engineering-swarms.mp4?v=25badd747b305939 106 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/jev-is-not-deterministic https://media.thesatyajit.com/ai/thumbs/jev-is-not-deterministic.jpg?v=e37bdb74adfc8a37 Jev is not deterministic, and option order is not noise https://media.thesatyajit.com/ai/films/jev-is-not-deterministic-poster.webp?v=cf658a8dd37d0c0f A public complaint about Jev packs three assertions into four sentences: it is not deterministic, the option order drastically changes the probabilities… https://media.thesatyajit.com/ai/films/jev-is-not-deterministic.mp4?v=3cc1185878507bb3 115 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/jev-omni https://media.thesatyajit.com/ai/thumbs/jev-omni.jpg?v=432a9dbc5ef38012 Jev-Omni: a 256-way classifier with no encoders https://media.thesatyajit.com/ai/films/jev-omni-poster.webp?v=1c7073b6fae94ef5 The first System One model that takes text, images, audio and video. Read against its own artifacts rather than its announcement: the multimodal part is… https://media.thesatyajit.com/ai/films/jev-omni.mp4?v=18be8e5ba209a379 112 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/kev https://media.thesatyajit.com/ai/thumbs/kev.jpg?v=664bba005f11713e Kev, and the price of not being able to see the other options https://media.thesatyajit.com/ai/films/kev-poster.webp?v=1d86d00d8cfcfd38 Jared Palmer refactored Kev onto Qwen3.5 and shipped 0.8B, 4B and 9B checkpoints with a fine-tuning script. https://media.thesatyajit.com/ai/films/kev.mp4?v=18b2e5f8c469a3dc 109 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/mini-agi https://media.thesatyajit.com/ai/thumbs/mini-agi.jpg?v=396f5926fe265689 mini-AGI: what the 169 experts and the 0.0067 nats actually are https://media.thesatyajit.com/ai/films/mini-agi-poster.webp?v=667dee74280049d7 A byte-level model that trains from scratch on one 8 GB laptop GPU and keeps learning while it reads, three days old, no weights published, 4.5% of the way… https://media.thesatyajit.com/ai/films/mini-agi.mp4?v=7dbf000bbd7db36f 108 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/modular-llm-inference-handbook https://media.thesatyajit.com/ai/thumbs/modular-llm-inference-handbook.jpg?v=723e8d9ac71799d6 Modular's LLM Inference Handbook: right formula, wrong variable https://media.thesatyajit.com/ai/films/modular-llm-inference-handbook-poster.webp?v=80e0afe5224b7dd5 Modular's LLM Inference Handbook is 70,000 words across 57 pages, and it is better than its genre deserves — the metrics chapter makes a distinction most… https://media.thesatyajit.com/ai/films/modular-llm-inference-handbook.mp4?v=ab806862e352bf70 121 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/needle-3-finetune https://media.thesatyajit.com/ai/thumbs/needle-3-finetune.jpg?v=b75e0297fccb199a Cactus Needle 3: the free fine-tune deletes the confidence head https://media.thesatyajit.com/ai/films/needle-3-finetune-poster.webp?v=9db7bc40f4ef1a66 Cactus ships two ways to customise Needle 3. The local one is real, runs on your laptop and writes a deployable binary. https://media.thesatyajit.com/ai/films/needle-3-finetune.mp4?v=d347886b202f47b9 109 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/nvfp4-kv-cache https://media.thesatyajit.com/ai/thumbs/nvfp4-kv-cache.jpg?v=cc543402a09bdcdc SGLang's NVFP4 KV cache: the 56% is exact, the 1.78x is untested https://media.thesatyajit.com/ai/films/nvfp4-kv-cache-poster.webp?v=9994cefa758574e1 NVFP4 KV lands in SGLang behind one flag and the arithmetic checks out perfectly — 56.25% of FP8 already includes the block scale, and the ratio is tight… https://media.thesatyajit.com/ai/films/nvfp4-kv-cache.mp4?v=7194fee0ee4f307e 121 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/openmuse https://media.thesatyajit.com/ai/thumbs/openmuse.jpg?v=9795ddda8d3c65a8 OpenMuse: "compatible with any agent harness" is one ternary branch https://media.thesatyajit.com/ai/films/openmuse-poster.webp?v=50f12138bb7ec8da CopilotKit's OpenMuse is an MIT, self-hostable personal agent whose README leads with harness portability. I read the repository at commit c79178f. https://media.thesatyajit.com/ai/films/openmuse.mp4?v=e9b8b4e94615461d 112 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/penjing-27b https://media.thesatyajit.com/ai/thumbs/penjing-27b.jpg?v=5b0962bc9e25887d Penjing-27B: reading a quantisation claim out of the GGUF header https://media.thesatyajit.com/ai/films/penjing-27b-poster.webp?v=ac27d35b6219f84c PollardWeights shipped a six-rung ladder of Qwen3.8-27B quants under the banner 'bits placed by per-layer sensitivity, not a uniform crush.' https://media.thesatyajit.com/ai/films/penjing-27b.mp4?v=11692d3038cbc853 116 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/qwen38-pd-serving https://media.thesatyajit.com/ai/thumbs/qwen38-pd-serving.jpg?v=50d114ad35dce4a1 vLLM serving Qwen3.8-2.4T: a Pareto frontier is not a deployment https://media.thesatyajit.com/ai/films/qwen38-pd-serving-poster.webp?v=e42654da5cf9ad75 vLLM published 5,000 total tokens/s/GPU and 180 output tokens/s/user for a 2.4T model on GB300 NVL72, and the two numbers belong to different machines — 48… https://media.thesatyajit.com/ai/films/qwen38-pd-serving.mp4?v=01459095f7bd491b 117 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/scaffolding-gets-eaten https://media.thesatyajit.com/ai/thumbs/scaffolding-gets-eaten.jpg?v=86abd7beec292cbe Thinking in Blender: which half of the 3D scaffolding gets eaten https://media.thesatyajit.com/ai/films/scaffolding-gets-eaten-poster.webp?v=71794480be218d72 The authors of a staged 3D coding harness went back to it with a stronger base model, found the gains almost gone, and warned that all of today's… https://media.thesatyajit.com/ai/films/scaffolding-gets-eaten.mp4?v=505018833633cfbf 115 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/skill2env https://media.thesatyajit.com/ai/thumbs/skill2env.jpg?v=9dc55a6a1e2fed82 NVIDIA's Skill2Env: collective is a distribution claim, and the benchmark disagrees https://media.thesatyajit.com/ai/films/skill2env-poster.webp?v=4824c2e8ee61e719 NVIDIA's Skill2Env compiles 3.4k public Agent Skills into 7,971 containerised RL environments, each with programmatic tests and a rubric distilled from the… https://media.thesatyajit.com/ai/films/skill2env.mp4?v=0ccdf1ab7be4e07e 117 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/tinfield-1 https://media.thesatyajit.com/ai/thumbs/tinfield-1.jpg?v=0b15a05ce317d294 Tinfield 1: 1.85% of a model, and a 51-billion-parameter hash table https://media.thesatyajit.com/ai/films/tinfield-1-poster.webp?v=61d5c9455ad2a79f Bad Theory Labs shipped a 360 GB checkpoint claiming 33.0 on Terminal-Bench 4.0 and 62.0 on DeepSWE v1.1. https://media.thesatyajit.com/ai/films/tinfield-1.mp4?v=fc72ff4963c6609d 117 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/voxel-tto https://media.thesatyajit.com/ai/thumbs/voxel-tto.jpg?v=408f4a850121c782 VoxelTTO spends 93% of a scene fixing the poses https://media.thesatyajit.com/ai/films/voxel-tto-poster.webp?v=00b7db52fc68b8f5 Feed-forward 3D Gaussian Splatting degrades when the poses drift, and the damage is invisible on input views — it only shows on the views you did not hand… https://media.thesatyajit.com/ai/films/voxel-tto.mp4?v=1f4f5177fe8a01ab 114 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/where-to-use-jev https://media.thesatyajit.com/ai/thumbs/where-to-use-jev.jpg?v=79ebe2ba6b489adc Where to use Jev: the test is whether one option can read another https://media.thesatyajit.com/ai/films/where-to-use-jev-poster.webp?v=e58c864f955a2b6a Eighteen pieces on this site have measured decision models in somebody else's production code. https://media.thesatyajit.com/ai/films/where-to-use-jev.mp4?v=56ef19b13ed704c4 106 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/xgen-jing-control-axes https://media.thesatyajit.com/ai/thumbs/xgen-jing-control-axes.jpg?v=e74210efc10d825d XGEN-JING wired up two more control axes, four days later https://media.thesatyajit.com/ai/films/xgen-jing-control-axes-poster.webp?v=621b09cc81b6e165 Four days ago this site read XGEN-JING's control parser and found that four of its six camera-control dimensions were hard-coded zero. https://media.thesatyajit.com/ai/films/xgen-jing-control-axes.mp4?v=c97afeab93c64e07 117 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/xgrammar-structural-tags https://media.thesatyajit.com/ai/thumbs/xgrammar-structural-tags.jpg?v=d7a4744999bd3b26 XGrammar structural tags: a grammar guarantees it parses, a menu guarantees it was on the menu https://media.thesatyajit.com/ai/films/xgrammar-structural-tags-poster.webp?v=b0a4b9fb73e304dc XGrammar-2's Structural Tags bring strict tool calling to SGLang and vLLM, enforcing tool-argument schemas through DeepSeek's own XML wire format rather… https://media.thesatyajit.com/ai/films/xgrammar-structural-tags.mp4?v=72793457ce5855d4 111 2026-10-07 yes 2026-09-22 monthly 0.8 https://ai.satyajit.si/articles/zcode-disclosure https://media.thesatyajit.com/ai/thumbs/zcode-disclosure.jpg?v=ca2bcba4a339cf10 ZCode, open-sourced: what a remediation can and cannot establish https://media.thesatyajit.com/ai/films/zcode-disclosure-poster.webp?v=e5a774c38a49aa40 Z.ai has published the ZCode client under Apache-2.0 after community reports that it generated and uploaded snapshots of local repositories. https://media.thesatyajit.com/ai/films/zcode-disclosure.mp4?v=7d51f686ad44d54a 93 2026-10-07 yes 2026-09-22 monthly 0.7 https://ai.satyajit.si/articles/laya-vs-jev https://media.thesatyajit.com/ai/thumbs/laya-vs-jev.jpg?v=c2af2a1250617759 Laya vs Jev: the slower model gets an easier question https://media.thesatyajit.com/ai/films/laya-vs-jev-poster.webp?v=b3d1aabe8e40731b A repo puts a local MLX model and a hosted API on the same T-Rex course and races them. https://media.thesatyajit.com/ai/films/laya-vs-jev.mp4?v=38a4aced793c18a0 114 2026-10-07 yes 2026-09-21 monthly 0.8 https://ai.satyajit.si/articles/jev-linkmap https://media.thesatyajit.com/ai/thumbs/jev-linkmap.jpg?v=8f426dec6b43e436 jev-linkmap: the rubric rewrite was mostly a threshold https://media.thesatyajit.com/ai/films/jev-linkmap-poster.webp?v=d1c21f772f4fd984 jev-linkmap rebuilt a 566-page site's internal link map in 5.9 seconds for $0.27, and the number I keep hearing quoted is 240x cheaper than Opus. https://media.thesatyajit.com/ai/films/jev-linkmap.mp4?v=85b771e1b5ae763d 97 2026-10-07 yes 2026-09-20 monthly 0.8 https://ai.satyajit.si/articles/laya-mlx https://media.thesatyajit.com/ai/thumbs/laya-mlx.jpg?v=eefd1586f7b767fb Laya on Apple silicon: 1.39×, not 50×, and K is a literal again https://media.thesatyajit.com/ai/films/laya-mlx-poster.webp?v=035adba6ad573dfc Yesterday's browser piece hit a wall — no GPU in the machine, every WebGPU number refused. https://media.thesatyajit.com/ai/films/laya-mlx.mp4?v=2e59ceda837b9cc4 97 2026-10-07 yes 2026-09-20 monthly 0.8 https://ai.satyajit.si/articles/any-model-can-be-jev https://media.thesatyajit.com/ai/thumbs/any-model-can-be-jev.jpg?v=8e3fdfc8f8b05e25 Any model can be Jev, for the price of a serving flag https://media.thesatyajit.com/ai/films/any-model-can-be-jev-poster.webp?v=a9838d26dbed9e50 SGLang has shipped a /v1/score endpoint since June 2025 that returns a probability for each candidate token you name, on an ordinary generative checkpoint… https://media.thesatyajit.com/ai/films/any-model-can-be-jev.mp4?v=397eaf214462befe 102 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/generative-ui-by-decision https://media.thesatyajit.com/ai/thumbs/generative-ui-by-decision.jpg?v=b2417cf5b7c79792 json-render + Jev: generative UI in two model calls https://media.thesatyajit.com/ai/films/generative-ui-by-decision-poster.webp?v=90eb753db8353afb Vercel Labs put a decision model behind json-render: instead of an LLM emitting a component tree, Jev picks from options your code enumerated. https://media.thesatyajit.com/ai/films/generative-ui-by-decision.mp4?v=4ad8ca995f4c0a0d 111 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/jev-as-substrate https://media.thesatyajit.com/ai/thumbs/jev-as-substrate.jpg?v=4c89b2234a272327 Jev as a substrate: every pixel is a Choice https://media.thesatyajit.com/ai/films/jev-as-substrate-poster.webp?v=7f91f27427cb811c Three projects this week used a typed-decision model as a general-purpose substrate. https://media.thesatyajit.com/ai/films/jev-as-substrate.mp4?v=18fa1e23025b50a8 107 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/jev-in-the-browser https://media.thesatyajit.com/ai/thumbs/jev-in-the-browser.jpg?v=fe383a3fe6cf6f88 Jev in the browser: N stops being a loop bound https://media.thesatyajit.com/ai/films/jev-in-the-browser-poster.webp?v=8682fd364f9752cc I built it: a per-option System One scorer on WebGPU/WASM through transformers.js, loaded on click, live in the page. It runs. https://media.thesatyajit.com/ai/films/jev-in-the-browser.mp4?v=54e261ace308f2c9 103 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/jev-scores-zero https://media.thesatyajit.com/ai/thumbs/jev-scores-zero.jpg?v=587cb698b2a3f865 Jev scores zero, and zero is the informative number https://media.thesatyajit.com/ai/films/jev-scores-zero-poster.webp?v=81bd9abc5fd5d810 An independent benchmark put Jev at 0 of 100 on relational choice — questions where one option's content decides another. Not low: zero, well below chance. https://media.thesatyajit.com/ai/films/jev-scores-zero.mp4?v=044f7e2c5aae1f32 99 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/orpo https://media.thesatyajit.com/ai/thumbs/orpo.jpg?v=48c76ff8f5ea1f2b ORPO: the odds ratio is the good part https://media.thesatyajit.com/ai/films/orpo-poster.webp?v=b79e023cad2d16ce ORPO appends a log-odds-ratio term to the SFT negative log-likelihood and drops the reference model. https://media.thesatyajit.com/ai/films/orpo.mp4?v=bc852a660cb7106a 111 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/qwen-livetranslate https://media.thesatyajit.com/ai/thumbs/qwen-livetranslate.jpg?v=1afad0b93c8a12f7 Qwen3.8-LiveTranslate: half a second off the lag, and 2.7 points on https://media.thesatyajit.com/ai/films/qwen-livetranslate-poster.webp?v=ccdd708b493f69ac Qwen3.8-LiveTranslate is announced on one number — average lagging down from 2.8s to 2.3s — and the number is nearly always quoted without the thing that… https://media.thesatyajit.com/ai/films/qwen-livetranslate.mp4?v=be54e780cb589a67 99 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/sam-3-1-object-multiplex https://media.thesatyajit.com/ai/thumbs/sam-3-1-object-multiplex.jpg?v=14889a3b9e9a8586 SAM 3.1: the 7× is at 128 objects https://media.thesatyajit.com/ai/films/sam-3-1-object-multiplex-poster.webp?v=3432a2dbbdd4c916 Meta shipped Object Multiplex, a shared-memory rewrite of SAM 3's video tracker that packs sixteen objects into one memory-path forward pass. https://media.thesatyajit.com/ai/films/sam-3-1-object-multiplex.mp4?v=f01fd3e0c1620b1d 102 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/search-by-meaning https://media.thesatyajit.com/ai/thumbs/search-by-meaning.jpg?v=c51f06c64deb8e45 jev-semgrep: grep, but the pattern is a proposition https://media.thesatyajit.com/ai/films/search-by-meaning-poster.webp?v=0f8311f0fc23b39a jev-semgrep is 309 lines of zero-dependency Node that asks a decision model, line by line, 'does this line match the meaning of X?' https://media.thesatyajit.com/ai/films/search-by-meaning.mp4?v=889854eb8ae100a5 107 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/splash-engine https://media.thesatyajit.com/ai/thumbs/splash-engine.jpg?v=1ff631b14e07a5c3 Splash on Apple silicon: 144 tok/s needs 502 GB/s, and Apple sells two M5 Max chips https://media.thesatyajit.com/ai/films/splash-engine-poster.webp?v=fff7ed86689523f9 Inco AI open-sourced Splash, a Metal inference engine for two Qwen models, under Apache-2.0; LM Studio shipped it a day later. https://media.thesatyajit.com/ai/films/splash-engine.mp4?v=36f450652805e8ab 95 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/three-tiers https://media.thesatyajit.com/ai/thumbs/three-tiers.jpg?v=3e13b19d0dce8a72 Decision models in three-tier agents: the middle tier is not in the middle https://media.thesatyajit.com/ai/films/three-tiers-poster.webp?v=98c6d37004bc5bcb Two posts in two days argue that agents now split three ways — deterministic code, a decision model, a reasoning model — and the split is real: I read three… https://media.thesatyajit.com/ai/films/three-tiers.mp4?v=5201f0f561720e23 106 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/webmcp-windtunnel https://media.thesatyajit.com/ai/thumbs/webmcp-windtunnel.jpg?v=afa6a962e209b8ba Jev on WindTunnel's WebMCP board: the 245x is 15x of interface and 16x of price https://media.thesatyajit.com/ai/films/webmcp-windtunnel-poster.webp?v=86f67adf318df9de WindTunnel's newest board has a decision model at rank one — Jev picks the tool, Mercury 2.5 writes the arguments, 49 of 49 tasks solved over WebMCP. https://media.thesatyajit.com/ai/films/webmcp-windtunnel.mp4?v=d6021d524557a7ad 102 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/what-decision-models-cannot-do https://media.thesatyajit.com/ai/thumbs/what-decision-models-cannot-do.jpg?v=770b296a6d876e7b What a decision model cannot do https://media.thesatyajit.com/ai/films/what-decision-models-cannot-do-poster.webp?v=1fe4eb145667030f Every System One piece this week was about what these models can do. This is the negative space, and it is the more useful half, because the limits are… https://media.thesatyajit.com/ai/films/what-decision-models-cannot-do.mp4?v=db831e95990c4bdc 89 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/cua-s1-forms https://media.thesatyajit.com/ai/thumbs/cua-s1-forms.jpg?v=0ce64ec8cceee764 cua-s1-forms: a System One model in 706,048 parameters https://media.thesatyajit.com/ai/films/cua-s1-forms-poster.webp?v=c7b63c7c8830ea2b Three open releases in one week rebuilt the Jev shape in public: cua-s1-forms trains a decision scorer from scratch at 706,048 parameters and reports 100%… https://media.thesatyajit.com/ai/films/cua-s1-forms.mp4?v=80cafeb311fe8951 110 2026-10-07 yes 2026-09-19 monthly 0.8 https://ai.satyajit.si/articles/monolith-1-0 https://media.thesatyajit.com/ai/thumbs/monolith-1-0.jpg?v=c2062dc80f1d051f 2026-09-19 monthly 0.7 https://ai.satyajit.si/articles/auto-gpu-kernel https://media.thesatyajit.com/ai/thumbs/auto-gpu-kernel.jpg?v=11c1870fef5f5b65 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/bend-2 https://media.thesatyajit.com/ai/thumbs/bend-2.jpg?v=98df32ed47be3827 Bend 2: proof-checked, and no longer an interaction net https://media.thesatyajit.com/ai/films/bend-2-poster.webp?v=2c1063aa92597da1 Victor Taelin's Bend 2 launched with three claims welded together: it 'blocks AI mistakes via proof checking — the same technique big AI labs used to solve… https://media.thesatyajit.com/ai/films/bend-2.mp4?v=560ef06281edfb28 108 2026-10-07 yes 2026-09-18 monthly 0.8 https://ai.satyajit.si/articles/bonsai-2-27b https://media.thesatyajit.com/ai/thumbs/bonsai-2-27b.jpg?v=6641c468b50a4828 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/browserskill https://media.thesatyajit.com/ai/thumbs/browserskill.jpg?v=ef609af2ae283baf 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/chesslfm https://media.thesatyajit.com/ai/thumbs/chesslfm.jpg?v=9a450dc50f6c125f 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/confidence-gated-fraud-detection https://media.thesatyajit.com/ai/thumbs/confidence-gated-fraud-detection.jpg?v=f4bad079f6cb7535 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/cuda-rust https://media.thesatyajit.com/ai/thumbs/cuda-rust.jpg?v=3ddb6f86b2664e4e 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/depth-anything-3-ros2 https://media.thesatyajit.com/ai/thumbs/depth-anything-3-ros2.jpg?v=cfd4b8ce30207202 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/dream-rsi https://media.thesatyajit.com/ai/thumbs/dream-rsi.jpg?v=1dc0477507117448 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/edge0-streaming-moe https://media.thesatyajit.com/ai/thumbs/edge0-streaming-moe.jpg?v=7c3714ac0f1f842e 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/fly-connectome-computing https://media.thesatyajit.com/ai/thumbs/fly-connectome-computing.jpg?v=0afaf4908ef27ce5 The fly connectome is a substrate, not a model https://media.thesatyajit.com/ai/films/fly-connectome-computing-poster.webp?v=c047a94ce9a2dd65 In September 2026 Google Research and HHMI Janelia published the complete male fruit-fly central nervous system — 166,691 neurons, brain to nerve cord. https://media.thesatyajit.com/ai/films/fly-connectome-computing.mp4?v=cb4350c364c40868 104 2026-10-07 yes 2026-09-18 monthly 0.8 https://ai.satyajit.si/articles/generated-web-at-2000-tokens-per-second https://media.thesatyajit.com/ai/thumbs/generated-web-at-2000-tokens-per-second.jpg?v=21e8348cd3727b37 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/glm-built-its-own-inference-infra https://media.thesatyajit.com/ai/thumbs/glm-built-its-own-inference-infra.jpg?v=bf46958202314240 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/goat-optimal-transport-attention https://media.thesatyajit.com/ai/thumbs/goat-optimal-transport-attention.jpg?v=b1134cc50a814bc7 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/grouped-value-attention https://media.thesatyajit.com/ai/thumbs/grouped-value-attention.jpg?v=111b2ea38c914c9d 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/jev-ecosystem https://media.thesatyajit.com/ai/thumbs/jev-ecosystem.jpg?v=3bb17b46aae384f7 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/jev-system-one-models https://media.thesatyajit.com/ai/thumbs/jev-system-one-models.jpg?v=d0fd4c1a8116a926 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/jina-ocr-v1 https://media.thesatyajit.com/ai/thumbs/jina-ocr-v1.jpg?v=b527e294d07caf87 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/lexing-on-the-gpu https://media.thesatyajit.com/ai/thumbs/lexing-on-the-gpu.jpg?v=0094653fd6d82dee 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/needle-3 https://media.thesatyajit.com/ai/thumbs/needle-3.jpg?v=9104ca89189ee396 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/open-1b-auditable-training https://media.thesatyajit.com/ai/thumbs/open-1b-auditable-training.jpg?v=94276d595dba69c4 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/parallel-constrained-decoding https://media.thesatyajit.com/ai/thumbs/parallel-constrained-decoding.jpg?v=cf21570e0c28ff97 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-omni-flash https://media.thesatyajit.com/ai/thumbs/qwen3-8-omni-flash.jpg?v=08539e8821fade6f Qwen3.8-Omni-Flash: skipping most of the video on purpose https://media.thesatyajit.com/ai/films/qwen3-8-omni-flash-poster.webp?v=00bb3ed289763973 Alibaba's first omni model built around agentic delivery: it decides what to watch instead of watching everything, and Qwen's own numbers say so — 45.7%… https://media.thesatyajit.com/ai/films/qwen3-8-omni-flash.mp4?v=551fea09364dffb1 105 2026-10-07 yes 2026-09-18 monthly 0.8 https://ai.satyajit.si/articles/relate-anything https://media.thesatyajit.com/ai/thumbs/relate-anything.jpg?v=a2182f682a81ad0f 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/retrieve-for-train https://media.thesatyajit.com/ai/thumbs/retrieve-for-train.jpg?v=7535fec201d42b83 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/rlcd-calibrated-decisions https://media.thesatyajit.com/ai/thumbs/rlcd-calibrated-decisions.jpg?v=ceed25298e91a0d0 RLCD is not constrained decoding https://media.thesatyajit.com/ai/films/rlcd-calibrated-decisions-poster.webp?v=d9e88f3ae8ac81bb TypeSafe's launch post makes three claims about Jev — a new architecture, a parallel sampler, and a training method called RLCD. https://media.thesatyajit.com/ai/films/rlcd-calibrated-decisions.mp4?v=abb7a113ddfb4295 101 2026-10-07 yes 2026-09-18 monthly 0.8 https://ai.satyajit.si/articles/runtime-abliteration https://media.thesatyajit.com/ai/thumbs/runtime-abliteration.jpg?v=04799e467d1db68f 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/tiny-browser-models https://media.thesatyajit.com/ai/thumbs/tiny-browser-models.jpg?v=b469997e31fb7984 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/ultradata-code https://media.thesatyajit.com/ai/thumbs/ultradata-code.jpg?v=08b444ab632baff7 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/uno-diffusion-augmented https://media.thesatyajit.com/ai/thumbs/uno-diffusion-augmented.jpg?v=517afd0f35f7c9ea 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/vram-is-a-policy https://media.thesatyajit.com/ai/thumbs/vram-is-a-policy.jpg?v=262821abdff6d29a 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/xgen-jing-dao https://media.thesatyajit.com/ai/thumbs/xgen-jing-dao.jpg?v=2272d3e3d0271cd4 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/zdtaichu-5-9b https://media.thesatyajit.com/ai/thumbs/zdtaichu-5-9b.jpg?v=513d418f9fd60663 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/auk-speech-editing https://media.thesatyajit.com/ai/thumbs/auk-speech-editing.jpg?v=dcaa91ae41676d82 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-2b-distill https://media.thesatyajit.com/ai/thumbs/qwen3-8-2b-distill.jpg?v=6fcee99b5fd67c36 2026-09-18 monthly 0.7 https://ai.satyajit.si/articles/soup-cli https://media.thesatyajit.com/ai/thumbs/soup-cli.jpg?v=20b82f2398f2d32b 2026-09-13 monthly 0.7 https://ai.satyajit.si/articles/marigold-v2 https://media.thesatyajit.com/ai/thumbs/marigold-v2.jpg?v=4e976069f680babe 2026-09-10 monthly 0.7 https://ai.satyajit.si/articles/activation-probe-monitors https://media.thesatyajit.com/ai/thumbs/activation-probe-monitors.jpg?v=153a8f8eea054338 2026-09-09 monthly 0.7 https://ai.satyajit.si/articles/cross-model-kv-transfer https://media.thesatyajit.com/ai/thumbs/cross-model-kv-transfer.jpg?v=938581181aef560d 2026-09-09 monthly 0.7 https://ai.satyajit.si/articles/desert-ant-labs https://media.thesatyajit.com/ai/thumbs/desert-ant-labs.jpg?v=c71f9b955367188c 2026-09-09 monthly 0.7 https://ai.satyajit.si/articles/flasepformer-jaec https://media.thesatyajit.com/ai/thumbs/flasepformer-jaec.jpg?v=ec6a3d98aaf7112a 2026-09-09 monthly 0.7 https://ai.satyajit.si/articles/fly-connectome-reservoir https://media.thesatyajit.com/ai/thumbs/fly-connectome-reservoir.jpg?v=3326f77d7deedf15 2026-09-09 monthly 0.7 https://ai.satyajit.si/articles/geoguessr-rl-environment https://media.thesatyajit.com/ai/thumbs/geoguessr-rl-environment.jpg?v=36c546f1c6c521be Stop looking, just guess: recomputing a GeoGuessr RL run from its raw data https://media.thesatyajit.com/ai/films/geoguessr-rl-environment-poster.webp?v=3071c6b2eaa088f5 Adithya Kolavi's Hugging Face article trains a Qwen3.5-4B + LoRA with TRL's GRPOTrainer inside an OpenEnv GeoGuessr environment, finishing second of eleven… https://media.thesatyajit.com/ai/films/geoguessr-rl-environment.mp4?v=0c0a4dd7f89b00bd 104 2026-10-07 yes 2026-09-09 monthly 0.8 https://ai.satyajit.si/articles/looped-transformers-matched-compute https://media.thesatyajit.com/ai/thumbs/looped-transformers-matched-compute.jpg?v=7371185a29a6378a SMELT: looping wins even after you pay for the FLOPs https://media.thesatyajit.com/ai/films/looped-transformers-matched-compute-poster.webp?v=11930b369952f33c This site's critique of virtual logic depth (arXiv 2506.18233) was that every looped-transformer comparison in the literature is parameter-matched, not… https://media.thesatyajit.com/ai/films/looped-transformers-matched-compute.mp4?v=ae40268222f8ac05 105 2026-10-07 yes 2026-09-09 monthly 0.8 https://ai.satyajit.si/articles/qwen-drive-1-0 https://media.thesatyajit.com/ai/thumbs/qwen-drive-1-0.jpg?v=cedd431ef353435f 2026-09-09 monthly 0.7 https://ai.satyajit.si/articles/aira-research-agents https://media.thesatyajit.com/ai/thumbs/aira-research-agents.jpg?v=49c338b9d6465302 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/audio8 https://media.thesatyajit.com/ai/thumbs/audio8.jpg?v=7032bc4ff113e188 Audio8: eight checkpoints, one broken link, and a 3B model with nowhere to ship https://media.thesatyajit.com/ai/films/audio8-poster.webp?v=bafc74254cba3767 Audio8's own framing for its two-month run of on-device audio releases is that it's more than a pile of checkpoints — model design, a technical report… https://media.thesatyajit.com/ai/films/audio8.mp4?v=cd56c0a7c2434869 116 2026-10-07 yes 2026-09-08 monthly 0.8 https://ai.satyajit.si/articles/extropic-z1t https://media.thesatyajit.com/ai/thumbs/extropic-z1t.jpg?v=fd0393df358acb7f Z1T: sparse transformers for a chip that samples, not multiplies — and what 100x actually measures https://media.thesatyajit.com/ai/films/extropic-z1t-poster.webp?v=46a0337efcd4be6c Extropic's Z1T pairs a probabilistic sampling chip with an FPGA to run transformer-like models, leading with 'over 100x energy efficiency gains vs GPUs'… https://media.thesatyajit.com/ai/films/extropic-z1t.mp4?v=89d1dbbb88c3979f 113 2026-10-07 yes 2026-09-08 monthly 0.8 https://ai.satyajit.si/articles/ling-3-0-flash-fin https://media.thesatyajit.com/ai/thumbs/ling-3-0-flash-fin.jpg?v=ce97fc09be238ce7 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/next-latent-world-models https://media.thesatyajit.com/ai/thumbs/next-latent-world-models.jpg?v=e25c466cdd34085e 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/sparda https://media.thesatyajit.com/ai/thumbs/sparda.jpg?v=7e2329a6037060b4 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/tgrep https://media.thesatyajit.com/ai/thumbs/tgrep.jpg?v=f55842f8fe11bcc6 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/virtual-logic-depth https://media.thesatyajit.com/ai/thumbs/virtual-logic-depth.jpg?v=272a22833ac23063 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/vit-compression-plant-disease https://media.thesatyajit.com/ai/thumbs/vit-compression-plant-disease.jpg?v=4701c224dfeaf920 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/whisper-hallucination-projection https://media.thesatyajit.com/ai/thumbs/whisper-hallucination-projection.jpg?v=5b1128c449595ec8 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/paint-with-code https://media.thesatyajit.com/ai/thumbs/paint-with-code.jpg?v=94f9dc0d3b2c7cef 2026-09-08 monthly 0.7 https://ai.satyajit.si/articles/scaling-agentic-rl https://media.thesatyajit.com/ai/thumbs/scaling-agentic-rl.jpg?v=a3d3ce1302be0b5d Scaling agentic RL: 365,000 environments behind one contract https://media.thesatyajit.com/ai/films/scaling-agentic-rl-poster.webp?v=cada02296b8b046c Prime Intellect integrated 23 agentic tasksets — software-engineering, terminal, and web-search — behind a single taskset API: ~365,000 tasks with prebuilt… https://media.thesatyajit.com/ai/films/scaling-agentic-rl.mp4?v=40ffa86c5a983989 110 2026-10-07 yes 2026-09-08 monthly 0.8 https://ai.satyajit.si/articles/dart-realtime-detection https://media.thesatyajit.com/ai/thumbs/dart-realtime-detection.jpg?v=eec6d6a53e6916da 2026-08-30 monthly 0.7 https://ai.satyajit.si/articles/glm-5-3-flash-uncensored https://media.thesatyajit.com/ai/thumbs/glm-5-3-flash-uncensored.jpg?v=cd54a6d42202957c 2026-08-30 monthly 0.7 https://ai.satyajit.si/articles/hauhaucs-qwen-fastmtp https://media.thesatyajit.com/ai/thumbs/hauhaucs-qwen-fastmtp.jpg?v=a0e208f84694ee14 2026-08-30 monthly 0.7 https://ai.satyajit.si/articles/levjepa https://media.thesatyajit.com/ai/thumbs/levjepa.jpg?v=eb080255eba47d28 LeVJEPA: one encoder, zero collapse-prevention machinery, and what 5.6-20.8x actually measures https://media.thesatyajit.com/ai/films/levjepa-poster.webp?v=7b42c2bfb7917ee0 LeVJEPA trains a single ViT video encoder with no EMA target network, no stop-gradient, and no predictor — replacing all three with SIGReg, a distributional… https://media.thesatyajit.com/ai/films/levjepa.mp4?v=9a8fb6ef5abfc508 105 2026-10-07 yes 2026-08-30 monthly 0.8 https://ai.satyajit.si/articles/hy4-preview https://media.thesatyajit.com/ai/thumbs/hy4-preview.jpg?v=fec5e239449b61d6 2026-08-29 monthly 0.7 https://ai.satyajit.si/articles/glm-5-3-flash-mlx https://media.thesatyajit.com/ai/thumbs/glm-5-3-flash-mlx.jpg?v=3bc25aee94d89249 2026-08-28 monthly 0.7 https://ai.satyajit.si/articles/phonellm-alpha-1 https://media.thesatyajit.com/ai/thumbs/phonellm-alpha-1.jpg?v=fffbef325355e533 2026-08-28 monthly 0.7 https://ai.satyajit.si/articles/vllm-speculative-decoding-amd https://media.thesatyajit.com/ai/thumbs/vllm-speculative-decoding-amd.jpg?v=04a9ab515f7f1b41 2026-08-28 monthly 0.7 https://ai.satyajit.si/articles/glm-5-3 https://media.thesatyajit.com/ai/thumbs/glm-5-3.jpg?v=9dde3476cdd5800a 2026-08-28 monthly 0.7 https://ai.satyajit.si/articles/diffusionopsd https://media.thesatyajit.com/ai/thumbs/diffusionopsd.jpg?v=37d58cdc2d193fb4 2026-08-27 monthly 0.7 https://ai.satyajit.si/articles/voicemem https://media.thesatyajit.com/ai/thumbs/voicemem.jpg?v=d17e0cd7977f6d05 2026-08-27 monthly 0.7 https://ai.satyajit.si/articles/breeze-tts-2 https://media.thesatyajit.com/ai/thumbs/breeze-tts-2.jpg?v=2a4fe6dcd196af71 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/flash-attention-3 https://media.thesatyajit.com/ai/thumbs/flash-attention-3.jpg?v=49f3c34216b17c2c 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/glm-5-3-flash https://media.thesatyajit.com/ai/thumbs/glm-5-3-flash.jpg?v=3aeb83dceb2cb04b 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/needle-environments https://media.thesatyajit.com/ai/thumbs/needle-environments.jpg?v=67340d72a0e5d3f3 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-whittle-moe https://media.thesatyajit.com/ai/thumbs/qwen3-8-whittle-moe.jpg?v=5e1bd2518b5e88e3 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/sana https://media.thesatyajit.com/ai/thumbs/sana.jpg?v=507fd47205d5896a Sana: the autoencoder does more of the work than the linear attention https://media.thesatyajit.com/ai/films/sana-poster.webp?v=cc0ce8194c3a1917 NVIDIA's Sana is sold as a linear-attention diffusion transformer that is 100× faster than FLUX. https://media.thesatyajit.com/ai/films/sana.mp4?v=b6af6776e5c48164 114 2026-10-07 yes 2026-08-26 monthly 0.8 https://ai.satyajit.si/articles/sglang https://media.thesatyajit.com/ai/thumbs/sglang.jpg?v=1b719bc8fd378153 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/tiel-coder-35b-a3b https://media.thesatyajit.com/ai/thumbs/tiel-coder-35b-a3b.jpg?v=d75aff2aad099934 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/vllm https://media.thesatyajit.com/ai/thumbs/vllm.jpg?v=4d0e24fa4d56a135 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/freetoken https://media.thesatyajit.com/ai/thumbs/freetoken.jpg?v=e042022b7d9472d8 FreeToken: a 753B model on one GPU, and the two bandwidths that decide everything https://media.thesatyajit.com/ai/films/freetoken-poster.webp?v=ad0352363fa2f1b3 An edge-native MoE engine that refuses to pick an offloading strategy and measures one instead. https://media.thesatyajit.com/ai/films/freetoken.mp4?v=cf3c2cded661d55b 100 2026-10-07 yes 2026-08-26 monthly 0.8 https://ai.satyajit.si/articles/colibri https://media.thesatyajit.com/ai/thumbs/colibri.jpg?v=7e6f946684032a08 2026-08-26 monthly 0.7 https://ai.satyajit.si/articles/apodex-frontier-agent https://media.thesatyajit.com/ai/thumbs/apodex-frontier-agent.jpg?v=ba8670258dbc12d6 2026-08-25 monthly 0.7 https://ai.satyajit.si/articles/gliner-2-5 https://media.thesatyajit.com/ai/thumbs/gliner-2-5.jpg?v=028c90065647eb06 2026-08-25 monthly 0.7 https://ai.satyajit.si/articles/pipette https://media.thesatyajit.com/ai/thumbs/pipette.jpg?v=9ab68f59c822b387 2026-08-25 monthly 0.7 https://ai.satyajit.si/articles/spec-ptc https://media.thesatyajit.com/ai/thumbs/spec-ptc.jpg?v=fde5fb6d2e1a7fc7 2026-08-25 monthly 0.7 https://ai.satyajit.si/articles/scaling-laws-2026 https://media.thesatyajit.com/ai/thumbs/scaling-laws-2026.jpg?v=01b487b91745205b Scaling laws in 2026: the line held, the recipe didn't https://media.thesatyajit.com/ai/films/scaling-laws-2026-poster.webp?v=d5d355169eb2ece6 The standard retelling of scaling laws ends in 2022 with 'twenty tokens per parameter' and treats it as the industry's blueprint. https://media.thesatyajit.com/ai/films/scaling-laws-2026.mp4?v=45691a5d666ed318 104 2026-10-07 yes 2026-08-24 monthly 0.8 https://ai.satyajit.si/articles/kimi-k3 https://media.thesatyajit.com/ai/thumbs/kimi-k3.jpg?v=a1b94df6afeb30e6 Kimi K3: a 2.8T open model that turns compute into intelligence 2.5× better https://media.thesatyajit.com/ai/films/kimi-k3-poster.webp?v=343744279e84629a Moonshot's Kimi K3 is a 2.8T-parameter, 104B-active open MoE with a 1M-token context — now with weights, code and a 47-page technical report on the table. https://media.thesatyajit.com/ai/films/kimi-k3.mp4?v=223e05f946606e55 106 2026-10-07 yes 2026-08-24 monthly 0.8 https://ai.satyajit.si/articles/flex-pi https://media.thesatyajit.com/ai/thumbs/flex-pi.jpg?v=262086d68b577c7d 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/lfm25-dspark https://media.thesatyajit.com/ai/thumbs/lfm25-dspark.jpg?v=38e290a13ace0db0 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/matryoshka-lm-suites https://media.thesatyajit.com/ai/thumbs/matryoshka-lm-suites.jpg?v=e40d73af49003c31 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/mimimodel https://media.thesatyajit.com/ai/thumbs/mimimodel.jpg?v=397a5036be014de2 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/nar-tts https://media.thesatyajit.com/ai/thumbs/nar-tts.jpg?v=6e7eb6486f2fc2da 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/needle-finetune https://media.thesatyajit.com/ai/thumbs/needle-finetune.jpg?v=c801fc4c7b1f424c 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/sensenova-u15 https://media.thesatyajit.com/ai/thumbs/sensenova-u15.jpg?v=d1ac25142292d171 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/taimi-14b-med https://media.thesatyajit.com/ai/thumbs/taimi-14b-med.jpg?v=3edc9227e5e1056d 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/ultra-fineweb https://media.thesatyajit.com/ai/thumbs/ultra-fineweb.jpg?v=49918fdf25932914 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/verifier-frontier https://media.thesatyajit.com/ai/thumbs/verifier-frontier.jpg?v=c1b3a308809abf6a 2026-08-23 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-open-weights https://media.thesatyajit.com/ai/thumbs/qwen3-8-open-weights.jpg?v=7fb19bba2fa7e1ef 2026-08-15 monthly 0.7 https://ai.satyajit.si/articles/arcee-open-models-api https://media.thesatyajit.com/ai/thumbs/arcee-open-models-api.jpg?v=07e160a1a3a30cf4 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/deepseek-harness https://media.thesatyajit.com/ai/thumbs/deepseek-harness.jpg?v=dabf75c4157fce2a DeepSeek Harness: an agent harness that refuses to send what it didn't log https://media.thesatyajit.com/ai/films/deepseek-harness-poster.webp?v=19573fef01dc1a19 DeepSeek open-sourced dsh, an agent harness built on Cordis where everything — including the agent loop — is a replaceable plugin across 219 packages. https://media.thesatyajit.com/ai/films/deepseek-harness.mp4?v=4d2afb1688e4c07d 100 2026-10-07 yes 2026-08-14 monthly 0.8 https://ai.satyajit.si/articles/dots3 https://media.thesatyajit.com/ai/thumbs/dots3.jpg?v=787ab319a6e13935 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/full-bandwidth-transformer https://media.thesatyajit.com/ai/thumbs/full-bandwidth-transformer.jpg?v=9cda03da4c2718dc 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/lfm2-5-vl-3b https://media.thesatyajit.com/ai/thumbs/lfm2-5-vl-3b.jpg?v=4b3807f42a845296 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/magi-2-preview https://media.thesatyajit.com/ai/thumbs/magi-2-preview.jpg?v=78d9164444272694 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/minimax-music3 https://media.thesatyajit.com/ai/thumbs/minimax-music3.jpg?v=1dc128b960ca5d7e 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/nac https://media.thesatyajit.com/ai/thumbs/nac.jpg?v=d48e2fca7f67c5e0 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/softmax-bottleneck https://media.thesatyajit.com/ai/thumbs/softmax-bottleneck.jpg?v=e60b98ad69d7eb72 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/speech-to-speech https://media.thesatyajit.com/ai/thumbs/speech-to-speech.jpg?v=08f25fe629cc47de 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/toast-1 https://media.thesatyajit.com/ai/thumbs/toast-1.jpg?v=dddb4418e87987d8 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/worldclaw https://media.thesatyajit.com/ai/thumbs/worldclaw.jpg?v=aa107b8e92a1b28d 2026-08-14 monthly 0.7 https://ai.satyajit.si/articles/ltc-gated-delta https://media.thesatyajit.com/ai/thumbs/ltc-gated-delta.jpg?v=c5930b8f8c2e62ff Liquid time constants and gated delta rules: two literatures, one recurrence https://media.thesatyajit.com/ai/films/ltc-gated-delta-poster.webp?v=01b6b907d59490dd Liquid Time-constant Networks (2020) and Gated DeltaNet (2024) come from different fields, cite different ancestors, and use different notation. https://media.thesatyajit.com/ai/films/ltc-gated-delta.mp4?v=febd0733b9638057 110 2026-10-07 yes 2026-08-10 monthly 0.8 https://ai.satyajit.si/articles/muse-glimmer https://media.thesatyajit.com/ai/thumbs/muse-glimmer.jpg?v=d7efb7cf64b48e64 Muse Glimmer: an agentic model designed backwards from a 24 GB budget https://media.thesatyajit.com/ai/films/muse-glimmer-poster.webp?v=ec0f001e399960f8 Meta's 30B open-weight agent model is a distillation of Muse Spark built to run on one consumer GPU. https://media.thesatyajit.com/ai/films/muse-glimmer.mp4?v=8ecf6033e1ab4063 110 2026-10-07 yes 2026-08-10 monthly 0.8 https://ai.satyajit.si/articles/skaling-law https://media.thesatyajit.com/ai/thumbs/skaling-law.jpg?v=13cca4a2c7d8d16c The Skaling law: Chinchilla assumes model size and data don't interact, and they do https://media.thesatyajit.com/ai/films/skaling-law-poster.webp?v=bad5657c130cdcb0 FAIR's Skaling law adds one exponent to the Chinchilla form — L = (A/Nᵅ + B/Dᵝ)^k + E — and cuts prediction error 1.5–3×. https://media.thesatyajit.com/ai/films/skaling-law.mp4?v=003b7c466cc8270f 108 2026-10-07 yes 2026-08-10 monthly 0.8 https://ai.satyajit.si/articles/btl-4 https://media.thesatyajit.com/ai/thumbs/btl-4.jpg?v=1b13fac2f899d68f 2026-08-06 monthly 0.7 https://ai.satyajit.si/articles/intern-s2-mobius https://media.thesatyajit.com/ai/thumbs/intern-s2-mobius.jpg?v=d4504f1263588138 2026-08-06 monthly 0.7 https://ai.satyajit.si/articles/maple-preview https://media.thesatyajit.com/ai/thumbs/maple-preview.jpg?v=af7082125128a084 2026-08-06 monthly 0.7 https://ai.satyajit.si/articles/pokee-isaac-28b https://media.thesatyajit.com/ai/thumbs/pokee-isaac-28b.jpg?v=c1c37adb0a7a5163 2026-08-06 monthly 0.7 https://ai.satyajit.si/articles/prime-agent https://media.thesatyajit.com/ai/thumbs/prime-agent.jpg?v=cd3277d2cde3187b Prime Agent: the interface is a Python REPL, not a tool-call schema https://media.thesatyajit.com/ai/films/prime-agent-poster.webp?v=3dbe65f13902b751 Prime Intellect's open-source coding and research agent exposes exactly one model tool — a persistent IPython kernel — and treats subagents, files, shell… https://media.thesatyajit.com/ai/films/prime-agent.mp4?v=c740831fa9789389 105 2026-10-07 yes 2026-08-06 monthly 0.8 https://ai.satyajit.si/articles/recursive-language-models https://media.thesatyajit.com/ai/thumbs/recursive-language-models.jpg?v=78825f3066f09483 2026-08-06 monthly 0.7 https://ai.satyajit.si/articles/tencentdb-agent-memory https://media.thesatyajit.com/ai/thumbs/tencentdb-agent-memory.jpg?v=54e0ba8509aaa600 2026-08-06 monthly 0.7 https://ai.satyajit.si/articles/abot-world https://media.thesatyajit.com/ai/thumbs/abot-world.jpg?v=a606f375253b5550 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/angelspec https://media.thesatyajit.com/ai/thumbs/angelspec.jpg?v=74c6e019a0bcf69e 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/autocompact https://media.thesatyajit.com/ai/thumbs/autocompact.jpg?v=cb3a8dea1c581b27 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/ax-k2 https://media.thesatyajit.com/ai/thumbs/ax-k2.jpg?v=d95d84096450fc30 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/chimera-diffusion https://media.thesatyajit.com/ai/thumbs/chimera-diffusion.jpg?v=56fd82ffbd362cfa 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/dcformer https://media.thesatyajit.com/ai/thumbs/dcformer.jpg?v=06ee9ac18e2d9afa 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/dream-cubed https://media.thesatyajit.com/ai/thumbs/dream-cubed.jpg?v=c7db796faa961123 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/explorative-modeling https://media.thesatyajit.com/ai/thumbs/explorative-modeling.jpg?v=a3c7825db4a78348 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/fara-1-5 https://media.thesatyajit.com/ai/thumbs/fara-1-5.jpg?v=f3819e6bf888f60b 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/injective-language-models https://media.thesatyajit.com/ai/thumbs/injective-language-models.jpg?v=e3185753adcd27b8 Language models are injective, so a KV-cache is not a summary — it's the prompt, in another basis https://media.thesatyajit.com/ai/films/injective-language-models-poster.webp?v=9d14c17067db34af A proof that decoder-only transformer LMs are almost-surely injective — distinct prompts never collide to the same hidden state — confirmed across billions… https://media.thesatyajit.com/ai/films/injective-language-models.mp4?v=a6365c4765189f8d 103 2026-10-07 yes 2026-08-03 monthly 0.8 https://ai.satyajit.si/articles/inkling-small https://media.thesatyajit.com/ai/thumbs/inkling-small.jpg?v=7937bcf57f73b804 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/instella-moe https://media.thesatyajit.com/ai/thumbs/instella-moe.jpg?v=327420929abdb367 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/josie-2 https://media.thesatyajit.com/ai/thumbs/josie-2.jpg?v=33c028c5845cadfc 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/kda-half-life https://media.thesatyajit.com/ai/thumbs/kda-half-life.jpg?v=ebc163f00e636b97 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/kimi-k3-in-c https://media.thesatyajit.com/ai/thumbs/kimi-k3-in-c.jpg?v=dddca88caac7b416 kimi-k3-in-c: 2.78 trillion parameters, one CPU, 8 GB of RAM https://media.thesatyajit.com/ai/films/kimi-k3-in-c-poster.webp?v=f06d9d5f1b611088 A single-author C99 engine — no BLAS, no framework, no GPU — reimplements Kimi K3 inference from the checkpoint's own bytes: 2.78T parameters, a 1.56 TB… https://media.thesatyajit.com/ai/films/kimi-k3-in-c.mp4?v=89b9a17e61ea1048 108 2026-10-07 yes 2026-08-03 monthly 0.8 https://ai.satyajit.si/articles/lfm2-5-encoders https://media.thesatyajit.com/ai/thumbs/lfm2-5-encoders.jpg?v=416ad8403616a725 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/looped-models-done-right https://media.thesatyajit.com/ai/thumbs/looped-models-done-right.jpg?v=79679764d07fe293 Towards looped models done right: what actually separates Ouro from Huginn https://media.thesatyajit.com/ai/films/looped-models-done-right-poster.webp?v=a9222928db1c5edb IFM Research isolates the three design axes tangled inside every looped transformer — iteration envelope, input injection, recurrent-state init — with… https://media.thesatyajit.com/ai/films/looped-models-done-right.mp4?v=cd50147f73df8b45 99 2026-10-07 yes 2026-08-03 monthly 0.8 https://ai.satyajit.si/articles/memharness https://media.thesatyajit.com/ai/thumbs/memharness.jpg?v=69f0183daac5f6a8 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/minimax-h3 https://media.thesatyajit.com/ai/thumbs/minimax-h3.jpg?v=dcc78dbea808bebd 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/pdf-inspector https://media.thesatyajit.com/ai/thumbs/pdf-inspector.jpg?v=ef26c49b44c2c797 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/qwen-cua https://media.thesatyajit.com/ai/thumbs/qwen-cua.jpg?v=1c16fe9ba5e9bc08 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/qwen3-8-max https://media.thesatyajit.com/ai/thumbs/qwen3-8-max.jpg?v=7c05b082d7c0fba9 Qwen3.8-Max: 16 days, 265 commits, zero humans in the loop https://media.thesatyajit.com/ai/films/qwen3-8-max-poster.webp?v=f3adaeb6ce5cddb2 Alibaba's Qwen3.8-Max scales to 2.4T parameters (95B active) on the Qwen 3.5 architecture and becomes the first Qwen-Max-class model getting open weights… https://media.thesatyajit.com/ai/films/qwen3-8-max.mp4?v=7215e2963ebce976 116 2026-10-07 yes 2026-08-03 monthly 0.8 https://ai.satyajit.si/articles/recursive-harness-self-improvement https://media.thesatyajit.com/ai/thumbs/recursive-harness-self-improvement.jpg?v=3e055cfedde1a29d 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/sol-attn https://media.thesatyajit.com/ai/thumbs/sol-attn.jpg?v=8ba264b747be1fc0 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/uber-adr https://media.thesatyajit.com/ai/thumbs/uber-adr.jpg?v=02d39afc21a684f0 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/macaron-v1 https://media.thesatyajit.com/ai/thumbs/macaron-v1.jpg?v=e807c3e8b1f19201 2026-08-03 monthly 0.7 https://ai.satyajit.si/articles/neutrino-1 https://media.thesatyajit.com/ai/thumbs/neutrino-1.jpg?v=29916d463e49b947 2026-07-27 monthly 0.7 https://ai.satyajit.si/articles/btl-3 https://media.thesatyajit.com/ai/thumbs/btl-3.jpg?v=ec59a24d27caa383 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/first-order-rl https://media.thesatyajit.com/ai/thumbs/first-order-rl.jpg?v=e949e68b52edeb42 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/flux-3 https://media.thesatyajit.com/ai/thumbs/flux-3.jpg?v=3ef612d9780c7368 FLUX 3: when an image model decides to become a world model https://media.thesatyajit.com/ai/films/flux-3-poster.webp?v=db706d7ff72a5609 Black Forest Labs turned the image-only FLUX line into one multimodal flow-matching model that generates image, video (up to 20s with native audio), and… https://media.thesatyajit.com/ai/films/flux-3.mp4?v=04219d4eb13905fc 103 2026-10-07 yes 2026-07-24 monthly 0.8 https://ai.satyajit.si/articles/gepa-optimize-anything https://media.thesatyajit.com/ai/thumbs/gepa-optimize-anything.jpg?v=14b9f8351cdfb597 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/lanyon-neurosymbolic https://media.thesatyajit.com/ai/thumbs/lanyon-neurosymbolic.jpg?v=077663f4c98f2a02 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/ling-3-0-flash https://media.thesatyajit.com/ai/thumbs/ling-3-0-flash.jpg?v=3acbf78e556aae43 Ling-3.0-flash: a 124B open MoE that runs like a 5B and reaches for 1M tokens https://media.thesatyajit.com/ai/films/ling-3-0-flash-poster.webp?v=536c3f8f9a7a25f4 inclusionAI's Ling-3.0-flash is a 124B-parameter, ~5.1B-active open MoE that reaches a 1M-token context by interleaving Kimi Delta Attention (linear-time)… https://media.thesatyajit.com/ai/films/ling-3-0-flash.mp4?v=3b8c3014b010c2a5 104 2026-10-07 yes 2026-07-24 monthly 0.8 https://ai.satyajit.si/articles/mai-image-2-5-voice-2 https://media.thesatyajit.com/ai/thumbs/mai-image-2-5-voice-2.jpg?v=46b182f2278d7e2d 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/sana-video2 https://media.thesatyajit.com/ai/thumbs/sana-video2.jpg?v=96d9b527e7663de0 SANA-Video 2.0: keeping video attention linear without losing the picture https://media.thesatyajit.com/ai/films/sana-video2-poster.webp?v=aafe586ffecd1377 NVIDIA's SANA-Video 2.0 generates 720p, multi-second video on a single H100 — 13.06s for a 720p/5s clip, a claimed 120× over Wan 2.2-A14B. https://media.thesatyajit.com/ai/films/sana-video2.mp4?v=66d6feb34b36c016 107 2026-10-07 yes 2026-07-24 monthly 0.8 https://ai.satyajit.si/articles/solar-open2-250b https://media.thesatyajit.com/ai/thumbs/solar-open2-250b.jpg?v=d30dabd9e80d4fca 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/ternary15m https://media.thesatyajit.com/ai/thumbs/ternary15m.jpg?v=e15fc630f3030a11 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/nanbeige-4-2-3b https://media.thesatyajit.com/ai/thumbs/nanbeige-4-2-3b.jpg?v=d55fe456f49a98d7 2026-07-24 monthly 0.7 https://ai.satyajit.si/articles/antares https://media.thesatyajit.com/ai/thumbs/antares.jpg?v=37630510150a077f 2026-07-22 monthly 0.7 https://ai.satyajit.si/articles/gigatoken https://media.thesatyajit.com/ai/thumbs/gigatoken.jpg?v=617ae9eda8bfd3e4 2026-07-22 monthly 0.7 https://ai.satyajit.si/articles/mage-flow https://media.thesatyajit.com/ai/thumbs/mage-flow.jpg?v=4d02eea4c9bebace 2026-07-22 monthly 0.7 https://ai.satyajit.si/articles/nvidia-rubin https://media.thesatyajit.com/ai/thumbs/nvidia-rubin.jpg?v=9e8e3057df9a8a34 2026-07-22 monthly 0.7 https://ai.satyajit.si/articles/harness-compositional-generalization https://media.thesatyajit.com/ai/thumbs/harness-compositional-generalization.jpg?v=ddd981da7f645e0d 2026-07-21 monthly 0.7 https://ai.satyajit.si/articles/laguna-s-2-1 https://media.thesatyajit.com/ai/thumbs/laguna-s-2-1.jpg?v=e1bd10724009f563 2026-07-21 monthly 0.7 https://ai.satyajit.si/articles/longstraw-2m-rl https://media.thesatyajit.com/ai/thumbs/longstraw-2m-rl.jpg?v=1780ab5740051d40 2026-07-21 monthly 0.7 https://ai.satyajit.si/articles/qwen-image-3 https://media.thesatyajit.com/ai/thumbs/qwen-image-3.jpg?v=5b98feada813ee57 2026-07-21 monthly 0.7 https://ai.satyajit.si/articles/eagle-3-speculative-decoding https://media.thesatyajit.com/ai/thumbs/eagle-3-speculative-decoding.jpg?v=6dff5c2341c5a299 2026-07-20 monthly 0.7 https://ai.satyajit.si/articles/motif-2-6b https://media.thesatyajit.com/ai/thumbs/motif-2-6b.jpg?v=e5de162395fc418a 2026-07-20 monthly 0.7 https://ai.satyajit.si/articles/nemotron-audex https://media.thesatyajit.com/ai/thumbs/nemotron-audex.jpg?v=f1b7a0e224d02d06 2026-07-20 monthly 0.7 https://ai.satyajit.si/articles/qwen-audio-3-tts https://media.thesatyajit.com/ai/thumbs/qwen-audio-3-tts.jpg?v=680be77baacbb09b 2026-07-20 monthly 0.7 https://ai.satyajit.si/articles/cosmos-world-model https://media.thesatyajit.com/ai/thumbs/cosmos-world-model.jpg?v=415cefdb8a161717 2026-07-20 monthly 0.7 https://ai.satyajit.si/articles/harness-effect https://media.thesatyajit.com/ai/thumbs/harness-effect.jpg?v=fbddfa08ab4e062e 2026-07-18 monthly 0.7 https://ai.satyajit.si/articles/diffusing-blame https://media.thesatyajit.com/ai/thumbs/diffusing-blame.jpg?v=83d0f4d87c1ea8a3 2026-07-17 monthly 0.7 https://ai.satyajit.si/articles/intern-s2 https://media.thesatyajit.com/ai/thumbs/intern-s2.jpg?v=fe16a2fcfd668458 2026-07-17 monthly 0.7 https://ai.satyajit.si/articles/limssr https://media.thesatyajit.com/ai/thumbs/limssr.jpg?v=9ee2c377ff70e702 2026-07-17 monthly 0.7 https://ai.satyajit.si/articles/videochat3 https://media.thesatyajit.com/ai/thumbs/videochat3.jpg?v=133e52eec1815e29 2026-07-17 monthly 0.7 https://ai.satyajit.si/articles/zuna-1-1 https://media.thesatyajit.com/ai/thumbs/zuna-1-1.jpg?v=f26076248694b623 2026-07-17 monthly 0.7 https://ai.satyajit.si/articles/grape-position-encoding https://media.thesatyajit.com/ai/thumbs/grape-position-encoding.jpg?v=91adafeaa1d9b131 2026-07-16 monthly 0.7 https://ai.satyajit.si/articles/bonsai-27b https://media.thesatyajit.com/ai/thumbs/bonsai-27b.jpg?v=732b5d0e530a3bc1 2026-07-15 monthly 0.7 https://ai.satyajit.si/articles/inkling https://media.thesatyajit.com/ai/thumbs/inkling.jpg?v=dd63d7e814e2d433 2026-07-15 monthly 0.7 https://ai.satyajit.si/articles/lotus-latent-reasoning https://media.thesatyajit.com/ai/thumbs/lotus-latent-reasoning.jpg?v=9211e28e1b97d501 LOTUS: reasoning in the hidden states, not the token stream https://media.thesatyajit.com/ai/films/lotus-latent-reasoning-poster.webp?v=ea17f5c075f495ca Explicit chain-of-thought writes every reasoning step out token-by-token, which is slow; latent CoT reasons in hidden states instead — but past 1B… https://media.thesatyajit.com/ai/films/lotus-latent-reasoning.mp4?v=c969a53dee1e5585 93 2026-10-07 yes 2026-07-15 monthly 0.8 https://ai.satyajit.si/articles/ring-zero-trillion-scale-rl https://media.thesatyajit.com/ai/thumbs/ring-zero-trillion-scale-rl.jpg?v=990692cdf42d1047 Ring-Zero: what a trillion-parameter model learns from reward alone https://media.thesatyajit.com/ai/films/ring-zero-trillion-scale-rl-poster.webp?v=53477f9abea496f4 Zero RL — reinforcement learning from verifiable rewards, no human labels, no SFT — has mostly been studied on small models. https://media.thesatyajit.com/ai/films/ring-zero-trillion-scale-rl.mp4?v=53ee3ec42538440b 99 2026-10-07 yes 2026-07-15 monthly 0.8 https://ai.satyajit.si/articles/mach-mind-4-flash https://media.thesatyajit.com/ai/thumbs/mach-mind-4-flash.jpg?v=84cff302c26329d5 2026-07-13 monthly 0.7 https://ai.satyajit.si/articles/soofi-s https://media.thesatyajit.com/ai/thumbs/soofi-s.jpg?v=f206e2e4940de194 2026-07-13 monthly 0.7 https://ai.satyajit.si/articles/dgx-spark-batching https://media.thesatyajit.com/ai/thumbs/dgx-spark-batching.jpg?v=41925cc1cf912307 2026-07-10 monthly 0.7 https://ai.satyajit.si/articles/kat-coder-agentic-training https://media.thesatyajit.com/ai/thumbs/kat-coder-agentic-training.jpg?v=bb0c9dd124af494f 2026-07-10 monthly 0.7 https://ai.satyajit.si/articles/musacoder-gpu-kernels https://media.thesatyajit.com/ai/thumbs/musacoder-gpu-kernels.jpg?v=57f0d4de57ce519e 2026-07-10 monthly 0.7 https://ai.satyajit.si/articles/switch-transformer https://media.thesatyajit.com/ai/thumbs/switch-transformer.jpg?v=7c648f6e3d7d26b6 2026-07-10 monthly 0.7 https://ai.satyajit.si/articles/swe-1-7 https://media.thesatyajit.com/ai/thumbs/swe-1-7.jpg?v=d69a96d7015c0ce0 2026-07-09 monthly 0.7 https://ai.satyajit.si/articles/a6b-k-expansion https://media.thesatyajit.com/ai/thumbs/a6b-k-expansion.jpg?v=d1422e0aef13d7fa 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/agent-harness https://media.thesatyajit.com/ai/thumbs/agent-harness.jpg?v=7196e9f9541093c0 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/antidoom https://media.thesatyajit.com/ai/thumbs/antidoom.jpg?v=b8109c0bf7d8c7ee 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/frontier-rl-cheaper https://media.thesatyajit.com/ai/thumbs/frontier-rl-cheaper.jpg?v=dae69514e6c78d4a 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/gemma-4 https://media.thesatyajit.com/ai/thumbs/gemma-4.jpg?v=ec9e82457116c476 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/hunyuan-hy3 https://media.thesatyajit.com/ai/thumbs/hunyuan-hy3.jpg?v=22c5c616100e9454 Hunyuan Hy3: Tencent's 295B-A21B MoE, and the community 1M GGUF https://media.thesatyajit.com/ai/films/hunyuan-hy3-poster.webp?v=e493870aa372349f Tencent's Hy3 (preview) is a 295-billion-parameter Mixture-of-Experts model that activates only 21B per token, ships open weights with a 256K native… https://media.thesatyajit.com/ai/films/hunyuan-hy3.mp4?v=7e9dd031c3ed4a47 112 2026-10-07 yes 2026-07-08 monthly 0.8 https://ai.satyajit.si/articles/jspace-open https://media.thesatyajit.com/ai/thumbs/jspace-open.jpg?v=d92498d06488fb52 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/muon-optimizer https://media.thesatyajit.com/ai/thumbs/muon-optimizer.jpg?v=9dcbb5c72491ee39 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/rollout-routing-replay https://media.thesatyajit.com/ai/thumbs/rollout-routing-replay.jpg?v=175e50faac5f5ef5 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/torch-profiler https://media.thesatyajit.com/ai/thumbs/torch-profiler.jpg?v=2d53efad174ed824 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/zvec https://media.thesatyajit.com/ai/thumbs/zvec.jpg?v=a018b22011b790ed 2026-07-08 monthly 0.7 https://ai.satyajit.si/articles/laguna-model-factory https://media.thesatyajit.com/ai/thumbs/laguna-model-factory.jpg?v=51903f2537eec942 Laguna's Model Factory: treating model development as an industrial process https://media.thesatyajit.com/ai/films/laguna-model-factory-poster.webp?v=484397adeeeb774d Poolside's Laguna M.1 (225B/23B-active) and XS.2 (33B/3B-active) are competent agentic-coding MoE models — but the tech report is really about the thing… https://media.thesatyajit.com/ai/films/laguna-model-factory.mp4?v=f464148b0f003d56 105 2026-10-07 yes 2026-07-08 monthly 0.8 https://ai.satyajit.si/articles/longcat-2 https://media.thesatyajit.com/ai/thumbs/longcat-2.jpg?v=e9c0852cdf0224eb LongCat 2.0: a 1.6T open-weights MoE, and the sparse attention behind it https://media.thesatyajit.com/ai/films/longcat-2-poster.webp?v=6114b87c9de0ee9a Meituan's LongCat 2.0 is a 1.6-trillion-parameter Mixture-of-Experts model that activates only ~48B per token, ships under MIT, and was trained end-to-end… https://media.thesatyajit.com/ai/films/longcat-2.mp4?v=f59b83306b027fb1 115 2026-10-07 yes 2026-07-05 monthly 0.8 https://ai.satyajit.si/articles/minimax-sparse-attention https://media.thesatyajit.com/ai/thumbs/minimax-sparse-attention.jpg?v=7725110546914f37 2026-07-04 monthly 0.7 https://ai.satyajit.si/articles/mrflow-diffusion-acceleration https://media.thesatyajit.com/ai/thumbs/mrflow-diffusion-acceleration.jpg?v=595269dc02d62883 2026-07-04 monthly 0.7 https://ai.satyajit.si/articles/nemotron-nvfp4 https://media.thesatyajit.com/ai/thumbs/nemotron-nvfp4.jpg?v=3e7a15f3d912056f 2026-07-04 monthly 0.7 https://ai.satyajit.si/articles/program-as-weights https://media.thesatyajit.com/ai/thumbs/program-as-weights.jpg?v=f9ccaacae16552ea 2026-07-04 monthly 0.7 https://ai.satyajit.si/articles/tabular-foundation-model https://media.thesatyajit.com/ai/thumbs/tabular-foundation-model.jpg?v=1b5fbe8cd75f03a6 2026-07-04 monthly 0.7 https://ai.satyajit.si/articles/leanstral-formal-proofs https://media.thesatyajit.com/ai/thumbs/leanstral-formal-proofs.jpg?v=3310481eae18d9bb 2026-07-03 monthly 0.7 https://ai.satyajit.si/articles/mimo-v2-flash https://media.thesatyajit.com/ai/thumbs/mimo-v2-flash.jpg?v=c11065ce3ca8fd30 2026-07-03 monthly 0.7 https://ai.satyajit.si/articles/set-diffusion https://media.thesatyajit.com/ai/thumbs/set-diffusion.jpg?v=ace941e66a2a9225 2026-07-03 monthly 0.7 https://ai.satyajit.si/articles/bm25 https://media.thesatyajit.com/ai/thumbs/bm25.jpg?v=cdb343b1e7ad5ef5 2026-07-02 monthly 0.7 https://ai.satyajit.si/articles/nemotron-twotower https://media.thesatyajit.com/ai/thumbs/nemotron-twotower.jpg?v=771b2db42dcbe457 2026-07-02 monthly 0.7 https://ai.satyajit.si/articles/turboquant-kv-cache https://media.thesatyajit.com/ai/thumbs/turboquant-kv-cache.jpg?v=7dcaa0a289c188c0 2026-07-02 monthly 0.7 https://ai.satyajit.si/articles/turbovec https://media.thesatyajit.com/ai/thumbs/turbovec.jpg?v=4fbd46d79c863355 2026-07-02 monthly 0.7 https://ai.satyajit.si/articles/hydrahead https://media.thesatyajit.com/ai/thumbs/hydrahead.jpg?v=074413f7c1b1a81b 2026-07-01 monthly 0.7 https://ai.satyajit.si/articles/tapered-language-models https://media.thesatyajit.com/ai/thumbs/tapered-language-models.jpg?v=5f182cb7d0637d16 2026-07-01 monthly 0.7 https://ai.satyajit.si/articles/agents-a1 https://media.thesatyajit.com/ai/thumbs/agents-a1.jpg?v=47cc786975e4398f 2026-06-30 monthly 0.7 https://ai.satyajit.si/articles/fast-lio2-lidar-inertial-odometry https://media.thesatyajit.com/ai/thumbs/fast-lio2-lidar-inertial-odometry.jpg?v=408b9a902f51acc0 2026-06-28 monthly 0.7 https://ai.satyajit.si/articles/how-llm-inference-works https://media.thesatyajit.com/ai/thumbs/how-llm-inference-works.jpg?v=86958dc142bff577 2026-06-28 monthly 0.7 https://ai.satyajit.si/articles/illada-diffusion-language-model https://media.thesatyajit.com/ai/thumbs/illada-diffusion-language-model.jpg?v=d970f3c4935ca8b9 2026-06-28 monthly 0.7 https://ai.satyajit.si/articles/kalman-filter https://media.thesatyajit.com/ai/thumbs/kalman-filter.jpg?v=7acbc16f98da5676 2026-06-28 monthly 0.7 https://ai.satyajit.si/articles/megatrain-single-gpu-training https://media.thesatyajit.com/ai/thumbs/megatrain-single-gpu-training.jpg?v=722fcaa672ce6e32 2026-06-28 monthly 0.7 https://ai.satyajit.si/articles/deepseek-dspark https://media.thesatyajit.com/ai/thumbs/deepseek-dspark.jpg?v=56317b0699cf408b 2026-06-27 monthly 0.7 https://ai.satyajit.si/articles/multi-token-prediction https://media.thesatyajit.com/ai/thumbs/multi-token-prediction.jpg?v=1e8c28b3e5ceff24 2026-06-27 monthly 0.7 https://ai.satyajit.si/articles/nous-hermes-moa https://media.thesatyajit.com/ai/thumbs/nous-hermes-moa.jpg?v=35eeb0b0bf0cf1ab 2026-06-27 monthly 0.7 https://ai.satyajit.si/articles/unconventional-un-0 https://media.thesatyajit.com/ai/thumbs/unconventional-un-0.jpg?v=6d1aa2cf2a96f424 2026-06-27 monthly 0.7 https://ai.satyajit.si/articles/glm-5-2 https://media.thesatyajit.com/ai/thumbs/glm-5-2.jpg?v=02a49709df6540e7 2026-06-23 monthly 0.7 https://ai.satyajit.si/articles/sakana-fugu https://media.thesatyajit.com/ai/thumbs/sakana-fugu.jpg?v=10352a9547ac105e 2026-06-23 monthly 0.7 https://ai.satyajit.si/articles/mixture-of-experts-from-scratch https://media.thesatyajit.com/ai/thumbs/mixture-of-experts-from-scratch.jpg?v=fdfb4a9e9d20db4e 2026-06-10 monthly 0.7 https://ai.satyajit.si/articles/coroutines-in-c https://media.thesatyajit.com/ai/thumbs/coroutines-in-c.jpg?v=82c8047ab554531c 2026-06-09 monthly 0.7 https://ai.satyajit.si/articles/how-transformers-attention-works https://media.thesatyajit.com/ai/thumbs/how-transformers-attention-works.jpg?v=37731f3877fc323a 2026-06-02 monthly 0.7 https://ai.satyajit.si/architectures/transformer https://media.thesatyajit.com/ai/thumbs/architectures/transformer.jpg?v=63f18ca4b54f3078 The Transformer, from first principles https://media.thesatyajit.com/ai/films/architectures/transformer-poster.webp?v=0177a132f4b30e1d How a decoder-only Transformer turns token ids into a next-token distribution: embeddings, scaled dot-product attention and why it divides by √d_k… https://media.thesatyajit.com/ai/films/architectures/transformer.mp4?v=396edb64229f1f52 111 2026-10-07 yes 2026-11-12 monthly 0.7 https://ai.satyajit.si/architectures/attention-kv https://media.thesatyajit.com/ai/thumbs/architectures/attention-kv.jpg?v=9dda3a69a2043a15 Attention and the KV cache: MHA, MQA, GQA and MLA, byte by byte https://media.thesatyajit.com/ai/films/architectures/attention-kv-poster.webp?v=6a00f700ccde98dd How the key-value side of attention shrinks to cut the cache a decoder rereads at every step: the 2·n_kv·d_h numbers per token per layer that multi-head… https://media.thesatyajit.com/ai/films/architectures/attention-kv.mp4?v=419b25c06a8a9ec0 103 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/diffusion-transformer https://media.thesatyajit.com/ai/thumbs/architectures/diffusion-transformer.jpg?v=0916e584a61e9d57 The Diffusion Transformer, from first principles https://media.thesatyajit.com/ai/films/architectures/diffusion-transformer-poster.webp?v=e920b0cc21c21596 How a Diffusion Transformer denoises: a VAE's latent space, patches as tokens, the timestep and the prompt steering every block through adaLN-Zero's scale… https://media.thesatyajit.com/ai/films/architectures/diffusion-transformer.mp4?v=5ed6e31cf1368ec4 105 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/encoder-bert https://media.thesatyajit.com/ai/thumbs/architectures/encoder-bert.jpg?v=aa66e7a318b87631 The bidirectional encoder (BERT), from first principles https://media.thesatyajit.com/ai/films/architectures/encoder-bert-poster.webp?v=0bccddbcffa3d4c9 How BERT reads a whole sequence at once: the Transformer encoder with no causal mask, the three summed input embeddings, masked language modelling with its… https://media.thesatyajit.com/ai/films/architectures/encoder-bert.mp4?v=2ac2dc48e5d908fa 112 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/encoder-decoder https://media.thesatyajit.com/ai/thumbs/architectures/encoder-decoder.jpg?v=4f4e066a802ec046 The encoder–decoder Transformer, from first principles https://media.thesatyajit.com/ai/films/architectures/encoder-decoder-poster.webp?v=b7862ce70f8f7530 How a sequence-to-sequence Transformer turns one text into another: an encoder that reads the source with full attention, a decoder with a causal mask and… https://media.thesatyajit.com/ai/films/architectures/encoder-decoder.mp4?v=6f94275f88894b54 116 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/kimi-k3 https://media.thesatyajit.com/ai/thumbs/architectures/kimi-k3.jpg?v=2e1e8381add0804e Kimi K3's hybrid block, from first principles https://media.thesatyajit.com/ai/films/architectures/kimi-k3-poster.webp?v=d61e5ea7c871163e How Kimi K3 mixes information along three axes: Kimi Delta Attention, a gated delta rule derived from linear attention, on 69 of 93 layers; Gated MLA with… https://media.thesatyajit.com/ai/films/architectures/kimi-k3.mp4?v=f794b701628e26ea 111 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/looped-transformer https://media.thesatyajit.com/ai/thumbs/architectures/looped-transformer.jpg?v=470417248c9ce802 Looped Transformers: depth from passes, not parameters https://media.thesatyajit.com/ai/films/architectures/looped-transformer-poster.webp?v=4b31f1db0164adee How a looped, recurrent-depth Transformer runs one block of layers again and again: the prelude, core and coda, why depth then costs FLOPs, latency and KV… https://media.thesatyajit.com/ai/films/architectures/looped-transformer.mp4?v=bcb2e1fe4d646844 110 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/mamba-ssm https://media.thesatyajit.com/ai/thumbs/architectures/mamba-ssm.jpg?v=5a9eca8a1667003b Mamba and selective state spaces, from first principles https://media.thesatyajit.com/ai/films/architectures/mamba-ssm-poster.webp?v=4ddc6f6a27c9c339 How a state space model turns a sequence into a fixed-size state: the continuous SSM and its zero-order-hold discretisation, why S4's time-invariant form is… https://media.thesatyajit.com/ai/films/architectures/mamba-ssm.mp4?v=1b8e7792c79e9b13 115 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/masked-diffusion-lm https://media.thesatyajit.com/ai/thumbs/architectures/masked-diffusion-lm.jpg?v=5ee4f923dfde8a0c Masked-diffusion language models, from first principles https://media.thesatyajit.com/ai/films/architectures/masked-diffusion-lm-poster.webp?v=e4bfcf648ae964c4 How a masked-diffusion language model works: the absorbing [MASK] forward process, a training loss that is a masked-language-model loss weighted by 1/t over… https://media.thesatyajit.com/ai/films/architectures/masked-diffusion-lm.mp4?v=337bfb79668dc96c 110 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/mixture-of-depths https://media.thesatyajit.com/ai/thumbs/architectures/mixture-of-depths.jpg?v=9d52696a623c9e93 Mixture-of-Depths: a router decides which tokens a block computes https://media.thesatyajit.com/ai/films/architectures/mixture-of-depths-poster.webp?v=4974cc0449d0bbc0 How Mixture-of-Depths caps a routed block at a fixed number of tokens: a scalar router per block, top-k over the sequence, the chosen tokens sent through… https://media.thesatyajit.com/ai/films/architectures/mixture-of-depths.mp4?v=35f3cc93a9086054 108 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/mixture-of-experts https://media.thesatyajit.com/ai/thumbs/architectures/mixture-of-experts.jpg?v=35aebc8cba95fb65 Mixture-of-Experts, from the router to the all-to-all https://media.thesatyajit.com/ai/films/architectures/mixture-of-experts-poster.webp?v=5b3d719f34d7e350 How a sparse MoE layer replaces one feed-forward network with many experts and a router: softmax top-k gating, fine-grained and shared experts, why total… https://media.thesatyajit.com/ai/films/architectures/mixture-of-experts.mp4?v=3e1bd7f9bb409f1c 111 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/architectures/rope https://media.thesatyajit.com/ai/thumbs/architectures/rope.jpg?v=ea776b65e9767c8d Rotary position embeddings (RoPE), from first principles https://media.thesatyajit.com/ai/films/architectures/rope-poster.webp?v=2368a9ab59c4f10a How RoPE tells attention where tokens are: why attention needs position at all, absolute versus relative schemes, rotating each pair of query and key… https://media.thesatyajit.com/ai/films/architectures/rope.mp4?v=3a3d91678b226adb 111 2026-10-07 yes 2026-09-26 monthly 0.7 https://ai.satyajit.si/logs/2026-06-03-scaffolding 2026-06-03 https://ai.satyajit.si/projects/vritti 2026-06-09 https://ai.satyajit.si/projects/torch-point-ops 2026-05-18 https://ai.satyajit.si/projects/fabrik 2026-03-23 https://ai.satyajit.si/projects/rentree 2026-03-17 https://ai.satyajit.si/projects/atoms 2026-03-06 https://ai.satyajit.si/projects/botcha 2026-02-23 https://ai.satyajit.si/projects/penora 2026-02-22 https://ai.satyajit.si/projects/hyr 2026-02-20 https://ai.satyajit.si/projects/mocklab 2026-02-13 https://ai.satyajit.si/projects/tapcn 2026-02-13 https://ai.satyajit.si/arxiv/2026-07-22 2026-07-22 https://ai.satyajit.si/arxiv/2026-07-15 2026-07-15 https://ai.satyajit.si/arxiv/2026-07-10 2026-07-10 https://ai.satyajit.si/arxiv/2026-07-03 2026-07-03 https://ai.satyajit.si/arxiv/2026-07-01 2026-07-01 https://ai.satyajit.si/arxiv/2026-06-30 2026-06-30 https://ai.satyajit.si/arxiv/2026-06-28 2026-06-28 https://ai.satyajit.si/arxiv/2026-06-27 2026-06-27 https://ai.satyajit.si/arxiv/2026-06-23 2026-06-23 https://ai.satyajit.si/arxiv/2026-06-10 2026-06-10 https://ai.satyajit.si/arxiv/2026-06-09 2026-06-09 https://ai.satyajit.si/arxiv/2026-06-03 2026-06-03 https://ai.satyajit.si/notes/lidar-odometry 2026-05-28 https://ai.satyajit.si/notes/point-cloud-registration 2026-05-28