~/satyajit

Changelog

What shipped, and when — straight from git history.

2026-07-24

  • 9133dfeUpdate Nanbeige4.2-3B article with the technical report
  • e41a2ddAdd 8 more architecture diagrams; show only slugs with diagrams
  • 3063074Curate /architectures to unique model architectures with diagrams only
  • bc63284Fix architecture diagram modal clipping: portal to body, proper scroll, above nav
  • fa0a781Clean up all architecture diagrams: fix overlaps, arrows-over-labels, and cramped wiring
  • 89dda20Fix Kimi K3 diagram: move backbone title off the Block n-1 box, fix caption spacing
  • ccc4db1Fix Ling architecture diagram: Q/K/V/α/β labels no longer overlap the inlet arrows
  • c2a3ea1Fix missing space in Ling architecture diagram caption (KDA (a…)
  • 9b63f76Add flagship article: SANA Video2 (NVIDIA single-GPU 720p video model)
  • 09ea337Make /architectures a diagram-first interactive grid; add ⌘K search loading indicator
  • d1b7108Add flagship article: FLUX 3 (Black Forest Labs unified image+video model)
  • f45f45aUpgrade transformer-block diagram with reference-style annotation callouts

2026-07-22

  • 1b7c9e5Add Nanbeige4.2-3B article
  • 7653547Add Antares article (Cisco Foundation AI vulnerability localization)

2026-07-21

  • 1004b42Add Laguna S 2.1 article
  • 90eac7aAdd Qwen-Image-3.0 article
  • 3be1409Agent: GPT-5.4 model router + fallback, and a rate limiter
  • dfe6493Ignore the agent scratchpad working directory
  • 951db92Switch site agent from Gemini to OpenAI GPT-5.6 Luna
  • e35e887Add stylometric similarity heatmap to /models
  • 830aad7Add LongStraw article: million-token GRPO on a fixed GPU budget
  • 03c256dAdd article: the harness as compositional generalizer

2026-07-18

  • b36d6adAdd a pull request template
  • ae66e6fAdd LIMSSR article (missing-modality action-quality scoring)
  • 848c563Add Monolith 1.0 article (Basalt Labs 1.6T open MoE)

2026-07-17

  • 11a7180Add Kimi K3 flagship + 4 model/paper articles, and a LOTUS network diagram

2026-07-16

  • 5694340Inkling: call out the architecture surprises (embedding RMSNorm, kernel-4 convs, relative bias)

2026-07-15

  • 0e8fc6eRename arXiv digest to 2026-07-15 (its actual authoring date)
  • 76cd254Add GRAPE explainer (arXiv 2512.07805): position encodings as one group action
  • d03dda3Fix article dates to reflect when each was actually written

2026-07-13

  • 248b3c2Add Mach-Mind-4-Flash explainer (arXiv 2607.09375): 100B-class scores from a 35B/3B-active MoE, by post-training alone

2026-07-11

  • 3d9584dfetch-arxiv: dedup against existing digests + anchor date window to arXiv's newest
  • 5d3f82earXiv digest 2026-07-10: 6 papers (speculative decoding, quant honesty, low-bit, few-step video, deformable SLAM, anytime LiDAR)

2026-07-10

  • 466528cArticles: add Featured filter toggle + solid gold star badge
  • ad47efaArticles list: sort strictly by date, keep ★ as a non-positional badge

2026-07-08

  • 6ee0a13Add 4 RL/optimizer articles and a featured-star for articles
  • 02f9f2dAdd 10 flagship articles, per-article citation, and Laguna AutoMixer deep-dive

37 changes · every entry shipped as a PR, most authored by the Claude crew.