AI research wiki
social-stream · 2026-06-11

2026-06-11

Summary

The one story that spanned both slots was Kilo's MiniMax M3 coding plans, pitched as near-Opus-4.8 quality at a tenth of the price and as insurance against vendor churn (Roo Code shutting down, Copilot moving to usage-based billing). Around it the two slots split cleanly: morning was an industry-and-safety feed, evening was the research feed. The morning standout was Anthropic making its invisible Fable 5 throttling visible, a three-account cluster confirming flagged frontier-AI-development requests will now openly fall back to Opus 4.8 and return a refusal reason. The evening standout was a small research cluster from Hugging Face's Elie Bakouch on world-modeling during RL, led by Prime Intellect's "true agents model the world" post combining RL on action tokens with SFT on tool-response tokens. Real secondary signal: Google's compute-bound DiffusionGemma (4x faster generation, open weights), Claude Managed Agents shipping cron deployments plus vault env-vars, and a run of @ns123abc breaking-news on capacity and capital (Bezos' Prometheus at $41B, Google shifting next-gen TPU to Samsung, Crusoe's Wyoming datacenter collapsing). The heavy tail in both slots was noise: long French political threads from @brivael, a Texas criminal-justice thread, and assorted promo.

Posts

  • Kilo launches MiniMax M3 coding plans (@kilocode · blog) [morning + evening] (cluster of 2). Buy external coding-plan subscriptions with existing Kilo credit, starting with MiniMax M3, claimed to bench near Opus 4.8 at a tenth of the price. Framed as insurance against vendor churn after Roo Code shut down and Copilot moved to usage-based billing. See the MiniMax M3 sparse-attention summary.
  • Anthropic makes Fable 5 safeguards visible (@ClaudeDevs · @eliebakouch · @ns123abc) [morning] (cluster of 3). Flagged frontier-LLM-development requests will now visibly fall back to Opus 4.8 and the API will return a refusal reason, an explicit walkback of "invisible safeguards." Bakouch welcomed it and argued it strengthens the case for open models; @ns123abc covered it as "secret sabotage" that Anthropic apologized for. The feed face of today's lead daily digest items.
  • Prime Intellect: true agents model the world (@eliebakouch · blog) [evening]. Combines RL over assistant action tokens with SFT over tool-response tokens (restated as RL with constant positive advantage, no extra cost), letting an agent predict its environment's reaction. Builds on ECHO and PaW; relates to the agentic environment survey.
  • DiffusionGemma: 4x faster text generation, open weights (@tinygrad quoting @sundarpichai · blog) [morning]. Apache-2.0 diffusion LLM that refines all tokens in parallel from a random canvas, making single-user generation compute-bound instead of memory-bound. @tinygrad's comment was a hype check on Google's compute scale.
  • Claude Managed Agents: scheduled deployments + vault env-vars (@ClaudeDevs · blog) [evening] (cluster of 5). Run agents on a cron schedule with a fresh session per run, plus an env-var credential type where secrets are swapped in at the network boundary so Claude never sees them. Apple Foundation Models framework can now call Claude.
  • Grok Build 0.1 solves git-leak-recovery (@kilocode · writeup) [evening] (cluster of 6). On a Terminal-Bench task it ran git fsck, found the orphaned commit, expired the reflog, and pruned: 27 steps, 41 seconds, $0.09 versus a ~30-minute human estimate. Notable for mid-task verification instead of declaring victory early.
  • Speedruns as a testbed for recursive self-improvement (@eliebakouch) [evening]. Bakouch argues Karpathy/Keller Jordan-style training speedruns are a strong environment for improving how models do AI research, with the bottleneck being the environment rather than raw capability.
  • Macrodata Labs: a new data-layer startup (@eliebakouch · repo) [evening]. The FineWeb / FinePDF team launches a multimodal/robotics data company, open-sourcing a large-scale processing framework called Refiner. Bakouch is a small investor.
  • Cursor Bugbot: 3x faster, 22% cheaper, 10% more bugs (@cursor_ai · blog) [morning]. 90% of runs now finish under three minutes, with a pre-push local /review mode that deduplicates against the GitHub/GitLab review.
  • Poetic raises $50M at $500M for hallucination-free automation (@_sholtodouglas) [morning]. Anthropic's Sholto Douglas boosted Markie Wagner's launch: complex multi-hour tasks at 99%+ accuracy and 10x fewer tokens, aimed at high-stakes back-office work like anti-money-laundering. Backed by Kleiner Perkins and Founders Fund.
  • Google Research: auditing machine unlearning and differential privacy (@GoogleResearch) [morning]. Regularized f-Divergence Kernel Tests for verifying a model forgot deleted data, claimed more sensitive to localized data shifts with fewer samples. Relevant to the responsible-AI privacy-audit thread.
  • Bezos' Prometheus raises $12B at $41B (@ns123abc) [evening]. New venture aiming at an "artificial general engineer" for the physical world. Large round, no technical detail yet.
  • Google reportedly shifting next-gen TPU to Samsung (@ns123abc · The Information) [evening]. Talks to fab part of its "Icefish" TPU at Samsung because TSMC is overbooked by Nvidia, alongside Tesla AI6 and an Nvidia LPU. Capacity-allocation signal worth tracking.
  • Crusoe Wyoming datacenter reportedly collapses (@ns123abc) [evening]. Per Bloomberg, Google (the main customer) walked over cost and timeline, with Oracle and OpenAI declining to expand with Crusoe in Texas. Crusoe called it a "pause."
  • NVIDIA GTC Taipei news cycle (@nvidia) [morning]. Stanford's 248-Hopper "Marlowe" SuperPOD, Vera CPUs powering NYSE infrastructure, and an AI Podcast with Mistral CTO Timothée Lacroix. Product and partnership signal, no research substance.
  • OpenAI weighing "drastic" price cuts and Stargate questions (@ns123abc) [morning]. OpenAI reportedly mulling price cuts to win the user war with Anthropic, and a claim the $500B Stargate Ohio figure is a full-buildout estimate, not a commitment. Rumor-grade but consistent with the inference price-war theme.
  • Anthropic CEO on scaling to ASI (@ns123abc) [evening]. Amodei quoted saying one to two more years of scaling laws holding would unlock ASI. Commentary, no new substance.
  • Mermaid + LaTeX rendering in-editor (@theskory) [evening]. A coding tool now renders Mermaid diagrams and LaTeX with open-as-image and copy-raw. Minor tooling note.
  • Opaque X long-form repost (@magicsilicon) [morning]. Linked an x.com/i/article on parallels between jetliners and transistors; body unfetchable (expired cookies). Click through to read.
  • Jensen Huang interview promo (@nvidia) [evening]. Jensen with Condoleezza Rice on Hoover's "Only in America." Skip.
  • Personal and political posts (@brivael · @AustinJustice · @bcherny · @lynnmartin) [morning + evening] (cluster of ~25). French political/economic commentary, a Texas criminal-justice thread, a Tokyo greeting, and a Knicks post. No AI relevance. Skip.