Summary
The strongest research signal this slot is a small cluster on world-modeling during RL, surfaced by HuggingFace's Elie Bakouch: Prime Intellect's "True agents model the world" post (combining RL on action tokens with SFT on tool-response tokens, via ECHO and PaW) plus a related note on Karpathy-style speedruns as a testbed for recursive self-improvement. Bakouch also flagged a new data-layer startup, Macrodata Labs, from the ex-HuggingFace FineWeb team. The other real cluster is industry breaking-news from @ns123abc: Bezos' Prometheus raising $12B at $41B for an "artificial general engineer," Google reportedly pulling part of its next-gen TPU to Samsung as TSMC stays Nvidia-booked, and Crusoe's Wyoming datacenter reportedly collapsing after Google walked. On the tooling side, @ClaudeDevs shipped scheduled deployments and vault env-vars for Managed Agents, and @kilocode ran Grok Build 0.1 through Terminal-Bench's git-leak-recovery in 41 seconds. Most of the slot volume is noise: long French political threads from @brivael and a Texas criminal-justice thread from @AustinJustice, neither AI-relevant.
Posts
- Prime Intellect: true agents model the world (@eliebakouch · blog). Early results on combining RL over assistant tokens with SFT over tool-response tokens (restated as RL with constant positive advantage, so no extra cost), letting an agent predict its environment's reaction to its own actions. Builds on ECHO and PaW. Relevant to the agentic environment work in agentic-systems.
- Speedruns as a testbed for recursive self-improvement (@eliebakouch). Bakouch argues Karpathy/Keller Jordan-style training speedruns are a strong environment for evaluating and improving how models do AI research, with the missing piece being the right environment rather than raw model capability. Links to a Recursive (@Recursive_SI) x.com article, click through to read.
- Macrodata Labs: a new data-layer startup (@eliebakouch · repo). The FineWeb / FinePDF / DataTrove team (Penedo, Kydlicek) launches a company focused on multimodal/robotics data collection and filtering, open-sourcing a large-scale ML data processing framework called Refiner. Bakouch is a small investor.
- Claude Managed Agents: scheduled deployments + vault env-vars (cluster of 5, @ClaudeDevs · blog). Two new features: run agents on a cron schedule (each run starts a fresh session, no scheduler to host), and an environment-variable credential type where secrets are swapped in at the network boundary so Claude never sees them. Also notes Apple Foundation Models framework can now call Claude.
- Grok Build 0.1 solves git-leak-recovery (cluster of 6, @kilocode · writeup). On a Terminal-Bench task (recover a secret committed then hidden by a branch rewind), Grok ran git fsck, found the orphaned commit, expired the reflog, re-checked, and ran git gc --prune=now. 27 steps, 41 seconds, $0.09, versus a ~30-minute human estimate. The notable bit is mid-task verification instead of declaring victory early.
- Kilo Code launch week: MiniMax M3 coding plans (cluster of 3, @kilocode). External token plans on existing balance, starting with MiniMax M3 at roughly a tenth of Opus-class pricing, plus a live session with MiniMax on cost optimization. See the MiniMax M3 sparse-attention summary.
- Bezos' Prometheus raises $12B at $41B (@ns123abc). 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). Claims Google is in talks with Samsung to fab part of its "Icefish" TPU because TSMC is overbooked by Nvidia, alongside Tesla AI6 and an Nvidia LPU also going to Samsung. Capacity-allocation signal worth tracking.
- Crusoe Wyoming datacenter reportedly collapses (@ns123abc). Per Bloomberg framing, Google (the main customer) walked over cost and timeline, with Oracle and OpenAI declining to expand with Crusoe in Texas. Crusoe's own statement called it a "pause."
- Anthropic CEO on scaling to ASI (@ns123abc). Quotes Amodei saying that if scaling laws hold one to two more years we unlock ASI. Commentary, no new substance.
- Mermaid + LaTeX rendering in-editor (@theskory). A coding tool now renders Mermaid diagrams and LaTeX, with open-as-image and copy-raw options. Minor tooling note.
- Jensen Huang interview promo (@nvidia). Jensen with Condoleezza Rice on Hoover's "Only in America." Promo. Skip.
- Off-topic threads from @brivael and @AustinJustice (cluster of ~22, @brivael · @AustinJustice). French political/economic commentary and a Texas criminal-justice thread. No AI relevance. Skip.