Evening brief

Monday, July 6

2 people · 0 companies · 5 topics

TL;DR natolambert argues GLM-5.2 is the open-agent “DeepSeek moment” and co-drops TMax, an open RL recipe hitting 27% Terminal-Bench 2.0 at 9B params; simonw documents Claude Fable writing most of sqlite-utils 4.0rc2 for $149.

People

  • natolambert posted “GLM-5.2 is the step change for open agents” on Interconnects — first open model competing with frontier proprietary on agentic coding, lag from US closed labs now ~6-8 months. post · tweet

  • natolambert co-authored TMax (arXiv 2606.23321) from Ai2: open RL recipe for terminal agents, 27% on Terminal-Bench 2.0 with 9B params, open dataset 2.5x larger than prior work. substack · tweet · paper

  • simonw wrote up Claude Fable authoring most of sqlite-utils 4.0rc2 over 37 prompts and 34 commits (+1,321/-190 lines, 30 files) for $149.25 — Fable independently flagged 5 release-blockers he had not caught. post · HN

Tools & dev

  • openai/codex-plugin-cc (+1,532 stars today) is OpenAI’s official Claude Code plugin — lets you call Codex from inside a Claude Code session for code review, background delegation, and session handoffs. Interop between the two main agentic coding runtimes is now official. repo

Models & research

  • GPT-5.6 Sol/Terra/Luna preview details confirmed: Sol tops Terminal-Bench 2.1, Terra matches GPT-5.5 at roughly half the cost, Luna is the cheapest tier — still limited to ~20 government-cleared partners. source

  • Grok 5 still not coming in Q3; Polymarket closed its June 30 contracts at 3% probability of release. Colossus 2 at 1.5 GW is still training it. source

OSS

  • meetily (+1,409 stars today) — privacy-first AI meeting assistant, local Parakeet/Whisper transcription in Rust, no cloud uploads. repo

  • strix (+1,114 today) — open-source AI penetration testing tool, now at 37k stars. repo

  • system_prompts_leaks (+981 today) — extracted system prompts from Claude, ChatGPT, Gemini, Grok and others in one public repo, trending at 50k stars. repo