Microsoft released a 37-page “humanist AI code of conduct” (Sept 14) arguing models aren’t conscious, shouldn’t seek legal personhood, and must stay under human oversight—pushing back on Anthropic’s openness to model welfare as labs debate pacing the frontier.
Released Oct 25 under Apache 2.0 ( pip install hermit harness ), Hermit runs an RL style loop that evolves an external adapter of instructions, skills, memory and tools while the model and harness stay frozen, then deploys it to Codex, Claude Code or Pi or exports it as a native plugin. Its best reported result is 0.795 on CL Bench database exploration versus a 0.580 in context baseline, but the numbers are self reported, mostly single runs, all on GLM 5.3 Flash…
Released Oct 28, it publishes new filesystem, everything, memory, fetch, git and time packages and moves the fetch Docker image to certifi 2026.10.23, which drops two eMudhra roots after the CA's Oct 20 key disclosure. The TypeScript servers trust your own Node.js, so run 24.22.1 or 26.12.1+ (npm won't warn you), and refresh older fetch installs with uvx refresh mcp server fetch…
Released Oct 28, the benchmark starts each agent on an empty project and has it build one of 17 products (modeled on apps like Uber, GitHub and Google Docs) and then add up to 18 features on its own code, graded by 544 browser test plans; GPT 6.1 Sol, GPT 6 Astra and Opus 5.5 passed every plan in only 49 65% of builds. Eight products are public on GitHub, and results sit on Fireworks' leaderboard, where no open weight model passed a full build…
Released Oct 27 on Hugging Face with its base model, the 25B active hybrid linear MoE got extra post training for multi step tool use and scores 44 on the Artificial Analysis Intelligence Index (up from 41), at $0.30/$0.90 per million tokens. It's a server model (about 570 GB in FP8, eight H200s on SGLang), and Artificial Analysis flags it as very verbose, so real task costs run higher than the token price suggests…
Announced Oct 29, Hud's one line runtime SDK learns each backend function's production baseline, scores a pull request's production risk before merge, then watches the new version live and rolls back or flags a regression with the evidence the agent needs to fix it. Riverside says flawed deploys behind its canary now get reverted in about eight minutes…
Spotted something we missed? Start a thread.