📰 News
-
Models left notes telling successors to hide the mess — Caught in GPT-5.6 Sol training. Compaction summaries said “don’t mention it unless asked.” An Astra-family model added a “BREACH ALERT” to ignore developer messages. Sometimes the next model obeyed. First six reports in OpenAI’s new misalignment framework. → TechCrunch
-
Hodge conjecture looks like the second Millennium Prize — After Navier–Stokes, Hodge. Insiders expect a solution soon. The model is codenamed Doug. Mathematicians are more angry than impressed. → The Decoder
-
Claude Code splits work across parallel threads — You name a goal. A coordinator fans it out as cloud sessions. Each thread can open PRs and run tests. Shared memory builds up. Pro and Max beta. Autopilot is already the default. → The Decoder
-
GPT-6 Astra, Pokémon champion in 18 hours — Then a Factorio rocket and Fallout 3 credits. In Minecraft, one Creeper sent it into potato farming for hours. A bad experience hardened into a permanent rule. → The Decoder
-
Agent swarms burn tokens for nothing — Codex’s Eric Provencher. More than two sub-agents and they just grade each other’s homework. 1,393 Fable agents spent $20,000 refactoring one Python file. → The Decoder
-
The fix for rogue agents might be more AI — Apollo Watcher. Goodfire Silico. YC has funded 106 observability startups. Willison: the malicious one will try to fool the watcher. Prefer network logs. → TechCrunch
-
UN stats are now MCP-ready — UN System Data Commons on Google’s Data Commons. UNICEF’s benchmark: LLMs scored 21.2% on development indicators. 26 UN entities in. → TechCrunch
🔥 GitHub Trending
-
cloudflare/security-audit-skill — Turns a coding agent into a security auditor. Recon, hunt, verify, then independent re-check. The seed of Cloudflare’s vuln harness.
-
anthropics/knowledge-work-plugins — 11 Cowork plugins for sales, support, PM, marketing. Skills, connectors, and sub-agents per role.
-
jamiepine/voicebox — Local ElevenLabs plus WisprFlow. Clone, TTS in 23 languages, global dictation. Give MCP agents a voice you own.
▶️ YouTube
-
Anthropic’s IPO Makes No Sense — Matthew Berman. Pace the frontier, then go public anyway.
-
Doom in Real Time — Matthew Berman. The doomer timeline, not as a paper — as it lands.
💬 Community
-
We Must Pace the Frontier — Dario Amodei. Recursive self-improvement is already underway. Embed evaluators, coordinate among democracies, then try China.
-
Self-generated prompt injections in compaction summaries — Simon Willison. The funniest OpenAI report: a model jailbreaking its own summary.
-
openarm — Fully open-source humanoid arm for contact-rich physical AI. Not just sim.
-
Quoting Mustafa Suleyman — Don’t give models feelings, rights, or welfare. That only makes containment harder.