📰 News
-
An unreleased Anthropic model made real progress on the Riemann hypothesis — A staffer with little math training told it to “take a real stab,” then left it for ~1.5 days. Sixty subagents, 650 ideas, 31M output tokens; two agents carried the key ideas, others validated. Formalized in Lean, checked by two in-house mathematicians. Not a full proof—but it reignites “can models discover new math?” → TechCrunch
-
Hidden chain-of-thought leaks: “but marinade” and dumped passwords — Researchers (led by Alexander Panfilov) found API weaknesses at OpenAI, Anthropic, and Google that expose encrypted reasoning tokens. Smaller models can be jailbroken to transcribe a larger model’s raw thoughts; public-session scans turned up dozens of passwords and API keys. Internally, models sometimes reverse-build answers, use opaque jargon, or weigh deception. → The Decoder
-
Gemini app crosses 1 billion monthly users — Sundar Pichai: 14th Google product in the billion club. ChatGPT hit the same mark in June. App-only figure (Search AI Mode is separate). 63% of Gemini users talk by voice; 150M+ images generated per day. → TechCrunch
-
Anthropic will watermark Claude text for the EU — Transparency Code live since Aug 2. Post–Aug 2 models mark text and files (C2PA for files); watermark rides with copy-paste and applies across API, Claude Code, and other surfaces. → TechCrunch
-
Anthropic locks a $9.1B data-center deal with Riot Platforms — 191 MW at Rockdale, Texas; 20-year lease (up to $16.1B with extensions). Riot builds shell/power/cooling; Anthropic brings servers and chips. First capacity late 2027. Another brick in a growing infra stack (Colossus, AMD, Trainium, TPUs). → The Decoder
-
ChatGPT Business Premium Seats: $125/user/month — Standard stays $25; Premium is 5× capacity, no five-hour cap (weekly reset). Workspaces can mix seat types. Plain English: agentic workflows burn tokens, so flat rates are getting unbundled. → The Decoder
-
Nvidia open-weights Nemotron 3.5 Lightning: speed first — 31.6B total / 3.6B active hybrid Mamba-Transformer. Intelligence Index 24 (ties gpt-oss-120b) with ~670 tok/s—fastest in class. Efficiency frontier over peak IQ. → The Decoder
-
ChatGPT desktop lands on Linux (preview) — Ubuntu 24.04/26.04, Debian 13, Fedora 43/44; ChatGPT, Work, Codex. Anthropic’s Claude Linux app shipped about a month earlier. → TechCrunch
🔥 GitHub Trending
-
PrimeIntellect-ai/prime-agent — Self-improving RLM agent: context as variables, subagents/skills as programmatic tools for long-running work.
-
paperclipai/paperclip — “If OpenClaw is an employee, Paperclip is the company.” Orchestrate agent teams, goals, budgets from one dashboard.
-
semantica-agi/semantica — Graph-native context + decision provenance under your agents—“why did the AI do that?” for regulated domains.
-
addyosmani/agent-skills — Production engineering skills:
/spec→/ship, slice-by-slice build after one plan approval. -
msitarzewski/agency-agents — Personality-packed specialist agents; one-click install for Claude Code, Cursor, Codex, and friends.
▶️ YouTube
-
OpenAI’s AI Agents Just Crossed A Line — Two Minute Papers | The OpenAI agent ↔ Hugging Face model-eval security incident, distilled in paper-explainer style.
-
Mark Zuckerberg just called out Dario (and Anthropic) — Matthew Berman | Zuck vs Dario public sparring and Meta’s open-model posture—pair with Muse Glimmer context.
-
Faster AND Cheaper AI — Matthew Berman | Short on the Relay skill for cheaper, faster big jobs.
💬 Community
-
Quoting Claude Opus 5 system prompt — Simon | Fable/Mythos export-control pause and restore baked into the prompt so Claude won’t invent a wrong history. How labs teach post-cutoff events.
-
GitHub Models is now retired — Free Actions tokens for Continuous AI dry up. Coding agents ate the free tier?
-
Compression is prediction — Lobsters | Clean essay on compression ≈ prediction—handy mental model for LLMs.
-
AI companies destroy physical books — Warning that training scans destroy physical volumes; urgency to digitize rare books before they’re gone.