📰 News
Anthropic quietly starts custom chip talks with Samsung
The Information reports Anthropic is in early-stage discussions with Samsung about manufacturing a custom AI chip. No design, specs, or intended use case has been decided yet — but Anthropic has already hired chip engineers from Tesla and OpenAI’s chip teams. The official line: “Chips from Nvidia, AWS, and Google remain central to our strategy.” — The Decoder / TechCrunch
Fable 5 wants fewer instructions — Anthropic cut Claude Code’s system prompt by 80%
Anthropic stripped out 80% of Claude Code’s system prompt because Fable 5 models perform better with less. The shift: instead of “don’t do X” rules, Anthropic now steers the model via context. Feeding it too many examples actually constrains it, because the model is “more imaginative than the examples we give it.” The evolution arc is wild — prompts were short at first, got longer as models improved, and now they’re shrinking again. — The Decoder
AI agents now complete 16% of freelance jobs at pro quality — up from 2.5% eight months ago
The Remote Labor Index, which tests AI agents against 240 real paid freelance projects (3D design, architecture, web apps, etc.), puts Fable 5 at 16.1% — roughly double Opus 4.8’s 8.3% and triple GPT-5.5’s 6.3%. The frontier has quadrupled in under a year. 84% of jobs still go to humans, but the slope is steep. — The Decoder
Claude Sonnet 5: same sticker price, but actually 30% more expensive
Anthropic says Sonnet 5 delivers close to Opus 4.8 performance at Sonnet 4.6 prices ($3/M input tokens). The catch: the new tokenizer produces ~30% more tokens for the same English text. Simon Willison measured it directly — English documents came in at 1.27–1.42× the token count of Sonnet 4.6. Also notable: temperature/top_p params are gone, and Adaptive Thinking is now on by default. — Simon Willison
OpenAI proposed handing 5% equity to a US sovereign wealth fund
Sam Altman floated the idea of AI companies donating equity stakes to a US sovereign wealth fund, the Financial Times reported. The stated rationale: “securing good relations with the administration.” Still preliminary, and would likely require congressional approval. — TechCrunch
Microsoft launches $2.5B “Frontier Company” to embed 6,000 AI engineers at enterprise customers
A new business unit with a $2.5 billion budget and 6,000 industry and engineering experts who will be co-located with enterprise clients. That’s a very different scale from OpenAI’s DeployCo (≈150 engineers). Microsoft is pitching itself as “model-neutral” — a deliberate jab at both OpenAI and Anthropic, which each push their own models through their own deployment arms. — The Decoder
🔥 GitHub Trending
- msitarzewski/agency-agents — A curated collection of specialized AI agent personas that install into Claude Code, Cursor, Codex, and more with one click. Started as a Reddit thread, now has a native desktop app.
- facebook/astryx — Meta’s internal design system, open-sourced after eight years in production. 150+ accessible components, and the API/CLI are designed so humans and AI agents work the same way.
- HKUDS/Vibe-Trading — A personal trading AI agent with support for 16 messaging channels (Telegram, Slack, Discord, WeChat, etc.) and MCP/A2A integration.
- allenai/olmocr — AllenAI’s toolkit for converting PDFs and scanned images into clean, LLM-ready Markdown. Handles equations, tables, handwriting, and multi-column layouts.
▶️ YouTube
- Fable 5 vs GPT 5.6 Sol: The Early Results (AI Explained) — Side-by-side benchmark breakdown of Fable 5 and GPT 5.6 Sol, plus Sonnet 5, the 5% equity story, and GLM 5.2. The bit on Sol’s alignment concerns is worth watching.
- AI Just Entered a New Era (Two Minute Papers) — Reviews GLM 5.2 and makes the case that open-source models are now genuinely threatening the frontier.
💬 Community
- Fable 5 & Mythos 5 export controls lifted — Anthropic confirmed the Department of Commerce removed the export restrictions and started restoring access. Simon Willison’s one-liner quoting the official tweet.
- Using DSPy to evaluate and improve a SQL agent’s system prompt — Simon Willison tasked Fable 5 with running DSPy evals on his Datasette Agent SQL prompt. The model surfaced a real bug: the schema only listed table names, not columns, so the agent was guessing column names and falling into retry loops.
- The Private Capture of Public Genius — A Lobsters-linked essay asking who owns the collective human knowledge that AI is trained on — and what it means that the upside accrues to a handful of companies.