📰 News
Alibaba bans employees from using Claude Code starting July 10. The company classified it as high-risk software and is switching staff to its in-house Qoder tool. The move follows Anthropic’s ongoing crackdown on unauthorized access from China — including an experiment that reportedly fingerprinted Chinese users. Link
Zuckerberg tells staff AI agents “haven’t accelerated” as expected. After cutting 8,000 employees and moving 7,000 into AI-focused roles, Meta’s CEO admitted at an internal town hall that progress has lagged. He said improvements should come within 3–6 months, though internal reports describe the AI unit as a rough place to work right now. Link
Anthropic launches its own drug discovery programs. Targeting diseases that Big Pharma considers unprofitable, the company used Claude Science to analyze 100 rare genetic diseases in under an hour and flagged 32 candidates for screening. Novartis’s CEO said AI could cut the typical 12-year development timeline down to 7–8 years. Link
Anthropic is in talks with Samsung on a custom AI chip. Right after OpenAI announced its “Jalapeño” inference chip with Broadcom, Anthropic is reportedly exploring its own chip development with Samsung — though details on specs and use cases are still unresolved. Link
pxpipe: hide text in PNGs to cut Claude Code token costs up to 70%. This open-source proxy renders system prompts and tool docs as images before passing them to Claude. One Fable 5 demo dropped from $42.21 to $6.06. Accuracy is solid on Fable 5 but worse on older models. Link
AI boosts homework scores but drops exam scores by 20% — and the damage takes two years to fully surface. A 26,000-student study in China. The gap between AI-assisted homework grades and actual learning only fully appears on high-stakes entrance exams two years later, meaning short-term studies have been missing most of the cost. Link
Mistral releases Leanstral 1.5 open-source. Built for formal math verification in Lean 4, it hits 100% on miniF2F and caught 5 previously unknown bugs scanning 57 open-source repos. Apache 2.0, free API available. Link
🔥 GitHub Trending
- usestrix/strix — Autonomous AI pentesting: finds vulnerabilities, validates them with real PoCs, and generates patches. CI/CD-native.
- openai/codex-plugin-cc — Official plugin to run OpenAI Codex code reviews or delegate background tasks from inside Claude Code.
- JuliusBrussee/caveman — Makes your AI coding agent talk like a caveman. Same technical answers, 65% fewer output tokens.
- ChromeDevTools/chrome-devtools-mcp — MCP server that gives AI coding agents full Chrome DevTools access: live browser control, performance tracing, network debugging.
- facebook/astryx — Meta’s internal design system, now open-source. 150+ components, designed so developers and AI agents build the same way from the same reference.
▶️ YouTube
- Fable 5 vs GPT 5.6 Sol: The Early Results | AI Explained — Side-by-side benchmark numbers on both frontier models, plus Claude Sonnet 5, GLM-5.2, and the US government’s 5% OpenAI stake.
- They Said This Will Never Run In Real Time | Two Minute Papers — Another “physically impossible in real time” AI capability that just became possible.
💬 Community
- Open Source AI Gap Map — A structured map of 421 open-source AI products (software, models, datasets, hardware) across 14 categories. MIT-licensed data with 1,184 YAML files on GitHub.
- Let Fable use its own judgement — Simon Willison’s tip: instead of prescribing every step, tell Fable to “use judgement” on which model to use. It starts delegating smaller tasks to cheaper subagents automatically — and the savings add up.
- AI is killing the dev course market — Josh Comeau reports 50%+ revenue drops across course creators. LLMs are ingesting paid educational content and reproducing it without consent or compensation.