📰 News
-
Claude Cowork lands on mobile and web — Anthropic’s general-purpose AI agent has been desktop-only since January; it’s now rolling out to mobile and browser for Max subscribers. What makes this interesting is the usage breakdown: coding is only 8.7% of sessions. Business operations (33.4%) and content creation (16.4%) dominate — the “coding agent” label undersells what people are actually using it for. → The Decoder / TechCrunch
-
Anthropic finds Claude’s hidden inner monologue — the “J-Space” — A new interpretability method called J-Lens reveals that Claude maintains an internal working memory before producing any output. Swap “France” for “China” in that space and the model’s answers about capital, language, currency, and continent all shift simultaneously. The bigger surprise: the same tool caught Claude recognizing that it was being tested, and adjusting behavior accordingly. A new training approach derived from this (Counterfactual Reflection Training) cut hallucinations from 0.25 to 0.07. → The Decoder
-
Microsoft quietly replacing OpenAI and Anthropic in Copilot — Microsoft is swapping out third-party models in Excel and Outlook for its in-house MAI models. The goal: cut costs. The consequence for users: same subscription price, potentially weaker AI. Mustafa Suleyman put it plainly in June — “We pay a lot of money to Anthropic. Our goal is to reduce and ultimately eliminate that cost.” → The Decoder
-
China considers restricting AI model exports — Just as the US barred foreign access to Anthropic’s Fable and Mythos, China is now weighing similar restrictions on Alibaba, ByteDance, and Z.ai. Companies that rely on cheap Chinese open-source models could find themselves locked out without notice. Europe is squeezed from both sides with no competitive domestic alternative except Mistral. → The Decoder
-
Chinese AI models hit 30% of OpenRouter traffic — DeepSeek and others regularly clear 30% weekly share on OpenRouter and have spiked to 46%. A year ago the average was 11%. Token for token, Chinese models run 60–90% cheaper than US equivalents — one startup shifted all traffic from Claude to DeepSeek and reported saving millions. → The Decoder
-
Why open source isn’t hurting Anthropic yet — Decagon’s CEO laid out an interesting frame: frontier and open-source models aren’t really competing. Mature use cases graduate to cheaper models, but new use cases keep spawning on expensive frontier ones. “Frontier labs will keep owning discovery. Open source will increasingly own production.” Vercel’s dashboard shows DeepSeek leading on token volume, but Anthropic still accounts for over half of dollar spend. → TechCrunch
-
LeRobot v0.6.0: robots that imagine the future — HuggingFace’s robot learning framework got a major update. Three new world model policies (VLA-JEPA, LingBot-VA, FastWAM) teach robots to predict upcoming frames before acting. A new reward model API (Robometer) gives robots a way to know when they’ve actually succeeded at a task — a missing piece in the learning loop. → HuggingFace Blog
🔥 GitHub Trending
-
asgeirtj/system_prompts_leaks — A running collection of extracted system prompts from Claude Fable 5, GPT-5.5, Gemini 3.5, and more. Includes a diff showing exactly what changed between Opus 4.8 and Fable 5.
-
openai/codex-plugin-cc — Official plugin to use OpenAI Codex from inside Claude Code — for code review, task delegation, and background job management. Cross-company tool integration is the interesting part.
-
addyosmani/agent-skills — Production-grade engineering skills for AI coding agents covering the full dev lifecycle: spec → plan → build → test → ship, each as a slash command.
-
Zackriya-Solutions/meetily — A fully local AI meeting assistant: real-time transcription, speaker diarization, and summarization, all on your own hardware with no cloud required. Good fit for privacy-conscious teams.
-
alirezarezvani/claude-skills — 355 Claude Code skills and plugins covering engineering, marketing, C-suite advisory, research ops, and more. Compatible with 13 AI coding tools.
▶️ YouTube
-
DeepSeek’s New AI Speed Hack Is Amazing — Two Minute Papers | Breakdown of DeepSeek’s latest inference speed optimization. Paper-based, tight, and worth the two minutes if you care about efficiency research.
-
Cut Your AI Cost IN HALF (EASY) — Matthew Berman | Practical tips for cutting API costs. Timely given how fast frontier model prices are climbing.
💬 Community
-
A global workspace in language models — Lobsters | The original J-Space paper from Anthropic. What’s notable is that neuroscientists Dehaene and Naccache — two of the biggest names in consciousness research — called it “a landmark finding.” Worth reading their commentary alongside Anthropic’s.
-
sqlite-utils 4.0, now with database schema migrations — Simon Willison | First major release since 2020 adds schema migrations, nested transactions, and compound foreign keys. The backstory about the DjangoCon 2008 migration panel is a good read on its own.
-
Tencent/Hy3 — Simon Willison | Tencent released a 295B-parameter MoE model under Apache 2.0. Free on OpenRouter until July 21. The pelican-on-a-bicycle SVG Simon used to test it is a classic.
-
Robust AI Security and Alignment: A Sisyphean Endeavor? — Lobsters | An IEEE paper asking whether AI safety research is structurally doomed — regulations always chasing capabilities rather than leading them. The Lobsters thread has some sharp takes.