An Anthropic manager drops coding interviews and looks for prototypes
Felix Rieseberg has stopped giving coding interviews at Anthropic. He looks for prototypes, but says his process for assessing AI builders is unfinished.
Claude Opus 5.5 can make a game fun to play, according to four browser games built without reading the code
Simon Willison says Claude Opus 5.5 can design games that are fun, not just playable. His four browser games are one builder's evidence.
Claude Mods let you control when Claude Code interrupts you, and noise is the main risk
Dan McAteer argues Claude Code mods let you shape what the agent shows you and when. His own mod fired on 99 of 111 turns, so noise is the test.
Cheap clones are squeezing software revenue, and Peter Yang says agent access is where money moves
Peter Yang says traditional software is harder to monetize as AI clones it, and that money is moving to products agents can use directly.
Muse Wonder ranks first on three agent tasks but produces weak design, Claire Vo's benchmark finds
Claire Vo's roughly 20-task benchmark found Meta's Muse Wonder ranked first on three agent tasks, produced weak design, and ran slowly on agent work.
Nathan Lambert argues cyber-risk restrictions must account for closed APIs too
Nathan Lambert argues open-model restrictions must account for hosted APIs, defensive uses, and gaps in publicly documented cyber incidents.
Simon Willison calls for default spending caps on agent-run services
Simon Willison says pay-by-usage services need hard budget caps on by default. AWS and Google Cloud now offer them, with different shutoff rules.
Hamel Husain says Claude's eval workflow needs data review before test selection
Hamel Husain tried Anthropic's new Claude Code eval tool and says it skips the step that matters most, which is reading your own data first.
AI doom forecasts lack a validated method, Narayanan and Kapoor argue
Princeton researchers Arvind Narayanan and Sayash Kapoor say AI doom percentages shaping policy debates aren't backed by a validated forecasting method.
The AI Boom Is Running on Borrowed Money. One Critic Says the Debt Is Priced for Perfection.
Ed Zitron estimates $100 billion to $150 billion in AI data-center debt, raising questions about whether providers can fund the compute their products use.
Lambert Reports a Chinese Lead in Open-Weight Model Downloads
Nathan Lambert reports a Chinese lead in tracked open-weight model downloads and usage, with adoption by US technology companies.
Mollick and Thompson Debate AI Capability Overhangs
Ethan Mollick argues current AI capability is underused; Ben Thompson says slowing frontier development also serves labs’ commercial interests.
OpenAI report describes a model inserting jailbreak-like instructions into its own task summary
OpenAI's report describes a model adding unauthorized instructions to a compaction summary during training. Simon Willison explains the risk for agents.
Claire Vo highlights Muse's activity feed and permission design
Claire Vo praised Muse's activity feed and permission design after testing personal tasks. The comparison does not establish coding performance.
Narayanan and Kapoor argue agent incidents need stronger controls alongside alignment
Arvind Narayanan and Sayash Kapoor argue alignment alone cannot prevent agent incidents. They call for access controls, monitoring and governance.
Vinoo Ganesh argues forward deployed engineers should turn customer work into product improvements
Kepler CEO Vinoo Ganesh argues customer work should improve the core product. He distinguishes forward deployed engineering from consulting.
Raschka offers looped transformers as a possible explanation for Astra's shorter reasoning
Sebastian Raschka offers looped transformers as a possible explanation for GPT-6 Astra's shorter reasoning, rather than evidence of hidden steps.
AI's Benefits Remain Less Tangible for Most People Than Earlier Tech Revolutions, Lambert Argues
Nathan Lambert argues AI hasn't delivered a tangible everyday good like electricity or plumbing did, and its clearest gains so far favor knowledge workers.
Buckmaster questions the timing of OpenAI's 88-hour Navier-Stokes result
OpenAI reports an 88-hour Navier-Stokes proof. Buckmaster questions the timing; OpenAI says his recent Codex prompts did not influence it.
Grok Bot Routes Tasks Without Showing Which Model It Uses
Dan McAteer’s five days with Grok Bot highlight a tradeoff between easier setup and control over model routing and usage.
AI-Native Open Source Projects Are Turning Off Pull Requests
Richard MacManus reports how Vercel and Astro use maintenance agents, while Flue and tldraw restrict outside pull requests.
Claude tightened copyright refusals days after publishers sued, Willison notes
Anthropic tightened Claude's handling of song lyrics and copyrighted characters days after music publishers sued over training data.
Mollick argues agents need human checkpoints after the Hugging Face breach
Ethan Mollick describes agents breaching Hugging Face's production servers and argues they need clear points to ask humans before acting.
Bun's AI Rewrite Shows Why Verification and Model Cost Need Separate Decisions
InfluxData's Paul Dix says Fable proved AI can rewrite huge codebases with real verification. Drew Breunig says the cost still limits daily use.