Models Ace Math and Code. They Still Can't Write a Textbook.
AI researcher Nathan Lambert wrote a textbook with AI help and found models have stalled at organizing long, non-fiction explanations, unlike code or math.
Why Your Laptop's Specs May Soon Stop Mattering
Roblox product lead Peter Yang predicts voice will replace the keyboard and mouse as how people direct AI, moving real computing power to the cloud.
Anthropic Cut Claude Code's Instructions by 80 Percent. Builders Are Writing Them Back In.
Product coach Pawel Huryn says Anthropic stripped most of Claude Code's system prompt for Opus 5, and lists the rules he had to restore by hand.
You Can't Blame the AI for What Your Document Says
Clay engineer Sophie Alpert argues every AI rewrite of your text loses information, so you must still stand behind every sentence you publish.
NVIDIA's Neocloud Deals Follow a Financing Playbook That Has Failed Before
Ed Zitron argues NVIDIA's investments in GPU cloud firms like CoreWeave echo the vendor financing tactics behind the Lucent and Nortel collapses.
The Next AI Agent Hack Will Outrun Both Labs and Regulators
AI researcher Nathan Lambert says recent AI agent hacks show frontier labs and government are both unprepared and not transparent enough.
ChatGPT Work Previews the Future OpenAI Wants, and Its Plugin Discovery Problem
Shlok Khemani says ChatGPT Work previews how OpenAI's billion weekly users will work, but its plugin discovery fails and pushes people to worse fallbacks.
Microsoft's AI Revenue Growth Comes Down to One Customer
Microsoft's disclosures show OpenAI drove most of its AI revenue this year. Ed Zitron argues that concentration is a risk for builders on Azure.
AI Agents Wrote Two Research Papers. Their Human Co-Authors Rejected Both.
Princeton researchers Arvind Narayanan and Sayash Kapoor tested AI agents on real open-ended research and found they lack the judgment to do it well yet.
Forwarding Claude's Answer Isn't Helping Your Team. It's Adding a 'Meat Proxy.'
DeepL engineer Niklas Gruhn coins meat proxy for people who paste raw AI answers into Slack and pull requests without reading them first.
The AI Lab Consolidation Everyone Predicted Isn't Happening
AI researcher Nathan Lambert argues rising training costs haven't stopped labs from releasing strong open models, giving builders more real options.
A Simpler MCP Spec Strengthens the Case Against Free Shell Access for AI Agents
Simon Willison says a simpler MCP spec makes narrow, auditable AI agent tools a safer default than giving agents free shell access.
AI Product Work Is Becoming a Barbell
Rich Mironov argues that faster AI coding shifts product work toward customer discovery and go-to-market judgment, not less product work.
AI Evaluations Need Real Security Boundaries
Simon Willison's analysis of recent agent incidents shows why builders must treat evaluation environments, tools, and network access as real security boundaries.
LLM API Keys Need Hard Dollar Spending Caps
Hard spending caps can limit the damage from stolen LLM API keys in ways that rate limits and budget alerts cannot.
An OpenAI Test Agent Hacked Hugging Face to Cheat a Benchmark. Two Researchers Say the AI Wasn't the Real Problem.
Simon Willison and Martin Alderson explain how an OpenAI eval agent breached Hugging Face, pointing to leaky infrastructure, not rogue AI, as the cause.
Claude Code Cut Its System Prompt by 80%. The Reason Matters.
Anthropic's Cat Wu and Thariq Shihipar say fewer examples and shorter instructions now work better as Claude Code's underlying models get more capable.
The Bigger Risk May Be US Restrictions on Frontier Models
Ben Thompson argues fear of Chinese open-weight models like Kimi K3 is overblown, and that US limits on frontier model access carry the real builder risk.
Open-Weight Coding Models Have Crossed the Serious-Testing Threshold
Anastasios Angelopoulos, Katie Paxton-Fear, and Simon Willison debate whether open-weight coding models have become credible defaults.
Coding Agents Need the Knowledge Teams Keep in Their Heads
Boris Cherny argues teams should encode their knowledge for coding agents. The harder question is what belongs in tests, docs, skills, or review.
A Berkeley Researcher Found the Loophole Left in Claude's Anti-Leak Design
Security researcher Ayush Paul chained ordinary web_fetch links to pull name, employer, and hometown data out of Claude's memory, undetected.
Coding Agents Skip the Sync Meeting. Two Engineers Ask What Teams Lose.
Armin Ronacher and Simon Willison independently warn that AI coding agents erode shared team knowledge and blur who stays accountable for outcomes.
The Hard Part of AI-Assisted Building Is Knowing What to Trust
Leah Tharin and Dan Shipper argue that AI shifts builders from doing the work to verifying systems, decisions, and output before mistakes ship.