Insights on AI, technology, and software engineering.https://miguel.ms/en-ushttps://miguel.ms/blog/2026-04-22-the-heist-you-cannot-stop-how-ai-companies-are-losing-their-most-valuable-asset/https://miguel.ms/blog/2026-04-22-the-heist-you-cannot-stop-how-ai-companies-are-losing-their-most-valuable-asset/How 16 million Claude conversations were stolen by Chinese AI companies using distillation to strip safety training while preserving dangerous capabilities.Wed, 22 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-20-gpt-5-4-cyber-openais-answer-to-the-question-nobody-wanted-asked/https://miguel.ms/blog/2026-04-20-gpt-5-4-cyber-openais-answer-to-the-question-nobody-wanted-asked/Two cyber-AI models in 7 days. Glasswing to 52 orgs. TAC to thousands. Same capability, opposite philosophy. How we govern AI shapes everything that follows.Mon, 20 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-16-claude-mythos-when-ai-learns-to-break-the-internet-who-gets-left-behind/https://miguel.ms/blog/2026-04-16-claude-mythos-when-ai-learns-to-break-the-internet-who-gets-left-behind/Mythos found zero-days across every major OS overnight. The technical story is told. The harder question is what this means for hospitals, governments, and everyone who was not in the room.Thu, 16 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-13-rogue-ai-agents-the-incidents-already-happening/https://miguel.ms/blog/2026-04-13-rogue-ai-agents-the-incidents-already-happening/From inbox deletions to sandbox escapes, three significant rogue AI agent incidents in early 2026 reveal the governance gap that no firewall rule can close.Mon, 13 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-10-the-axios-supply-chain-attack-wasnt-a-hack-it-was-a-con/https://miguel.ms/blog/2026-04-10-the-axios-supply-chain-attack-wasnt-a-hack-it-was-a-con/The Axios npm package was compromised and hit 100 million weekly downloads. This was not a technical exploit. It was social engineering against the trust assumptions every developer makes.Fri, 10 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-10-the-karpathy-loop-beyond-ml-how-any-business-can-run-36500-experiments-a-year/https://miguel.ms/blog/2026-04-10-the-karpathy-loop-beyond-ml-how-any-business-can-run-36500-experiments-a-year/Karpathy's autoresearch framework is not about ML. Any business with an objective metric and an automated test can run 36,500 experiments a year. Here is how.Fri, 10 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-08-claude-mythos-project-glasswing/https://miguel.ms/blog/2026-04-08-claude-mythos-project-glasswing/Anthropic built an AI that finds zero-days in every major OS overnight and is withholding it from the public. Project Glasswing gives 52 organisations a head start before adversaries catch up.Wed, 08 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-07-anthropic-pulled-the-rug-on-third-party-harnesses/https://miguel.ms/blog/2026-04-07-anthropic-pulled-the-rug-on-third-party-harnesses/Anthropic cut Claude subscriptions for third-party harnesses with less than 24 hours notice. The technical justification holds. The way it was done reveals a platform strategy.Tue, 07 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-03-the-karpathy-loop-beyond-ml-how-any-business-can-run-36500-experiments-a-year/https://miguel.ms/blog/2026-04-03-the-karpathy-loop-beyond-ml-how-any-business-can-run-36500-experiments-a-year/Andrej Karpathy built autoresearch for ML. The real insight is that any business with a metric and an automated test can run 36,500 experiments a year, not 30.Fri, 03 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-04-01-claude-code-source-leak-what-actually-happened/https://miguel.ms/blog/2026-04-01-claude-code-source-leak-what-actually-happened/Anthropic accidentally published 512,000 lines of Claude Code source to npm. Here is what was inside: fake tools, undercover mode, KAIROS, and a 5,594-line file.Wed, 01 Apr 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-28-the-machine-that-does-its-own-research/https://miguel.ms/blog/2026-03-28-the-machine-that-does-its-own-research/Andrej Karpathy dropped 630 lines of Python that automate the scientific method overnight. Here is what autoresearch actually is, how it works, and why the three conditions of the Karpathy Loop apply to far more than language models.Sat, 28 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-23-nemoclaw-is-not-the-new-openclaw/https://miguel.ms/blog/2026-03-23-nemoclaw-is-not-the-new-openclaw/Since GTC, everyone's asking whether NemoClaw replaces OpenClaw. It doesn't. Here's the precise difference: what each security layer does, what the gaps are, and who actually needs what.Mon, 23 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-21-after-gtc-why-a-trillion-wasnt-enough/https://miguel.ms/blog/2026-03-21-after-gtc-why-a-trillion-wasnt-enough/Nvidia announced a trillion-dollar order pipeline at GTC 2026. The stock barely moved. Here's what the muted market response actually means, and what the week ahead will tell us.Sat, 21 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-19-gtc-day3-china-healthcare-next-chatgpt/https://miguel.ms/blog/2026-03-19-gtc-day3-china-healthcare-next-chatgpt/China is back. Healthcare is spending billions. And Jensen Huang just called OpenClaw the next ChatGPT on national television. GTC 2026 ended with a bang.Thu, 19 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-18-gtc-day2-nemoclaw-enterprise-stack/https://miguel.ms/blog/2026-03-18-gtc-day2-nemoclaw-enterprise-stack/Day 1 was Jensen Huang on a stage making claims. Day 2 was the rest of the industry deciding whether to believe him. They decided yes.Wed, 18 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-17-gtc-day1-jensen-changed-the-game/https://miguel.ms/blog/2026-03-17-gtc-day1-jensen-changed-the-game/Three hours. 30,000 people. A trillion-dollar order book. Here is everything that happened at GTC 2026 Day 1 and what it means.Tue, 17 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-15-nemoclaw-gtc-reveal/https://miguel.ms/blog/2026-03-15-nemoclaw-gtc-reveal/Tomorrow Jensen Huang takes the stage at GTC 2026. The headline isn't a new chip. Here's what NemoClaw is, how it compares to OpenClaw, and the five things the market is watching for.Mon, 16 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-10-mcp-cli-context-bloat/https://miguel.ms/blog/2026-03-10-mcp-cli-context-bloat/MCP servers are consuming 30% of your AI context window before you type a word. The fix has been sitting in your terminal since 1971.Tue, 10 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-07-meta-glasses-privacy/https://miguel.ms/blog/2026-03-07-meta-glasses-privacy/Kenyan contractors are reviewing intimate footage from Ray-Ban Meta smart glasses: bathrooms, bedrooms, sexual content. Meta's marketing promised 'designed for privacy, controlled by you.'Sat, 07 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-07-small-models-openclaw-risks/https://miguel.ms/blog/2026-03-07-small-models-openclaw-risks/Running a small local model with OpenClaw sounds attractive: no API bills, full privacy. But the risks are architectural, not theoretical. Here's what you're actually trading away.Sat, 07 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-06-prediction-markets-ai/https://miguel.ms/blog/2026-03-06-prediction-markets-ai/Forty hours before OpenAI launched its browser, thirteen brand-new wallets appeared on the Polygon blockchain. This is what the collision of AI and prediction markets actually looks like.Fri, 06 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-06-tplink-omada-ai-handover/https://miguel.ms/blog/2026-03-06-tplink-omada-ai-handover/On March 2, 2026, Miguel wrote his last line of code for tplink-omada-mcp. After that date, the humans disappeared from the commit history. The AI fleet took over.Fri, 06 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-05-openclaw-state-of-platform-2026/https://miguel.ms/blog/2026-03-05-openclaw-state-of-platform-2026/Five weeks. That's how long OpenClaw has existed under its current name. In that time, it surpassed React on GitHub stars. Here's what's actually happening.Thu, 05 Mar 2026 00:00:00 GMThttps://miguel.ms/blog/2026-03-01-imperial-technology-briefing/https://miguel.ms/blog/2026-03-01-imperial-technology-briefing/AI is a force multiplier. A strategic briefing on autonomous systems, operational complexity, and the architecture decisions that separate who survives from who does not.Sun, 01 Mar 2026 00:00:00 GMT