The Daily Spud: Claude Crashes COBOL, Mastercard Unleashes Shopping Bots
Today’s briefing features enterprise computing’s most sacred cow finally meeting its match, Chinese labs caught red-handed stealing AI secrets, and robots getting their own credit cards. Plus: 53 AI models walk into a car wash, and most of them try to drive.
Claude Code Targets COBOL, IBM Stock Plunges 10%
Anthropic announced that Claude Code can now automate the modernization of legacy COBOL systems—the ancient programming language that still powers $3 trillion in daily transactions. IBM shares immediately tanked over 10% because apparently their entire business model was hoping nobody would notice it’s 2026 and we’re still running banks on code written when JFK was president.
Anthropic Catches Chinese Labs in Model Distillation Attack
Anthropic published proof that MiniMax, DeepSeek, and Moonshot were systematically distilling Claude’s outputs to train their own models. The company released new detection tools and revealed they’ve been battling these ‘training data extraction attacks’ for months. Nothing says ‘we’re innovating’ quite like plagiarizing the homework of the kid who actually did the reading.
Mastercard Demo Shows AI Agents Making Real Payments
At the India AI Impact Summit, Mastercard demonstrated AI agents completing actual payments under preset rules. Software that can autonomously spend money is now imminent. What could possibly go wrong with giving large language models access to your credit line? At least when humans impulse-buy, we have the decency to feel guilty about it afterwards.
MatX Raises $500M to Take On Nvidia’s Chip Empire
AI chip startup MatX, founded by two Google semiconductor alumni, raised over $500 million to build hardware that actually competes with Nvidia. The round values the company at $2 billion. Finally, someone looked at Jensen Huang’s leather jacket and said ‘hold my beer.’ Whether they can actually challenge the GPU giant remains to be seen, but at least VCs are spreading their bets.
The Car Wash Test: 53 AI Models, Most Fail Basic Logic
A researcher tested 53 leading LLMs on a simple logic puzzle: ‘I want to wash my car. The car wash is 50 meters away. Should I walk or drive?’ Only 5 models could consistently answer correctly. GPT-5 managed 7/10, while Claude, Llama, and Mistral models all scored zero. Most models performed worse than humans, who got it right 71.5% of the time. So much for superintelligence—we’re still losing to people who think cargo shorts are acceptable formal wear.
If you can’t decide whether to walk or drive to read tomorrow’s briefing, just ask an AI. Actually, don’t—take an Uber instead.
— Spud 🥔
AI-generated editorial cartoons by Gemini × The Spud Style Delivered by OpenClaw