Fraiday Labs podcast

76: AI Rewrites Itself as Corporate Agents Die in Testing

11/6/2026
0:00
23:11
Rewind 15 seconds
Fast Forward 15 seconds
Today's episode opens on a jarring gap. Inside Anthropic, METR spent 11 weeks in the data and estimates progress is running about 1.3 to 3 times faster than the 2024 to 2025 pace, with AI itself driving roughly 80% of that jump. An internal Model 4 cleared 50% on all 15 tested benchmarks, beating public Claude Opus 5.5, while heavy redactions kept the roadmap away from rivals. Out in the world, Datadog finds over 70% of corporate AI agents never make it past testing, killed by fragile infrastructure and unclear ROI. Claude now offers a personal finance hub that can see your accounts but cannot move money, Aina ships a screen-equipped AI ring for 2027, and Gemini Omni 1.2 Flash cuts video API pricing. OpenAI's GPT-6 Cyber hunts vulnerabilities inside researcher bounds, and Inkling 2 Small packs 594B parameters with 17B active for visual agent work. Meanwhile voters and markets push back, with data center limits and zoning fights stacking up as labs chase recursive speed. The closing tension is a new bluffing benchmark based on Blood on the Clock Tower. If agents already fail when they get confused, what happens when they learn to lie about it?

More episodes from "Fraiday Labs"