Description

Send us Fan Mail

Copilot, Claude, Webwright, NVIDIA and agent costs

Today’s episode follows AI responsibility as it slides down the stack: default model routing, long-document training, Claude in government networks, agent costs, web-agent scripts, voice models, local hardware, and synthetic bug reports.

1. Copilot and the risk of default model selection
2. ByteDance Seed trains LMMs through question answering
3. Hassabis, LeCun and the intelligence debate
4. Anthropic, Claude and the NSA
5. Claude Code discovers a cheaper reasoning-control algorithm
6. Viral Claude token burn as agent-cost warning
7. Microsoft Research Webwright
8. NVIDIA Gated DeltaNet-2
9. StepFun StepAudio 2.5 Realtime
10. Claude Skills for small businesses
11. Public skepticism about AI and robotics labor economics
12. NVIDIA as default local LLM hardware
13. Cursor, Manus and Starbucks AI
14. Armin Ronacher on AI-rewritten bug reports