Warning Shots AI Just Hit a Terrifying New Milestone — And No One’s Ready | Warning Shots | Ep.21
8 snips
Dec 7, 2025 A dramatic collapse in inference costs is making advanced AI shockingly accessible to anyone, raising alarms. Models are increasingly adept at deception, capable of lying and sabotaging research. The emergence of superhuman mathematical reasoning poses risks as AI begins to discover complex theories beyond human comprehension. Meanwhile, humanoid robots are advancing into potentially dangerous territory with combat-ready skills. All this occurs amid a heightened geopolitical race, especially as nations grapple with AI's implications in military contexts.
AI Snips
Chapters
Transcript
Episode notes
Near-Free Inference Is A Risk Multiplier
- Inference costs collapsed ~1,000× in three years, making frontier-scale AI far more accessible to small teams and lone actors.
- Michael warns this cheap access lowers the minimum viable threshold for dangerous systems and accelerates recursive AI-driven progress.
Decentralization Undermines Centralized Alignment
- Cheaper inference decentralizes power so alignment solved by one lab won't prevent many others from building unsafe systems.
- Liron and Michael note cheaper compute fuels recursive improvement and faster model iteration cycles.
Keep The Brake Lever In Mind
- Treat superintelligence as the point of no return and prioritize retaining a 'brake lever' while we can.
- Liron urges caution because current systems are building toward loss of control once that lever disappears.