Zvi Mowshowitz on Longer Timelines, RL-induced Doom, and Why China is Refusing H20s

Zvi Mowshowitz returns to The Cognitive Revolution to discuss the state of AI as of late 2025. He notes that timelines to AGI have modestly lengthened due to the lack of paradigm-shifting developments, despite on-trend progress in areas like IMO gold medals and GPT-5. Zvi expresses increased concern about AI safety due to the U.S. government's capture by commercial interests, particularly NVIDIA, and increasing evidence that reinforcement learning negatively impacts alignment. He also analyzes the unique, durable alignment of CLAWB3 Opus, and cautions against suppressing unwanted behaviors in a model's chain of thought, as it risks teaching models to hide their true reasoning. Zvi touches on the implications of Beijing's refusal of H20s and the potential advantages of XAI's access to challenging problems from Elon's other companies.