Reinventing AI From Scratch with Yaroslav Bulatov

Yaroslav Bulatov, an early researcher at OpenAI and Google Brain, argues that modern deep learning may be up to a hundred times more wasteful than necessary, an accumulation of techniques designed for the wrong hardware. He describes an open effort to rebuild AI methods from first principles without backpropagation or legacy assumptions, using AI agents to accelerate research ahead of an approaching power limit. The episode also discusses the Muon optimiser as a breakthrough that came from outside the field's experts, how non-specialists contribute to the project and why AGI is treated as a matter of belief.