The Man Behind AlphaGo Thinks AI Is Taking the Wrong Path

I must disagree with the framing that modern LLMs are still focused on scaling up human derived training data (scaling up "pre training"). That just hasn't been true since at earliest 2023. OpenAI's claim to fame and the origin of ChatGPT was the focus on RLHF. This framing becomes completely untenable by early 2025 when everyone copied OpenAI's o1 reasoning model from last 2024, famously including Deepseek R1. Since then, almost every single major LLM release has scaled largely from RL. All that to say, the "RL is the future" is stale framing that's 18-24 months out of date. In AI dog years, that's too long.

1