#490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI

An AI can learn from trying a problem repeatedly when its answers can be scored. In this discussion of model training, researcher Nathan Lambert explains reinforcement learning with verifiable rewards: the model generates an answer, a checker grades it, and training uses that feedback to improve subsequent attempts. Math and coding supply prominent examples.

Extending the approach to open-ended work changes the difficulty of grading. Lambert describes using another model to develop a rubric and score answers against it. That can supply feedback where a simple correctness check is unavailable, but the judgment is less definite. The central engineering question becomes how well the scoring method represents a genuinely good result.

“the hype over Anthropic's Claude Opus 4.5 model has been absolutely insane”

Nathan Lambert

“I think that's the beauty behind coding. It doesn't lie. It's math, basically.”

Sebastian Raschka

“If you join a frontier lab and you want to have impact, the best way to do it is just find new data that's better.”

Nathan Lambert

“Like, I think there are ways that you can do this with a language model where, within five minutes of chatting with it, the model just gets you.”

Nathan Lambert

“One issue in society in the future will be: how do you become an expert if you never try to do the thing yourself?”

Sebastian Raschka