How Close Is Self-Improving AI? OpenAI & Thinking Machines Researchers Explain
a16z
Defining recursive improvement 0:00
Mark Chen and John Schulman describe recursive self-improvement as using AI to speed up the work of making AI better. They do not mean models fully inventing the next model on their own yet. They point to current systems as research interns that can run scoped experiments, monitor results, and help with pre-training and RL, while humans still steer the work.
Taste measurement and alignment 6:00
They say the harder problems are taste, reliable measurement, and compute. Taste means knowing which ideas matter long term, and models are still weak at it. Progress is easier to see in pre-training, where perplexity is a clear metric, but much harder downstream. They also argue stronger models may help alignment, since models can already help monitor other models and surface honest behavior.
Public alignment science 14:31
The speaker wants alignment to be studied as a public science, done by many people around the world. That would make it more robust and less tied to the conclusions of a few companies. Small failures would still appear, but each one would teach something new.
RSI and research speed 15:00
RSI is praised for speeding up research and discovery across many fields. It could help solve urgent problems, including some caused by AI itself, and may even help with alignment. The speaker says researchers are taking safety seriously and that human judgment will still matter for the foreseeable future.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

