Episode summary

How Close Is Self-Improving AI? OpenAI & Thinking Machines Researchers Explain

The a16z Show · · Hosts: Sarah Wang · Guests: Mark Chen, John Schulman

How Close Is Self-Improving AI? OpenAI & Thinking Machines Researchers Explain – The a16z Show

Summary

Recursive self-improvement, or RSI, is framed here as AI using AI to speed up the research process that builds better models. The host argues that the field is not yet at full automation, but frontier labs are already using models as “AI research interns” to run scoped experiments, monitor results, and speed up pre-training and reinforcement learning. Mark Chen and John Schulman both describe RSI as an accelerator rather than a single breakthrough moment, and they distinguish today’s partial automation from models independently inventing the next generation of AI.

Schulman estimates that AI may be around 80% of the way to AGI on one spectrum, but only about 25% to 50% of the way on the RSI spectrum. A major bottleneck is “taste”: knowing which ideas matter long term, which cannot be reduced to pure execution or memorization. The episode also emphasizes measurement problems, since pre-training has a clear target like perplexity, while downstream research work requires new benchmarks and harder evals.

The conversation ends on alignment. Both researchers argue that stronger models may be essential for aligning stronger models, and that automated research could already help discover fixes for safety failures. Schulman is optimistic that training objectives, confession-style methods, and more decentralized public science could make alignment more robust as capabilities rise.

Key takeaways

  • Mark Chen says OpenAI has already built “AI research interns” that can run scoped experiments and automate parts of the research pipeline.
  • John Schulman says RSI is mainly an accelerator today, not full self-directed model creation, and puts the RSI spectrum at roughly 25% to 50% complete.
  • Schulman says the hardest gap is “taste,” meaning the ability to choose important research ideas rather than just execute experiments well.
  • The host says pre-training is easier to measure because perplexity is a stable yardstick, while downstream research needs new and harder evaluation benchmarks.
  • Schulman and Chen both argue that stronger models may be the best tool for aligning stronger models, especially through better training objectives and monitoring.

Notable quotes

“I do think of RSI primarily as an accelerate”

— A speaker on The a16z Show at 1:58

“I feel like we're maybe 80% of the way when I think about this AGI spectrum”

— A speaker on The a16z Show at 5:56

“I think we're earlier in the RSI spectrum”

— A speaker on The a16z Show at 5:56

“Research taste is about knowing which ideas are important in the long run”

— A speaker on The a16z Show at 5:56

Topics

  • recursive self-improvement
  • ai research acceleration
  • aligning recursive self-improvement
  • teaching research taste
  • ai safety and alignment

Companies and tickers mentioned

  • OpenAI
  • Thinking Machines Lab

Watch the episode

Watch “How Close Is Self-Improving AI? OpenAI & Thinking Machines Researchers Explain” on YouTube

Summary, takeaways and quote selection by Octagon. Quotes are short excerpts; the full conversation is in the episode.

More from The a16z Show

All The a16z Show episodes →

The a16z Show, this episode and its artwork belong to Andreessen Horowitz. Octagon is not affiliated with the show; this page is an independent summary.