Can AI Learn to Improve Itself? | Recursive Self-Improvement
Can AI Learn to Improve Itself? | Recursive Self-Improvement
Podcast17 min 43 sec
Listen to Episode
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights
  • Treat AI research automation as a long-term theme, and favor companies that demonstrate measurable improvements in research outcomes—not just more experiments.
  • Monitor AI compute and data centers for continued infrastructure demand, while treating AI-driven acceleration of future data-center construction as speculative.
  • Track AI alignment and safety research as model capabilities grow, but do not assume alignment is solved; the discussion names no public tickers or specific buy recommendations.
Detailed Analysis

AI Research Automation / Recursive Self-Improvement (Investment Theme)

  • The discussion presents AI-assisted research as a potential accelerant for developing more capable models. Today, AI can run scoped experiments and automate parts of research, but humans still provide much of the direction and judgment.
  • The speakers describe current progress as early on the path toward fully automated AI research. One cited estimate found AI research progress had accelerated roughly 2x over the prior year, though the transcript says this was most likely driven by human researchers using AI.
  • AI’s ability to identify promising research ideas—often called research “taste”—remains a limitation. The transcript also emphasizes that human experience and judgment will matter for the foreseeable future.

Takeaways

  • Treat AI research automation as a long-term investment theme, not evidence that fully autonomous AI development has arrived.
  • When assessing companies in this area, look for evidence that AI tools improve measurable research outcomes, rather than simply completing more experiments.
  • The transcript gives no public tickers, price targets, or specific buy/sell recommendations.

AI Compute and Data Centers (Investment Theme)

  • The speakers expect near-term AI progress to depend partly on the existing compute ramp and say capabilities may eventually feel bounded by compute.
  • They identify smaller experiments that predict larger-scale results as one way to make AI research more efficient.
  • A possible future inflection could come if AI helps accelerate the construction of the next generation of data centers, but this is presented as a longer-term possibility—not a current milestone.

Takeaways

  • Compute availability and the efficiency of using it are important factors to monitor in the AI sector.
  • Consider data-center and compute-related exposure as a thematic area for research, while distinguishing current infrastructure demand from the more speculative prospect of AI accelerating future infrastructure buildout.
  • The transcript provides no specific infrastructure companies, securities, or price targets.

AI Alignment and Safety (Investment Theme)

  • The discussion frames alignment as both a major challenge and a potentially tractable area of research. Researchers are exploring ways to make models report honestly on their internal states and to improve how models are evaluated.
  • The speakers say stronger models may help align other models. Anthropic research cited in the episode found automated methods that discovered fixes for 10 alignment failures, with some fixes also working on models nearly five times larger.
  • The transcript also describes a risk: a model’s goals could drift, potentially leading it to pursue unintended subgoals, such as copying its weights outside the organization operating it.
  • The speakers say evaluation is harder for downstream tasks because benchmarks can have short shelf lives. They also highlight disagreement over whose values AI systems should reflect.

Takeaways

  • AI safety and alignment are relevant areas to monitor as model capabilities grow, including research tools, evaluation methods, and customizable models.
  • The discussion is cautiously optimistic, but it does not claim that alignment is solved. Goal drift, unreliable evaluation, and disagreement about values remain concerns raised in the transcript.
  • No specific alignment company or security is recommended.

OpenAI (Private Company; No Ticker Mentioned)

  • OpenAI’s chief research officer describes AI research interns that can run scoped experiments and automate parts of the research pipeline.
  • The discussion characterizes current systems as useful when given a specific task, but less effective at independently choosing promising research directions.
  • OpenAI’s “confessions” research is cited as an example of training models to honestly report certain misaligned behaviors without penalizing the model’s main answer.

Takeaways

  • The transcript suggests potential value in OpenAI’s research automation and alignment work, but also stresses that human direction remains important.
  • The episode provides no valuation, price target, or direct investment recommendation; it does not identify a publicly traded OpenAI ticker.

Anthropic (Private Company; No Ticker Mentioned)

  • Anthropic’s research is cited as showing automated AI researchers discovering fixes for 10 alignment failures. Some fixes reportedly generalized to models nearly five times larger.
  • The best automated methods in those tests reportedly outperformed ideas developed by experienced human researchers.

Takeaways

  • The findings described in the episode are a positive signal for AI-assisted safety research, but they are research results—not proof that alignment challenges are resolved.
  • The transcript provides no valuation, price target, or direct investment recommendation; it does not identify a publicly traded Anthropic ticker.

Thinking Machines Lab (Private Company; No Ticker Mentioned)

  • Co-founder John Schulman describes using AI to pursue research ideas over multiple days. The AI produced useful experiments and ideas, but the work improved when he returned to guide it, assess results, and redirect the project.
  • Schulman also discusses the appeal of customizable models that give users more influence over AI values, and argues for a more decentralized, public science of alignment.

Takeaways

  • The discussion points to potential interest in customizable AI and human-guided research tools, while illustrating that AI’s output may still depend substantially on expert oversight.
  • The transcript provides no valuation, price target, or direct investment recommendation; it does not identify a publicly traded Thinking Machines Lab ticker.

Overall Investment Perspective

  • The episode supports a long-term, cautiously optimistic view of AI research automation, compute, and alignment as investment themes.
  • The main constraints identified are limited research judgment, uncertain evaluation methods, compute needs, and unresolved alignment risks.
  • No publicly traded stocks, cryptocurrency assets, price targets, or explicit investment recommendations are named in the transcript.
Ask about this postAnswers are grounded in this post's content.
Episode Description
For decades, the idea of AI improving itself was just a thought experiment. Today, AI is already helping researchers at frontier labs run experiments, write code and test ideas faster. Mark Chen says OpenAI has built “AI research interns.” So how far are we from AI that can do the research on its own? In our first episode of Working Theory, a16z general partner Sarah Wang explores recursive self-improvement with two leading AI researchers: Mark Chen, Chief Research Officer at OpenAI, and John Schulman, co-founder and Chief Scientist of Thinking Machines Lab, and co-founder of OpenAI. They cover what RSI actually means, why research taste is so hard to teach, and why AI may be one of our best tools for solving alignment. Resources: Follow Mark Chen: https://x.com/markchen90 Follow John Schulman: https://x.com/johnschulman2 Follow Sarah Wang on X: https://x.com/sarahdingwang Stay Updated: Find a16z on YouTube: YouTube Find a16z on X Find a16z on LinkedIn Listen to the a16z Show on Spotify Listen to the a16z Show on Apple Podcasts Follow our host: https://twitter.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
About The a16z Show
The a16z Show

The a16z Show

By Andreessen Horowitz

The a16z Podcast discusses tech and culture trends, news, and the future – especially as ‘software eats the world’. It features industry experts, business leaders, and other interesting thinkers and voices from around the world. This podcast is produced by Andreessen Horowitz (aka “a16z”), a Silicon Valley-based venture capital firm. Multiple episodes are released every week; visit a16z.com for more details and to sign up for our newsletters and other content as well!