Summary
What does recursive self-improvement (RSI) in AI actually mean, and why are people talking about it now? David Shapiro takes on the topic, prompted by OpenAI’s release of hundreds of solutions to unsolved math problems and a new batch of Erdos problem results from Google’s AlphaSolve Nexus. He also touches on rumors that Anthropic may be nearing RSI, while cautioning that social media rumors deserve skepticism.
Shapiro argues that popular images of RSI drawn from Skynet, Ultron, and Cyberpunk 2077 misunderstand how AI works, since superintelligence needs large data centers and cannot simply spread across the internet or run on a phone. Instead, he defines RSI more practically: the share of the next AI generation’s code, data, and algorithms that is contributed by the current generation. He notes that OpenAI and Anthropic have increasingly acknowledged their models helping build successors, and that this is becoming the default in 2026. He also discusses efforts to automate alignment research.
The second half looks at implications, including how AI mastery of mathematics could unlock progress in materials science, medicine, chips, batteries, and fusion, and the role of formal verification in proving that code, hardware, and materials behave as intended.
šŗ Source: David Shapiro Ā· Published October 13, 2026
š·ļø Format: Opinion Editorial
