The avalanche that hit math is coming for AI development

The avalanche that hit mathematics is now on its way to artificial intelligence development. This transformation has caught many off guard. Earlier this year, I believed it would take at least a decade before AI reached superintelligence. However, in the past three months, my perspective has shifted dramatically. Through discussions with others in the field and observing AI progress in verifiable domains, I am now convinced we are only a few years or even months away from AI systems that match the best human experts in AI research.

I also believe that automating AI research will accelerate progress so rapidly that, by 2030 or possibly sooner, AI will have achieved expertise in every field, surpassing even the best human specialists. The evidence supporting this claim is compelling. Independent evaluations, such as those conducted by METR, have shown that AI progress is now twice as fast as it was a year ago, with a significant portion of this acceleration attributed to the extensive use of AI in research and development.

Major players in the AI industry, like OpenAI and Anthropic, have reported similar findings. For instance, OpenAI observed an increase in internal coding agent usage from $1,000 a day to over $7,000 a day over just a few months, while Anthropic reported that its AI model, Claude, led 26% of its AI R&D work by August, up from less than 1% in February.

In the realm of mathematics, the impact of AI is particularly striking. Current AI systems are solving complex mathematical problems with remarkable speed and efficiency, a task that previously took decades for top mathematicians and computer scientists to resolve. This rapid progress has transformed mathematics from a field dominated by theoretical conjectures to one where AI can now solve high-profile problems, such as a Millennium Prize problem, in just 88 hours.

While current AI systems may not yet be capable of posing powerful questions or discovering new mathematical fields, their ability to automate proof generation and solve well-scoped problems has profound implications for machine learning research. Unlike mathematics, where the pursuit of new theories is crucial, AI labs are primarily focused on developing algorithms that enhance training efficiency and produce smarter models.

The progress made in deep learning over the past decade and a half has not required extensive new theory, making AI development remarkably similar to the type of mathematical advancements that current AI systems excel at. This shift in AI capabilities suggests that the next few years will see AI making significant strides in automating and advancing various fields, including mathematics and AI research itself.