AI developing stronger AI itself, compressing four years of progress into one: Dwarkesh predicts superintelligence could emerge as early as next year.

Beating AI News Flash: Renowned AI podcast host Dwarkesh Patel has significantly moved up his prediction for superintelligence. Just a few months ago, he didn't believe AI could progress so quickly, but now he thinks superintelligence could emerge by 2030, or even as early as next year.

AI is most likely to match top human AI researchers around 2028, and could potentially do so within a few months. Once AI can independently develop stronger AI, technological progress that would normally take years could be compressed into a single year.

A recent METR survey of Anthropic found that its models' capability progress on certain quantifiable tasks is roughly twice as fast as before, with much of the acceleration likely coming from AI's involvement in R&D. Anthropic's own data also shows that in February of this year, Claude-led R&D work was less than 1%, and by August it had reached 26%, though human oversight is still required.

Many tasks in AI R&D can be verified through experiments. For example, after adjusting training algorithms or optimizing data, you can directly test whether the model has become stronger. AI can continuously propose solutions, run experiments, check results, and continue improving. Similar capability leaps have already occurred in mathematics and programming, and AI R&D may follow suit.

Once it reaches the level of top researchers, AI can also work around the clock and deploy large numbers of research agents simultaneously. According to Dwarkesh's estimate, a single AI researcher could complete 50 times the workload of a human per week. However, experiments still consume compute, so overall R&D speed won't directly increase 50-fold.

If AI R&D becomes automated, two scenarios could emerge:

One is that AI continuously improves itself and approaches the capability limit supported by existing chips within just a few months;

The other is that three to six years of technological progress is first compressed into one year, then slows down due to insufficient compute and experimental resources.

Dwarkesh assigns a 50% subjective probability to each of these two scenarios.

Currently, AI still cannot fully independently complete frontier R&D. Finding the right research direction is also far more complex than solving problems with standard answers. If this capability cannot be breakthrough for a long time, the timeline could be delayed. Dwarkesh believes there is still a 25% chance that AI won't match top human AI researchers until after 2032.