Article
AI develops stronger AI on its own, compressing 4 years of progress into 1 year: Dwarkesh predicts superintelligence could emerge as early as next year
Beating AI Flash News: 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 advance this quickly, but now he thinks superintelligence could appear by 2030—at the earliest, even next year.
AI is most likely to catch up with top human AI researchers around 2028, potentially achieving this within just a few months. Once AI can independently develop stronger AI, technological progress that would normally take years could be compressed into one year.
METR recently investigated Anthropic and found that its model's capability improvement speed on certain quantifiable tasks is about twice what it was previously, with much of this acceleration likely stemming from AI involvement in R&D. Anthropic's own data also shows that in February this year, Claude-led R&D accounted for less than 1%, rising to 26% by August—though human oversight remains necessary.
Much of the work in AI R&D can be experimentally verified. For example, after adjusting training algorithms or optimizing data, one can directly test whether the model has improved. AI can continuously propose solutions, conduct experiments, check results, and then continue refining. Similar leaps in capability have already occurred in mathematics and programming; AI R&D may follow suit.
Once reaching the level of top researchers, AI can also work around the clock while deploying numerous research agents. According to Dwarkesh's estimates, a single AI researcher could accomplish 50 times the weekly workload of a human. However, experiments still consume computing power, so overall R&D speed will not increase directly by a factor of 50.
If AI R&D becomes automated, two scenarios may arise:
One is that AI continuously improves itself, approaching the limits of what current chips can support within just a few months;
The other is that three to six years of technological progress are first compressed into one year, only to slow down later due to insufficient computing power and experimental resources.
Dwarkesh assigns a subjective probability of 50% to each scenario.
Currently, AI still cannot fully independently complete cutting-edge R&D. Identifying the correct research direction is far more complex than solving problems with standard answers. If this capability remains stuck for a long time, the timeline could be delayed. Dwarkesh believes there is still a 25% chance that AI won't catch up with top human AI researchers until after 2032.