📰 Key Takeaways

Dwarkesh Patel, the well-known AI podcast host, has moved up his superintelligence forecast sharply, saying it could show up as early as next year and no later than 2030. He thinks AI is most likely to match top human AI researchers around 2028, possibly in just a few months. Once AI can build better AI on its own, progress that used to take years could be compressed into a single year.

As supporting evidence, METR recently looked into Anthropic and found that its models are improving at roughly twice the previous pace on some quantifiable tasks, with most of the speedup coming from AI taking part in R&D. Anthropic’s own numbers show that Claude-led R&D work was under 1% in February this year and reached 26% by August, though it still needs human oversight.

Dwarkesh points out that a lot of AI R&D work can be verified through experiments. For example, you tweak a training algorithm or clean up the data, then test right away whether the model got stronger. That lets AI keep proposing ideas, running experiments, checking results, and improving, and we’ve already seen similar jumps in math and coding. Once it reaches top-researcher level, AI can also work around the clock and run huge numbers of research agents at once. He estimates a single AI researcher could get through 50x the weekly workload of a human, but experiments still burn compute, so overall R&D speed won’t simply go up 50x.

After automation, things could go one of two ways. In one, AI keeps improving itself and within just a few months approaches the capability ceiling that current chips can support. In the other, three to six years of progress get compressed into one, then things slow down because compute and experiment resources run short. Dwarkesh puts a subjective 50% on each. He also cautions that AI still can’t do frontier research independently, and that finding the right research direction is much harder than solving problems with a standard answer. If that ability takes a long time to crack, the timeline slips, and he thinks there’s still a 25% chance AI won’t match top human researchers until after 2032.

💬 JudyAI Lab Take

Dwarkesh Patel has moved his superintelligence timeline up significantly, saying it could arrive as early as next year and no later than 2030. His core call is that AI will most likely match top human AI researchers around 2028, and possibly within just a few months.

What’s worth noticing here isn’t only the date, but the reasoning behind it. A lot of AI R&D work gives you results you can verify directly. You tweak a training algorithm or optimize the data, then immediately test whether the model got better, so AI can keep proposing, experimenting, checking, and improving. The original report mentions that METR found Anthropic’s progress on some quantifiable tasks is about twice as fast as before, and Anthropic’s own data shows Claude-led R&D work went from under 1% in February to 26% in August, though it still needs human oversight. Another thing to notice is that Dwarkesh doesn’t overstate his case: experiments consume compute, so R&D speed won’t jump 50x, AI still can’t do frontier research on its own, and he thinks there’s a 25% chance it won’t match humans until after 2032. For AI builders, the takeaway is that tasks that can be verified quickly get automated first, while judging direction is the real bottleneck.

My suggestion: go through your own workflows, hand the steps that are “verifiable and repeatable to test” to AI first, and keep your people for deciding what to build.

📅 Source Info

🔗 Further Reading