Dwarkesh Patel publishes a 7,000-word long article predicting that superintelligence could arrive as early as next year
Dwarkesh Patel, host of the AI podcast Dwarkesh Podcast, published a 7,000-word long article predicting that AI's level in doing AI R&D will catch up with top human researchers in as soon as a few months and at the latest by 2028, and that superintelligence will emerge before 2030. Citing Anthropic's September data, he said that as of August, Claude had led 26% of Anthropic's AI R&D work, while in February this year that proportion was still less than 1%; METR's early-November on-site evaluation report said that Anthropic's speed of improving model capabilities was about twice that from March 2024 to October 2025, of which about 80% of the acceleration may be attributable to extensive use of AI for R&D. He also estimated that AI's one-week workload is about 50 times that of a human researcher, and believes that in the first year of AI R&D, roughly four years' worth of progress can be covered.
Why read this
AI score 55
The original article sorts out the changes in Dwarkesh Patel's arguments from being cautious last year to now predicting superintelligence before 2030, which can be compared against measured data on AI participation in R&D.
Topics
Body · AI translation
Add new crowd-testing promotional copy Add new crowd-testing promotional copy end Add App-exclusive traffic referral
(Source: AI Information Gap)
Just now, Dwarkesh Patel, host of the top AI podcast Dwarkesh Podcast, published a long article: "Superintelligence will arrive before 2030, possibly as early as next year." His reasoning: "AI's breakthroughs in mathematics are like an avalanche, and they are now rushing toward AI R&D itself."
How far away is superintelligence, really?
A big name who was still writing last year that AGI wasn't coming soon has just changed his answer to before 2030.
Data released by Anthropic in September shows that, as of August, Claude led 26% of Anthropic's AI R&D work; researchers give a rough direction, and Claude does most of the work from start to finish.
In February this year, that proportion was still less than 1%.
The person saying this is Dwarkesh Patel, host of the top AI podcast Dwarkesh Podcast. OpenAI co-founder Ilya Sutskever and Anthropic CEO Dario Amodei have both appeared on his show.
In June last year, he wrote an article specifically titled "Why I don't think AGI is just around the corner."
His latest 7,000-word article is titled "The avalanche sweeping mathematics is now rushing toward AI R&D."
According to Dwarkesh's prediction, AI's level in doing AI R&D catching up with the top human researchers may happen in 2028, or in a few months if things go fast. Not long after that, superintelligence that is no worse than top experts will emerge in every field.
"I know this sounds crazy; just a few months ago, I didn't believe it myself."
In 2023, AI models could not even do four-digit addition reliably; in 2024, they basically could not solve high school math competition problems; in 2025, they won a gold medal at the International Mathematical Olympiad; earlier this year, they began solving problems left by the mathematician Erdős.
By September, OpenAI used internal models to drive about 10,000 agents and spent 88 hours to crack the Navier-Stokes equation problem, one of the Millennium Prize Problems.
In mathematics, proving a series of conjectures is not the ultimate goal; coming up with new concepts like topology and group theory is far more important.
The Transformer in 2017 came about exactly this way. Components such as attention mechanisms and residual connections already existed; what was missing was someone to splice six or seven small improvements together and then run large numbers of experiments to verify them.
And this is precisely what AI is best at in mathematics. With a future Claude 7, if you ask it to improve the attention mechanism, it may be able to invent breakthroughs like the Transformer itself.
The third-party evaluation organization METR was embedded inside Anthropic for 3 months and released an evaluation report in early November.
On tasks with clear goals, the speed at which Anthropic is now improving model capabilities is about twice that of March 2024 to October 2025.
About 80% of that acceleration may be attributable to Anthropic's extensive use of AI for R&D.
OpenAI also published a set of data in September.
The top 10% of researchers who use agents most consume more than $7,000 worth of tokens per person per day, calculated at API prices.
OpenAI's goal for itself is to develop automated AI researchers before March 2028, basically consistent with Dwarkesh's prediction.
According to Dwarkesh's estimate, today's AI works about 10 times faster than humans, and it doesn't need sleep or weekends. By that calculation, one AI researcher's weekly workload is roughly 50 times that of a human researcher.
About 80% of algorithmic progress at a top AI lab comes from the top 50 people, which is also why Meta's Zuckerberg is willing to spend hundreds of millions of dollars to poach talent. Replace them with AI, and every AI researcher is at that level, and thousands of them can be copied to work together.
You might think, "No matter how many researchers there are, the compute for running experiments is limited. AI researchers work 50 times faster, but R&D may not speed up 50 times."
But Dwarkesh believes that even if compute stays the same, R&D can still speed up many times over.
In a footnote in his article, he answers this question with a thought experiment. Suppose there is a "sloth AI" company with as much compute as OpenAI, with only one difference: its researchers work only 1 hour per week.
How much slower would this company progress? Not dozens of times slower, because while the researchers sleep, experiments are still running. But if researchers have only 1 hour a week, they cannot carefully design and analyze experiments, and a lot of compute will be wasted. Dwarkesh estimates that Sloth AI's progress would be at least 5 times slower than OpenAI's.
This shows that with the same amount of compute, how much time researchers spend thinking still affects R&D progress.
The reverse is also true. AI researchers think faster than humans and don't need sleep. Compared with a company full of AI researchers, today's OpenAI is that sloth AI. Since less thinking time slows progress, more thinking time will speed it up.
Human thinking speed is just an accident of evolution. "If it happened to be stuck exactly at the point where 'thinking faster wouldn't help,' that would be too much of a coincidence."
This footnote went viral on X.
Dwarkesh himself was amused too: "Good grief, then maybe I really shouldn't have put it in a footnote, haha."
AI companies have to increase R&D compute to 4 times the original and double their headcount every year just to maintain the same pace of progress.
Dwarkesh believes the two outcomes are each about 50% likely.
According to his estimate, the first year of AI R&D could cover roughly four years' worth of progress, equivalent to going from GPT-3.5 to GPT-6 Astra in 12 months.
Back then, GPT-3.5 could only write a doggerel verse and complete a simple function; it was basically a toy. Today's GPT-6 Astra can develop a complete game from scratch and continuously produce Fields Medal-level mathematical results, and everyone has become used to it.
If the next equally large leap begins in 2027 or 2028, the AI at that time will most likely be better than Nvidia at designing the next generation of chips.
Give it a materials laboratory it can operate remotely, and it might find a room-temperature superconductor faster than all of humanity.
Dwarkesh says that in his next long article he will talk about what superintelligence looks like.
Let's mark the 2028 time point for now. Statement: This is an exclusive article for Sina.com. Reproduction without authorization is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin
Non-targeted 300*250 button end