I thought superintelligence was a decade away. Not anymore.
Dwarkesh Patel
AI research is speeding up 0:00
The speaker says his view has shifted fast. He now thinks superintelligence may be only a couple of years away, or even a couple of months away in the narrow case of AI matching the best humans at AI research. He points to signs that AI is already accelerating its own progress. Meter found Anthropic’s progress on well scoped tasks was twice as fast as a year earlier, with most of that boost likely coming from AI used for research and development. OpenAI said its heaviest researchers raised daily spending on internal coding agents from about $1,000 to more than $7,000 by late August. Anthropic also reported that Claude now does 26% of its AI R&D work, up from under 1% in February.
Two possible outcomes 15:00
Ideas get harder to find as the easy ones are used up. If AI research speed outruns that slowdown, he says you get a software-only singularity: AI improves itself fast and reaches superintelligence within months, limited only by current chips. If ideas get harder faster, you get a spark: three to six years of AI progress in one calendar year, then slower gains until more compute is built. He treats the two as about equally likely. Even the spark case is wild. A jump like GPT-3.5 to GPT-6 Astra would move from writing passable limericks and tiny Python functions to building games, making serious math, and becoming vastly superhuman in areas that can be checked, like chip design, trading, and lab work.
Verifiable research work 29:31
Automated research can be checked in several parts of AI work. Pre-training improvements can be tested by whether an idea that works at 100 million parameters still works at 10 billion. Post-training has clear proxies too: you can measure whether a method lowers compute, data, or time for RL, or whether a model learns faster in a new game or environment. Inference efficiency is even easier to verify through GPU use, KV cache size, forward-pass flops, and latency. The main objections are serious, but they do not look fatal. Continual learning may still be missing, yet the speaker now thinks scaling up context and longer sessions may be enough, and that models without long-term memory are already solving hard problems. Data can also scale, because AIs can help compile it. A DeepSeek paper is cited for showing that gains came mainly from better synthetic data and environments, not just RL changes.
Smaller models still matter 43:32
A smarter, faster automated company could keep looking for new scaling findings like MUP, which helps researchers guess the best hyperparameters for a larger run from a model 100 times smaller. The speaker wonders how many more discoveries like that exist, and recalls an AI researcher who said tiny neural networks can teach you a great deal. The line was, 'The gods above must work from the same principles as the ants below.'
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

