Why Dwarkesh Patel moved his AI-research automation forecast forward

Earlier this year I would have guessed that superintelligence was the better part of a decade away. After discussions over the past three months and watching progress in tasks whose answers can be checked, I have substantially updated that view.

My median guess is now that AI will match the best human AI researchers around 2028, assuming labs continue at full speed. I think this could happen as soon as 2027, but put at least a 25% chance on it taking until after 2032. Soon after research is automated, I expect models at least as capable as the best human experts across fields.

AI is already accelerating AI research, according to a METR assessment of Anthropic’s work on well-scoped tasks. I also point to reports from OpenAI and Anthropic about increased use of AI coding agents and the growing share of research work led by Claude.

Mathematics has been a major update for me. In tasks where results are readily checked, models have moved quickly from struggling with basic arithmetic and competitions to solving difficult open problems. I cite the recent reported solution of a Millennium Prize problem as an example of the speed of this change.

Toby Ord notes that evidence of automated proof is not yet evidence that AI can ask powerful questions or found important new fields. I agree this distinction matters in mathematics, where new concepts can matter more than individual solutions. But AI research often values an empirically better training recipe even without new theory.

Consider the transformer. Researchers had already developed many of its ingredients, including attention and residual networks. Its success required combining several known pieces and running many small experiments. Persistent, intelligent search across combinations is the kind of work I expect AI researchers to do well.

AI research is less directly verifiable than mathematics in some respects. Nevertheless, I think enough of it can be checked to train and assess powerful automated researchers. The 2028 median is uncertain: a harder automation problem could delay it, while current acceleration could make 2027 possible.

Some researchers I consider well informed expect automated researchers to surpass the best humans in 2027. Ajeya Cotra has noted that a simple model of the recently reported acceleration could imply a very fast takeoff. I do not take that extrapolation as certain, which is why my median is later.

Once AI matches the best human researchers, I expect progress to accelerate for three reasons: automated researchers can work faster, every copy could be highly capable, and many more copies can work at once. I estimate that a copy working around the clock at ten times human speed could fit roughly 50 times as much work into a calendar week.

That does not translate directly into a 50-fold increase in research progress. Experiment compute and other resources remain limited, and adding researchers brings diminishing returns. My estimate is that a relatively small group of top researchers accounts for a large share of frontier-lab algorithmic advances; copies at that level could still change the pace substantially.

Some tasks parallelize better than others. A swarm may not write one great novel much faster, but a formal proof can have thousands of independently checkable subtasks. AI labs similarly have teams working on data filtering, training environments and efficient kernels; agents could explore alternatives in parallel.

Even 100,000 collaborators would not make a project 100,000 times faster. Still, I find it unlikely that orders of magnitude more automated labor would add no more than a modest acceleration in work with separable experiments and verifiable measures.

After the initial acceleration, two forces compete. Recursive self-improvement could make automated researchers faster and smarter, while useful research ideas may become harder to find. Which force dominates determines whether acceleration continues or slows.

I assign roughly equal probability to two outcomes. In a software-only singularity, self-improving researchers rapidly reach the limits of existing chips within months. In a “spark,” roughly three to six years of today’s progress occur in one year, then progress slows until more computing capacity is built.

Even the slower spark would be dramatic by my lights: I compare four years of progress compressed into twelve months with the jump from GPT-3.5 to GPT-6 Astra. I expect another comparable jump could produce systems far beyond human performance in domains with reasonably verifiable results.

A natural objection is that many important skills require real-world experience rather than tasks that can be graded inside a data center. I do not think all the answers can be baked into model weights. Instead, I expect training on verifiable tasks to produce general abilities that let models learn rapidly in unfamiliar settings.

Those abilities include seeking relevant information, making and revising plans, tracking lessons and collaborating. I expect automated engineers to create far more varied, longer and more complex reinforcement-learning environments, including simulated economies in which agents build companies or compete over long periods.

Other exercises could ask a model to predict historical events or rediscover scientific ideas from information available before the discovery. I also expect improvements in online learning and sample efficiency—the ability to learn more from limited data—to transfer to unfamiliar real-world tasks.

Current coding agents already orient themselves in unfamiliar projects, and in my experience newer models help more with abstract research as well. I also use evolution as an analogy: selection for survival produced humans capable of activities far beyond the original pressures, while AI labs deliberately test for broader transfer.

Even a superintelligent learner would need some interaction with the real world. My guess is that it could nevertheless reach expert performance within months of deployment. I point to unusually capable young people who accomplished major things after only a few years in a field, while acknowledging that feedback arriving over months cannot be instantly obtained.

A future model shadowing a CEO might read company records and interview employees and clients in parallel, but it would still have to wait for outcomes such as quarterly earnings. In my median spark scenario, I expect top-human performance across domains around 2028 or 2029.

Even if software improvements slow after a spark, I expect AI to help expand compute by designing chips and facilities and assisting with construction. That path could take years to reach an endpoint that a software-only singularity might reach in months.

In the appendix I explain why I think key AI research problems have measurable proxies. Sample efficiency and online learning can be evaluated through the data, tasks or time required to reach a particular performance level. I divide AI research into pretraining, post-training and inference efficiency.

Verifiability helps in two ways: labs can train models through repeated practice on measurable tasks, and parallel agents can check whether their individual contributions improve a score. For pretraining, one possible objective is reducing the compute required to reach a given training loss.

Small-scale improvements do not necessarily carry over to large models. I suggest training models to predict which changes will scale, using comparisons between experiments at different sizes. That could make some aspects of research judgment measurable, though the experiments themselves still cost compute.

Better pretraining can improve general capabilities while allowing a smaller model to match a larger one. Smaller models can train and run experiments more cheaply, potentially allowing more automated researchers and a tighter feedback loop. I speculate that rapid small runs might overtake much longer training projects.

For post-training, measurable targets include the compute and data needed to improve a capability, how quickly a model learns an unfamiliar environment, and performance on real research tasks. For inference, quantities such as latency, GPU utilization and memory usage provide direct measurements.

I take objections seriously; they reduce my probability of the most extreme takeoff. The first is continual learning: today’s models do not reliably accumulate weeks or months of on-the-job experience, which was previously my main reason for doubting near-term automation of top researchers.

I have changed my mind for two reasons. Longer effective context might let a single session retain much more experience, and a model need not already possess continual learning in order to help invent it. Still, failure on this and related capabilities is part of why I assign a substantial probability to a timeline after 2032.

A second objection is that the right measurable objective might itself require research taste, especially across new paradigms. I respond that transformers improved existing translation benchmarks, and reasoning models improved existing math and coding benchmarks. Broader tests could also measure performance on new scientific problems or unfamiliar games.

A third objection concerns data. With Jerry Han, I found that improvements to training data contributed more than three times as much algorithmic progress as model changes over the period we studied. But filtering existing data and creating synthetic data are themselves research tasks that I expect agents increasingly to automate.

Finally, experiments require real compute and time, unlike reasoning through a mathematical proof. I estimate that, if frontier compute keeps growing at its recent rate, by late 2027 a large lab could provide thousands of agents substantial daily experimental capacity. This estimate depends on that growth continuing.

Research today is organized around scarce expert time relative to compute. If the bottleneck shifts, automated researchers could trade labor for compute: run carefully designed small experiments, inspect each training rollout, fix broken environments and analyze pooled experiments in much greater detail.

They could also improve the science of learning from smaller experiments. Scaling laws and methods for transferring hyperparameter choices from small models already show how a smaller run can inform a larger one. I expect automated researchers to search for further relationships of this kind.