AI Could Match the Best Human Researchers by 2028

Dwarkesh Patel says he would have put superintelligence at least most of a decade away earlier this year. After watching progress in AI research and in fields where answers can be checked, he has revised that estimate sharply: AI systems could match the best human AI researchers in a couple of years, perhaps even a couple of months. If research becomes automated, he argues, progress could then compress enough that AIs match the best human experts across fields by 2030—or, in a faster scenario, as early as next year.

His median estimate for AI systems matching the best human researchers at AI research is 2028. He puts a meaningful chance on 2027, but also estimates at least a 25% chance that the milestone does not arrive until after 2032. These dates assume the labs press ahead rather than deliberately slowing down or being slowed by government action. Patel says the forecast is difficult to make in advance; his point is that the prospect of very fast progress should now be taken seriously.

Patel points to signs that AI is already speeding up work at the frontier. He cites an assessment by METR, an independent evaluation organization that spent three months embedded at Anthropic. METR estimated that Anthropic’s rate of progress on well-scoped tasks was about twice what it had been a year earlier, and that a majority of the acceleration was likely attributable to the company’s extensive use of AI in research and development. Patel also points to OpenAI’s report that its most intensive users among researchers went from using roughly $1,000 a day in internal coding agents at the beginning of the summer to more than $7,000 a day by the end of August. Anthropic, he says, reported that Claude led 26% of its AI research work, up from less than 1% in February.

The larger question, Patel argues, is whether abilities models have demonstrated in mathematics and other checkable tasks can transfer to AI research.

Patel’s most important update came from the pace of AI progress in mathematics. He traces a rapid change: models that could not reliably add four-digit numbers in 2023 were weak on high-school competition mathematics in 2024, reached gold-medal standard at the International Mathematical Olympiad in 2025, and began solving niche open problems. He points to OpenAI’s claim that agents resolved a Navier–Stokes Millennium Prize problem in 88 hours, with formal verification taking another 17.

Patel notes that the gains have been concentrated in work where results can be checked, such as finding counterexamples and proving well-scoped claims. Toby Ord has argued that there is little evidence current systems can ask powerful questions or open profound new fields of inquiry. Patel’s response is that a system can make striking advances without those abilities: current AIs, he says, have solved a Millennium Prize problem without being able to choose profound new directions.

He sees the distinction between mathematical discovery and AI research as important. Mathematics is not chiefly about resolving long lists of conjectures; its great advances include new ways of thinking, such as topology or group theory. Machine-learning research, by contrast, is judged heavily by whether a method improves training efficiency or produces more capable models. Patel argues that much of deep learning’s progress over the past decade and a half did not require major new theory. Systematically trying combinations of ideas and testing which ones work may therefore be more valuable in machine learning than it is in mathematics.

His example is the Transformer. Before its 2017 publication, researchers knew about attention and other ingredients that became part of the architecture. Chris Olah described the Transformer as a forceful combination of ideas that had already proved effective independently. The difficulty was finding the right combination and making it work through many experiments. Patel argues that this kind of persistent search—connecting ideas from nearby fields, testing variations, and keeping what performs—resembles the work at which AI has become strong in mathematics. A future model asked to make attention work better, he suggests, might have discovered an architecture like the Transformer.

Patel acknowledges that top researchers make judgments that are not directly verifiable, including choosing research directions years before their value is clear. His case is that models can generalize beyond the tasks they were trained on: a model trained on tractable mathematics problems may solve a problem no human had cracked. He expects practice on thousands of checkable research tasks, combined with more capable successive models, to help systems transfer to less easily checked parts of AI research.

If AI systems reach the level of the best human researchers, Patel expects three sources of acceleration: speed, capability, and numbers. Today’s AI systems can already perform some work at around ten times human speed, he says. They can also operate around the clock. On that basis, an automated researcher might do roughly 50 times as much work in a calendar week as a human researcher. Patel cautions that this would not translate into a 50-fold increase in progress: experiments take time and compute, and returns diminish. But he expects the added working time to matter.

The second change would be the quality of the researchers. Patel estimates that about 80% of a frontier lab’s algorithmic progress comes from its 50 top researchers. An automated company in which every researcher was as capable as the best people in today’s labs would, he argues, add another major source of acceleration.

The third is parallel labor. Patel contrasts research with writing a novel, where a swarm of agents may not produce a good result much faster than one agent. Research, he says, has more modular tasks that can be attempted independently and checked locally. He points to a project in which agents worked on thousands of subtasks involved in formalizing Fermat’s Last Theorem. AI research already divides into separate efforts: improving pretraining data, building reinforcement-learning environments, and writing more efficient kernels, among others. A large group of persistent agents could test many approaches across such work, though Patel expects diminishing returns rather than a speedup proportional to the number of agents.

What happens after that initial acceleration is, for Patel, the harder question. Automated researchers could keep getting faster, smarter, and more numerous, while useful ideas become harder to find as the easy gains are exhausted. If AI capabilities compound faster than research becomes difficult, he calls the result a “software-only singularity”: a swarm improving AI with existing compute until models approach the physical limits of current chips. If ideas become harder to find sooner, the result would instead be a “spark”: several years of progress compressed into one calendar year, followed by a slowdown until more compute allows further experiments.

Patel puts roughly even odds on those two paths. Even the slower “spark” case, he argues, would be an extraordinary change: something like the jump from GPT-3.5 to GPT-6 Astra in a year. In his comparison, GPT-3.5 could write a passable limerick or complete a trivial Python function, while Astra can build video games from scratch and produce mathematical results that have become familiar enough to seem ordinary. A further jump of similar scale, he suggests, could yield systems far beyond humans in domains with even modestly checkable outcomes.

That would not mean instant mastery of every real-world job. Patel expects systems to need experience in unfamiliar settings. But he thinks they could accumulate it quickly: copies could read an organization’s historical documents, interview employees and clients in parallel, and then learn from work on the job. He likens the required ramp-up to that of unusually capable young people who have entered unfamiliar industries and achieved exceptional results within a few years, rather than needing decades of experience.

The difficult transfer problem is how a system trained on tasks that are easy to evaluate inside a data center could become effective at building a business, winning a court case, trading, or helping a candidate win an election. Patel rejects the idea that every answer to such work can simply be encoded in model weights. Instead, he argues that reinforcement learning on increasingly broad tasks could build general abilities: finding relevant information, planning, persisting through mistakes, retaining useful lessons, and collaborating.

He imagines future training environments that are longer, more open-ended, and more varied than today’s. Agents might compete or cooperate in simulated economies, learn organizational practices through long-running competition, or be trained on historical records to predict what happened next. One proposed task: give a model information only up to 1700 and see whether it can derive general relativity. Another: give it newspaper coverage up to 1962 and ask it to predict the Cuban Missile Crisis.

Patel points to two reasons to expect transfer. First, he says current systems already show some ability to orient themselves in unfamiliar codebases and handle projects they have not seen before. In his own experience, they have also improved at tasks beyond code and mathematics, such as helping with the abstract research questions involved in writing his essay. Second, he argues that human evolution shows how a general-purpose learner can emerge from optimization on a narrow set of tasks. Evolution did not train humans to write biographies, build advanced semiconductors, or govern large nations; it selected for survival and reproduction. Patel sees AI training as a more deliberate optimization process, aimed at tasks more closely related to modern capabilities and increasingly designed to improve transfer.

He still expects real-world deployment to matter. But he argues that systems may gain experience faster than people expect. An AI shadowing a company’s CEO could read years of email, messages, and documents in a day, while copies interviewed employees and clients in parallel. It would not immediately receive feedback that depends on months of outcomes, such as the next quarterly earnings. Patel’s claim is that even with those delays, post-“spark” systems could reach top human performance in many domains within months of deployment.

Patel’s confidence depends on research being verifiable enough to train against. He divides AI research into pretraining, post-training, and inference efficiency, and argues each has measurable proxies. Pretraining can be evaluated through loss on held-out data and experiments that test whether small-scale improvements transfer to larger models. Post-training can be assessed through performance on unfamiliar environments, the amount of data or compute needed to learn, and benchmarks of research capability. Inference efficiency has directly measurable quantities such as GPU utilization, cache size, compute, and latency.

These measures matter in two ways. They can provide training signals for models themselves, and they let groups of automated researchers check whether an experiment improved a result before combining their work. Patel argues that sample efficiency and online learning—how quickly a system learns from new data—are also measurable, even if they remain difficult research problems.

In pretraining, he describes tasks where a model could search for ways to reach a target loss with less compute, then test improvements on larger models. He notes that scale matters: an improvement at small scale may not survive at large scale. But he argues that judgment about which improvements will transfer can itself be trained by testing whether small-model gains persist at larger sizes. More efficient pretraining, in turn, would make it cheaper to run experiments and reinforcement-learning rollouts, and would allow more automated researchers to run at once.

For post-training, Patel emphasizes not only improving the reinforcement-learning algorithm, but improving the tasks and environments used to train models. He cites a DeepSeek paper that attributed gains largely to the scale, diversity, and verifiability of synthesized data and environments, rather than novel changes to the optimization procedure. He also points to benchmarks that test how well models learn in unfamiliar environments and to Anthropic’s internal benchmark of problems its researchers and engineers had solved. Such tasks, he argues, could let models practise replicating recent human progress.

The strongest objections, in Patel’s view, concern continual learning, choosing the right proxies, data, and experiment compute. He had previously considered the lack of continual learning a major obstacle: models forget between sessions and do not steadily accumulate weeks or months of experience. He now thinks longer effective context windows may be enough to let a single session learn on the job, and that solving online learning may itself be possible through verifiable research practice. But he retains a serious chance that models will not compensate for the missing capability in the next few years; this uncertainty is part of his estimate that top-level automation could take until after 2032.

The proxy objection is that a system cannot climb a hill researchers have not identified. Patel responds that the Transformer was developed to improve existing translation benchmarks, and that reasoning models also advanced on established coding and mathematics benchmarks. He further proposes broader measures—such as solving Millennium Prize problems, mastering unfamiliar games, or deriving later science from earlier data—that do not depend on a single current research paradigm. He would be surprised if choosing proxies, rather than other constraints, proved the decisive obstacle.

Data is a more substantial concern. Patel and Princeton student Jerry Han found, in research he cites, that data improvements contributed more than three times as much as model improvements to pretraining progress over six years. But he argues that improving data is itself automatable research: filtering existing material, synthesizing new examples, and testing their effect on held-out performance. For reinforcement learning, he points to work in which models help construct tasks and assess them for difficulty and correctness.

Experiment compute remains a physical constraint: AI research requires running tests, not just reasoning on paper, and each additional researcher could receive a smaller share of available compute. Patel’s counterargument is that frontier labs have been growing compute rapidly, and that automated researchers could use it more efficiently than human teams do. They could run more small experiments, inspect rollout transcripts, repair flawed environments, and spend more time designing and analyzing pooled experiments. He also points to scaling laws that allowed researchers to predict GPT-4’s final loss using models trained with far less compute, and to methods for transferring hyperparameters from smaller models to larger ones. Discoveries of that kind, he suggests, could make smaller experiments more informative.