The avalanche that hit math is coming for AI development
Summary
- Patel’s median forecast is that AI matches the best human AI researchers in 2028, assuming labs proceed without deliberate or government-imposed slowdowns. He assigns at least a 25% chance to this taking until after 2032, but considers 2027 plausible. His broader forecast reaches best-human expertise across fields by 2030, plausibly as early as next year—a “wild claim” he says he rejected only months earlier.
- Patel cites a doubling of Anthropic’s progress on well-scoped tasks as evidence that AI is already accelerating its own development. METR judged a majority of that acceleration likely attributable to AI use, while Anthropic reported that “Claude now leads 26%” of its AI R&D work, up from under 1% in February.
- AI research could benefit more than mathematics from machines that solve verifiable problems without inventing new theory. Empirically better training recipes are valuable regardless of whether they deepen human understanding. Patel argues that even the Transformer depended on “persistently and intelligently matching dozens of different combinations of ideas”—the capability he sees emerging in AI mathematics.
- A researcher operating at 10× human speed around the clock could perform roughly 50× as much work each calendar week. Patel estimates that about 80% of frontier-lab algorithmic progress comes from its top 50 researchers; automation could replicate that caliber of talent at scale. Faster, smarter, more numerous researchers would accelerate progress, although experiment compute and diminishing returns limit the gains.
- Patel assigns 50% to a temporary research “spark” and 50% to a software-only singularity. The distinction is whether improving researchers outpace the increasing difficulty of finding ideas. Even the spark compresses three to six years of progress into one calendar year—a jump he compares to “GPT-3.5 to GPT-6 Astra.”
- The forecast depends on AI becoming a general learner that acquires real-world expertise within months. Patel expects sufficiently advanced systems to outperform NVIDIA in accelerator design and Citadel and Jane Street with equal market access. His claim is that training teaches models to “ramp up quickly in new contexts,” while real-world experience and delayed feedback still matter.
- Continual learning and experiment compute remain substantive bottlenecks in Patel’s forecast. He has softened his objection that models cannot accumulate experience, because much longer context might suffice and AI might solve online learning itself. On compute, his response is to “trade labor for compute”: abundant researchers could design smaller, more informative experiments and scrutinize their results far more thoroughly.
Deep dive
1. Mathematics changes the case for automating research
Patel begins with acceleration already underway. METR’s three months embedded at Anthropic reportedly found that progress on well-scoped tasks had doubled over a year, with a majority of the acceleration likely attributable to AI use. Anthropic separately reported Claude leading 26% of its AI R&D work, versus under 1% in February.
His mathematical timeline runs from unreliable four-digit addition in 2023, through weak high-school competition performance in 2024, to International Mathematical Olympiad gold-medal performance in 2025. After niche open problems earlier this year, he says OpenAI solved a Millennium Prize problem in 88 hours a month before the essay. He describes current AIs as slicing through well-scoped math problems “like a hot knife through butter.”
Toby Ord’s qualification matters: there is “very little evidence” that current systems can ask powerful questions or establish profound new fields. Patel accepts the distinction but reverses its significance for machine learning. Mathematics values conceptual understanding; AI labs value recipes that empirically work. Merely “grinding through random conjectures” could therefore be enormously productive in AI research.
The Transformer supplies his concrete mechanism. Before its 2017 publication, researchers had identified attention and residual networks as useful ideas. Chris Olah called the architecture a “natural combination of many ideas.” Making it outperform existing methods required combining roughly half a dozen known tweaks through many experiments. Patel argues that persistent, intelligent recombination resembles the mathematical work AI now handles well—and suggests Claude 3.7 could have invented the Transformer if prompted to make attention work better.
2. Automated researchers could compress years into one
Earlier this year, Patel expected most of a decade; his median for best-human AI research is now 2028, assuming labs proceed without deliberate or government-imposed slowdowns. He puts at least 25% on after 2032, while taking 2027 seriously. Ajeya Cotra’s extrapolation from METR reaches “extreme superintelligence within a year” if AI nearly doubles progress and acceleration is smooth. Patel’s uncertainty about such straightforward modeling keeps his median later.
His labor model combines 10× speed, continuous operation—roughly 50× weekly work—and copies matching the best human researchers. He estimates about 80% of frontier-lab algorithmic progress comes from its top 50 people. Formalizing Fermat’s Last Theorem illustrates parallelization: modular tasks with local checks let agents work independently. AI development similarly separates into data filtering, RL environments, kernels, and other workstreams, although experiment compute and diminishing returns limit the aggregate speedup.
The subsequent trajectory depends on recursive improvement outrunning diminishing research returns. Patel notes that AI companies already have to quadruple R&D compute and double their workforce each year just to maintain the same rate of progress. He assigns 50% to a “software-only singularity,” progressing rapidly without additional compute, and 50% to a “spark”: three to six years of progress in one year before larger experiments become necessary. Even four years compressed into twelve months means another “GPT-3.5 to GPT-6 Astra” leap, starting from today’s much stronger baseline.
3. General learning could carry the gains into the economy
Patel does not claim every real-world answer can be baked into model weights. His response to business-building, courtroom, and election examples distinguishes domain knowledge from general capabilities: seeking information, planning, persisting, remembering, and collaborating. At a tiny fraction of the RL scale he expects during the intelligence explosion, systems already generalize across unfamiliar coding projects. He treats that as early evidence that verifiable training can teach models to “ramp up quickly in new contexts.”
His proposed substrate includes tens of thousands more RL environments, with simulated economies where millions of AIs experience centuries of corporate competition. Historical backtests could ask models to discover general relativity using information available before 1700. His analogy is evolution: selection for survival on the savanna produced a general-purpose learner capable of semiconductor design, despite never directly optimizing that task.
The Fortune 500 example makes deployment concrete: an AI could read every corporate email, Slack message, and document on its first morning, while copies interview every employee and major client by day’s end. Quarterly feedback would still require calendar time. Nevertheless, Patel expects top-human performance across domains within months—2028 or 2029 in his median scenario.
After the spark, he expects superiority in accelerator design over NVIDIA, and in trading over Citadel and Jane Street given equal market access. Remote access to a materials laboratory could enable superior superconductor research. Software stagnation would not end the story: global AI compute has grown roughly 3× annually for half a decade, and AI-designed chips and fabs could support further expansion.
4. Research has measurable objectives across the stack
For pre-training, held-out loss supplies a direct objective. Patel acknowledges that small-scale improvements may fail at larger scale, but argues that this “research taste” is itself trainable: GPT-7 could practice predicting whether an improvement at 100 million parameters survives at 10 billion. Verification enables both repeated RL practice and independent checking before agents merge contributions.
Better pre-training also compounds: equally capable smaller models make training, RL rollouts, and researcher copies cheaper, tightening the improvement loop. Patel cites Magic’s announcement of a base model matching leading open models such as DeepSeek v4 Pro using half the original GPT-3 training compute—roughly $500,000 today. He considers further orders-of-magnitude improvements plausible.
Post-training offers measurable targets in compute, data, and time to reach a score; adaptation to unfamiliar environments; and research performance itself. Anthropic’s CoBench contains around 500 problems its researchers and engineers solved that year. Inference efficiency supplies GPU utilization, KV-cache size, FLOPs, and latency. Sample efficiency and online learning, Patel argues, likewise admit concrete measurements.
Data creation belongs inside this loop. Patel’s research with Princeton student Jerry Huang found that data improvements contributed more than three times as much pre-training algorithmic progress as model improvements over six years. He cites DeepSeek attributing “essentially all of the observed gains” in its post-training to better synthesized data and environments, with increasingly automated construction.
5. Learning and compute could constrain the loop
Patel’s central change of mind concerns continual learning. He spent a year and a half objecting that “every new session is their first day on the job.” Now he thinks months-long effective context might suffice, or models lacking continual learning might nevertheless discover it. He still sees a serious chance that it remains unsolved and other strengths cannot compensate.
Another objection is that researchers must invent the right objectives before optimizing them. Patel answers that existing translation benchmarks captured the Transformer’s breakthrough, just as coding and mathematics benchmarks captured reasoning improvements. Broader tests—discovering scientific results or mastering unseen games—could also evaluate new paradigms. He regards this objection as weaker than the others.
Experiment compute is a more serious constraint, but Patel argues the available amounts remain substantial. If current growth continues, by the end of 2027 OpenAI could give almost 10,000 agents enough compute each for a daily GPT-3-sized training run. The final model in the original Transformer paper took three days on eight P100s—about $50 at today’s rental prices, or one ten-thousandth of GPT-3’s training compute.
His response is to “trade labor for compute”: inspect rollouts, stop broken environments, run meticulous ablations, and design informative pooled experiments. Better extrapolation offers another lever: the GPT-4 paper says its final loss was predicted using runs with 10,000× less compute; μP transfers hyperparameter choices from models 100× smaller. One researcher’s framing stays with him: “the gods above must work from the same principles as the ants below.”