The avalanche that hit math is coming for AI development

1. Key Themes

AI is already accelerating AI research, with hard numbers

The foundation of Dwarkesh's updated view is that recursive self-improvement is not hypothetical; it is measurable at the frontier labs right now. Three independent data points: METR's embed at Anthropic, OpenAI's usage data, and Anthropic's own uplift numbers.

Dwarkesh: "METR, which is an independent evaluation organization, spent three months embedded inside of Anthropic, and afterwards concluded that Anthropic's rate of AI progress on well-scoped tasks is 2x what it was a year ago, and that a, quote, 'majority of this acceleration today is likely attributable to Anthropic's extensive use of AI to do R&D.'" 00:00:58

Dwarkesh: "the heaviest 10% of AI users amongst OpenAI's researchers went from using about $1,000 a day of internal coding agents at the beginning of the summer to over $7,000 a day by the end of August." 00:00:58 And: "Claude now leads 26% of Anthropic's AI R&D work, up from under 1% as recently as February." 00:00:58

Mathematics is the leading indicator for what happens to any verifiable domain

The "avalanche" in the title: in any domain where answers can be checked and models can practice thousands of rounds, AI goes from worse than most humans to better than the best humans in just a few years. Math is the proof of concept.

Dwarkesh: "As recently as 2023, these models couldn't reliably add four-digit numbers. In 2024, they could barely solve any problems in a high school math competition. In 2025, they could win a gold medal in the International Math Olympiad. To earlier this year, where they finally started to solve some niche open problems. To a month ago, where OpenAI solved a Millennium Prize problem in 88 hours." 00:02:04

He generalizes: "in every domain where we can check answers and give AIs thousands of rounds of practice, for example, competitive programming contests or well-scoped open math problems, they've gone from worse than most humans to better than the best humans within just a few years." 00:02:04

AI research is structurally better suited to this than math itself

The priorities are reversed between the fields. Mathematicians prize theory (topology, group theory); AI labs only care whether a recipe empirically lowers loss. That makes "grinding through random conjectures", the thing AIs are best at, directly valuable.

Dwarkesh: "AI labs mainly care about how any new algorithmic recipe empirically makes training more efficient or elicits smarter models. They don't care about theory for its own sake, and all the progress that we've made in deep learning over the last decade and a half has not required much new theory anyways." 00:02:04

The Transformer as a "combinatorial search" discovery

Dwarkesh argues that even the field's biggest breakthrough was essentially a persistent search over known ingredients, exactly what current AIs excel at. Chris Olah's view and Sholto Douglas's unpublished essay anchor the claim.

Dwarkesh, quoting Chris Olah: "At that time, my sense is that the transformer architecture was a natural combination of many ideas, like attention and residual networks, which had independently been effective, taken to quite aggressive lengths." 00:02:04 His conclusion: "even the Transformer... is the kind of thing that Claude 7 could have invented if Chris Olah asked it to try really hard to get this attention thing to work better." 00:02:04

Timeline: median 2028 for top-human-level AI researchers, with fat tails on both sides

Dwarkesh gives explicit probabilities: median 2028, plausible 2027, at least 25% chance of slipping past 2032. He treats METR's 2x uplift as implying 2027 under naive modeling, but discounts it.

Dwarkesh: "If I had to pick my median year for when automated researchers will be as good as the very best human AI researchers at doing AI research, I'd guess 2028... I'd guess that there's at least a 25% chance that it doesn't happen until after 2032." 00:02:04 He quotes Ajeya Cotra: "If we take at face value that AI is now nearly doubling the rate of AI capabilities progress, and assume that acceleration from AI is smooth, this naively implies that we would reach extreme superintelligence within a year." 00:02:04

Faster, smarter, more numerous: the three multipliers

Once AIs match top researchers, progress accelerates through speed (about 50x more work per calendar week), quality (80% of algorithmic progress comes from the top 50 researchers, and every agent would be at that level), and quantity (parallelizable, modular work).

Dwarkesh: "each individual automated researcher will be cramming roughly 50 times more work every calendar week than a human researcher can do." 00:11:13 And: "My sense is that around 80% of the algorithmic progress at a given frontier lab is driven by its 50 top researchers." 00:11:13 And: "AI research is closer to formalizing a theorem than it is to writing a novel, because there's a lot of independent modular work to do, and there are verifiable proxies for progress." 00:11:13

Spark vs. software-only singularity: a coin toss

The key uncertainty is whether AI getting smarter outpaces ideas getting harder to find. Dwarkesh frames two scenarios and assigns 50/50.

Dwarkesh: "AI companies have to quadruple their R&D compute and double their workforce every single year just to maintain the same rate of progress." 00:14:27 The spark: "We'd get something like 3 to 6 years of AI progress at the current rate in a single calendar year, and then things would simmer down until we build more compute." 00:14:27 "If pressed, I'd give a 50% chance to the spark scenario, and a 50% chance to the software-only singularity." 00:14:27

Superintelligence can get good at non-verifiable domains via generalization and simulated worlds

The central objection (real-world domains like business, trading, and law can't be verified in a data center) is answered by transfer, vastly scaled RL environments, and historical backtesting.

Dwarkesh: "an AI might be dropped into a simulated economy of millions of other AIs and told to co-found a successful company." 00:17:23 And: "Another task might be, given every single newspaper article until 1962, can AI predict how the Cuban Missile Crisis is going to unfold?" 00:17:23 On transfer: "Evolution did not specifically optimize us to be able to write a Pulitzer Prize-winning biography... But in optimizing us for survival on the savanna, evolution built a general-purpose learner, and in training these AIs, we're running a similar process." 00:17:23

AI research is verifiable across all three buckets, including "research taste"

In the appendix, Dwarkesh decomposes AI R&D into pre-training, post-training, and inference efficiency, each with measurable proxies. Even taste can be trained against.

Dwarkesh: "you can make an RL environment where GPT-7 has to predict whether an algorithmic improvement that it made at the 100-million parameter scale will also hold at the 10-billion parameter scale." 00:26:00 For inference: "you can easily measure GPU utilization, or KV cache size per token, or total FLOPs in a forward pass, or latency." 00:26:00 On sample efficiency: "AI's equivalent of a Millennium Prize problem. And mathematics has already shown us what happens to such problems, even when they are very difficult, as long as they are verifiable." 00:26:00

Objections (continual learning, proxies, data, compute) are real but not crippling

Dwarkesh walks through four objections and why each is likely survivable. Notably, he changed his own mind on continual learning.

Dwarkesh on compute: "by the end of 2027, OpenAI could almost give 10,000 agents... each enough compute to run a GPT-3-sized training run every single day." 00:33:32 On data: "more than three times the algorithmic progress has come from data improvements than model improvements" in pre-training (research with Jerry Xuan). 00:33:32

2. Contrarian Perspectives

Continual learning may not need to be "solved" by a continual learner

Most people (including Dwarkesh a few months ago) view the lack of on-the-job learning as the key blocker to automating AI research. His reversal: it's itself a verifiable ML research problem, solvable by amnesiac models, and in-context learning with months-long effective sessions may be enough.

Dwarkesh: "you might not need a continual learner to solve continual learning. As I explained earlier, online learning is an ML research problem with clean verifiable proxies." 00:33:32 And: "These models are out here solving Millennium Prize problems while still being these derpy savant interns that have no long-term memory. It's like watching somebody do cartwheels down the track and still keep pace with Usain Bolt." 00:33:32

AI solving a Millennium Prize problem without "asking powerful questions" is still a revolution

The standard skeptical framing (Toby Ord) says AI is only grinding verifiable subproblems, not creating new theory. Dwarkesh flips the framing and argues the limitation is irrelevant for AI research, where grinding is the whole game.

Dwarkesh: "another way to say that same exact sentence would be this: current AI systems have solved a Millennium Prize problem, all without the ability to ask powerful questions or find profound new fields of mathematical inquiry." 00:02:04

Superhuman performance in real-world domains will arrive within months of deployment, not decades

Conventional wisdom says expertise requires years of real-world experience. Dwarkesh argues humans are not at the learning-speed ceiling, and AIs can compress years of context into days.

Dwarkesh: "By 25, just two years after he got into the oil business, Rockefeller was running one of the world's largest refineries... Rockefeller and Napoleon needed experience, yes, but only a couple years' worth, not decades." 00:17:23 And: "AIs can accumulate years of deployed context in just a few days of calendar time... on its first morning, it could read every email, and Slack message, and document that the company's ever produced. By the end of its first day, copies of it have interviewed every employee and major client in parallel." 00:17:23

Compute constraints won't bottleneck RSI because labor can be traded for compute

The common objection is that experiments need compute, so more researchers won't help. Dwarkesh argues today's compute usage is wasteful precisely because researcher time is the scarce resource, and that flips under automation.

Dwarkesh: "The nature of research and experimentation at the labs today reflects that researcher time is extremely scarce relative to compute... So the way labs use compute today tells us very little about how efficiently they could use it if compute, rather than researcher time, was the key bottleneck." 00:33:32 Also: "the final model trained in the original Transformer paper in 2017 took 3 days on 8 P100s, which would cost about 50 bucks if you went to rent it today." 00:33:32

Smaller, faster-trained models tighten the RSI loop and make giant training runs obsolete

Counter to the "bigger runs" paradigm, Dwarkesh predicts that algorithmic gains will come so fast that long frontier pre-training runs would be overtaken by newer small-model runs started a week ago.

Dwarkesh: "by the time that a giant run finished, it would already have been made obsolete by a much smaller model that first began training only a week ago. It's similar to how if you launch a spaceship to Alpha Centauri with today's technology, it would probably be overtaken in transit by one that was launched a century from now." 00:26:00

3. Companies Identified

METR

Independent AI evaluation organization. Mentioned for producing the key external evidence that AI is already accelerating AI R&D. Dwarkesh: "METR... spent three months embedded inside of Anthropic, and afterwards concluded that Anthropic's rate of AI progress on well-scoped tasks is 2x what it was a year ago." 00:00:58

Anthropic

Frontier AI lab. Mentioned for using AI extensively to accelerate its own R&D, and for its internal Co-bench benchmark. Dwarkesh: "Claude now leads 26% of Anthropic's AI R&D work, up from under 1% as recently as February." 00:00:58 On benchmarks: "Anthropic, for example, has this internal benchmark called Co-bench, which contains around 500 real problems that their researchers and engineers have solved this year." 00:26:00

OpenAI

Frontier AI lab. Mentioned for internal coding-agent adoption and for the model that solved a Millennium Prize problem. Dwarkesh: "OpenAI solved a Millennium Prize problem in 88 hours." 00:02:04 And on spend: "went from using about $1,000 a day of internal coding agents... to over $7,000 a day by the end of August." 00:00:58

Claude Code / Codex

Agentic coding products from Anthropic and OpenAI. Mentioned as evidence of generalization to unseen contexts. Dwarkesh: "If you point Claude Code or Codex at a project that it's never seen before, it knows how to orient itself and figure out how everything fits together, and then tackle an ambitious assignment." 00:17:23

Magic (AI lab)

AI lab that announced a pre-training efficiency result. Mentioned as evidence that pre-training has orders of magnitude of headroom. Dwarkesh: "Magic recently announced that they pre-trained a model that matches the leading open-source base models today, like DeepSeek-V4-Pro, for just half of the compute that was used to train the original GPT-3, which would roughly cost $500,000 today." 00:26:00

DeepSeek

Chinese AI lab. Mentioned for its paper arguing that data/environment synthesis, not RL algorithm changes, drive post-training gains. Dwarkesh quoting the paper: "Systematic improvements in the scale, diversity, and verifiability of synthesized data and environments account for essentially all of the observed gains." 00:33:32 And: "We invest in large-scale automated pipelines for data synthesis and environment construction." 00:33:32

ByteDance (EdgeBench)

Released EdgeBench, a benchmark for measuring how fast models learn in unfamiliar environments. Dwarkesh: "check out this great benchmark that ByteDance released called EdgeBench for these kinds of tasks." 00:26:00

Nvidia

AI accelerator maker. Used as a benchmark for what a superintelligence could outperform. Dwarkesh: "it would be better than Nvidia at coming up with the next generation of AI accelerators." 00:14:27

Citadel and Jane Street

Elite quantitative trading firms, used as the bar for AI trading performance. Dwarkesh: "It would make more money than the current human versions of Citadel and Jane Street if it was given equal access to the market." 00:14:27

Meta (implicit via Zuckerberg)

Mentioned for paying hundreds of millions to acquire top AI talent. Dwarkesh: "It's not for nothing that Zuckerberg was willing to pay hundreds of millions of dollars to acquire the best talent in the industry." 00:11:13

3Blue1Brown

Educational YouTube channel. Recommended for explaining the compression-intelligence link. Dwarkesh: "I highly recommend this excellent video by 3Blue1Brown on the topic." 00:26:00

4. People Identified

Sholto Douglas

AI researcher and Dwarkesh's friend. Has an unfinished essay on how the Transformer could have been invented two years earlier. Dwarkesh: "My friend Sholto Douglas, who is an AI researcher himself, has this perennially half-finished blog post where he explains how you could have invented the Transformer two years earlier than people actually did." 00:02:04

Chris Olah

Interpretability researcher (Anthropic). Cited for his account of the Transformer as a natural combination of existing ideas. Dwarkesh quoting: "the transformer architecture was a natural combination of many ideas, like attention and residual networks, which had independently been effective, taken to quite aggressive lengths." 00:02:04

Toby Ord

Philosopher. Wrote an essay arguing AI can't yet ask powerful questions in math. Dwarkesh: "Toby Ord recently wrote a great essay about this." 00:02:04

Ajeya Cotra

Former guest and AI forecaster. Her blog post on METR's report implied extreme superintelligence within a year under smooth acceleration. Dwarkesh: "As former guest Ajeya Cotra put it in a blog post about this report..." 00:02:04

Dario Amodei

Anthropic CEO. Dwarkesh asked him two years ago why LLMs couldn't connect ideas across fields to make discoveries. Dwarkesh: "It was only two years ago that I asked Dario why LLMs aren't able to use their vast knowledge across so many different fields in order to connect together different ideas and make new discoveries." 00:02:04

Mark Zuckerberg

Meta CEO. Cited as proof of how much top AI talent is worth. Dwarkesh: "Zuckerberg was willing to pay hundreds of millions of dollars to acquire the best talent in the industry." 00:11:13

Napoleon and John D. Rockefeller

Historical examples of young people rapidly mastering new domains. Dwarkesh: "By the age of 27, Napoleon had stacked victory after victory against far larger armies from the other great powers of Europe." 00:17:23

Ilya Sutskever

Top AI researcher, used as the benchmark for "top human researcher." Dwarkesh: "we won't automate Ilya-level top human researchers until after 2032." 00:33:32 Also named alongside Noam Shazeer as researchers who would still make progress with limited compute. 00:33:32

Noam Shazeer

Transformer co-author and top AI researcher. Dwarkesh: "If the top human AI researchers, like Noam Shazeer and Ilya Sutskever, had enough experiment compute to train a GPT-3-sized model every single day, they might make slower progress... but it would be extremely surprising to me if they failed to make meaningful progress." 00:33:32

Jerry Xuan

Princeton student and Dwarkesh's research collaborator on data vs. model algorithmic progress. Dwarkesh: "Together with Jerry Xuan, who is a student at Princeton, I actually did some research into this." 00:33:32

Max Farrens, Gabe Freed, Thomas Larson

Friends who paired with Dwarkesh on writing the essay. 00:44:00

Usain Bolt

Referenced only as an analogy for continual learning models still keeping pace. 00:33:32

5. Operating Insights

Trade labor for compute: waste is optimal when the scarce resource is human time

Dwarkesh's point about labs generalizes: your resource-allocation patterns encode which input is currently scarce. When a constraint flips (e.g., labor becomes cheap through AI), the efficient frontier changes dramatically. Teams should audit processes that are "wasteful" only because human attention was the bottleneck.

Dwarkesh: "To save valuable researcher time, experiments today are, and should be, very wasteful with compute. So instead of painstakingly setting up and carefully monitoring tons of small-scale pre-training experiments, a human researcher today might instead kick off a couple of huge multi-agent RL runs in parallel. But if you're bottlenecked by experiment compute, you can just do more of the former." 00:33:32

Build verifiable proxies and leaderboards to let parallel agents (or people) merge work safely

Verifiability has a second benefit beyond RL training: it lets independent workers empirically check whether they beat a baseline before merging, which makes parallel work coordinate without heavy management.

Dwarkesh: "these proxies allow each agent to empirically check whether they've beaten some score on a leaderboard before they merge their contributions, and so they can work much more effectively in parallel." 00:26:00

Convert your own real problems into internal benchmarks that double as training environments

Anthropic's Co-bench is a template: take real problems your team solved recently, package them as a benchmark, and use it for both measurement and practice.

Dwarkesh: "Anthropic, for example, has this internal benchmark called Co-bench, which contains around 500 real problems that their researchers and engineers have solved this year. And every benchmark is also an RL environment, and an RL environment like this would allow Anthropic to have their models practice redoing the last 3 to 6 months of human progress at the frontier." 00:26:00

Onboard AI to a role by giving it full organizational context from day one

The "shadow the CEO" model: AI ramps fast because it can ingest every email, document, and message and run parallel interviews with every stakeholder. Operators should design deployment around maximal context ingestion rather than narrow task prompts.

Dwarkesh: "By the end of its first day, copies of it have interviewed every employee and major client in parallel." 00:17:23

Test at the scale you'll actually deploy at, rather than betting that small-scale results transfer

Faster iteration cycles let you eliminate the extrapolation risk entirely.

Dwarkesh: "since you're training smaller models faster and on a scale where you can run thousands of experiments daily, you won't have to bet that small-scale results will hold up in bigger models, you've already tested at the scale which you'll be training at." 00:26:00

6. Overlooked Insights

The cost of replicating the original Transformer is now about $50

Mentioned almost as a throwaway while making a compute-bottleneck argument, but it is a striking datapoint on how steep algorithmic/hardware deflation is: the breakthrough that built the modern AI industry can be reproduced for pocket change, which implies that today's "frontier" experiments will be trivially cheap soon, and that compute per researcher is far less binding than it appears.

Dwarkesh: "the final model trained in the original Transformer paper in 2017 took 3 days on 8 P100s, which would cost about 50 bucks if you went to rent it today, or about 1/10,000th of the training compute of GPT-3." 00:33:32

Data improvements have driven 3x more pre-training algorithmic progress than model changes, and RL gains are dominated by environment synthesis

Buried in the objections section is a significant finding that reframes what "AI research" is: most progress is data and environment engineering, which is exactly the kind of work that is automatable, verifiable, and already being automated (DeepSeek's pipelines that use "difficulty and correctness as reward signals" to iteratively build better tasks). This makes the "data wall" argument much weaker and points to environment-construction as a high-value investment area.

Dwarkesh: "for pre-training, over the last six years, more than three times the algorithmic progress has come from data improvements than model improvements." 00:33:32 And from DeepSeek: "Using difficulty and correctness as reward signals, we iteratively train the model to construct better tasks." 00:33:32