
Nov 13, 2026 · 44 min
Dwarkesh moves his AI researcher forecast to 2028
The avalanche that hit math is coming for AI development
The episode argues that automated AI researchers could accelerate not only software progress but also the chips, data centers, and robots needed to sustain it.
- 1Mathematical reasoning and AI-assisted coding have shortened Dwarkesh’s timeline for systems matching top human AI researchers.
- 2Automated researchers could work continuously, replicate widely, and divide verifiable tasks across large populations of agents.
- 3Continual learning, research proxies, data limits, and experiment compute remain unresolved constraints on recursive improvement.
Don't miss
Dwarkesh gives a median 2028 estimate for AI systems matching the best human AI researchers, while emphasizing meaningful uncertainty around earlier and later outcomes.
The brief
Recent advances in mathematical reasoning and AI-assisted coding have moved Dwarkesh’s median forecast for systems matching the best human AI researchers to 2028.
The argument rests on a shift in research economics: AI agents can work continuously, be copied, and divide substantial portions of machine-learning research into verifiable tasks.
The Transformer serves as a case study in how persistent experimentation and intelligent combination of existing ideas might let automated systems assemble major breakthroughs.
The forecast remains conditional. Continual learning, unknown research objectives, data bottlenecks, and experiment compute could all slow progress even as infrastructure expands.
If software progress reaches a takeoff, automated researchers might also design chips, data centers, and robots, extending acceleration beyond model improvements alone.
What was said on this episode
42 statements · 36 positive · 3 negative · 3 neutral
AI systems may match the best human AI researchers within months or years.
“we're only a couple years away, and possibly only a couple months away, from AIs that are as good as the best humans. at AI research.”
Listen at 0:22
AI systems may match the best human experts across every field by 2030.
“by 2030 and plausibly as early as next year, we'll have AIs that are at least as good as the best human experts in every field.”
Listen at 0:38
AI systems are producing hundreds of previously difficult mathematical results.
“hundreds of important results, which the top mathematicians and computer scientists in the world had spent decades trying to resolve, are being churned out”
Listen at 2:45
AI systems could win an International Math Olympiad gold medal in 2025.
“In 2025, they could win a gold medal in the International Math Olympiad.”
Listen at 3:08
Current AI systems have solved a Millennium Prize problem.
“Current AI systems have solved a millennium price problem”
Listen at 4:03
AI research values empirically useful algorithms over mathematical theory.
“in AI research, the priorities are exactly reversed”
Listen at 4:54
AI systems are especially skilled at persistent combination of ideas.
“persistently and intelligently matching dozens of different combinations of ideas, is exactly the kind of capability that AIs have shown themselves disproportionately skilled at in their contributions in mathematics.”
Listen at 7:09
Cloud7 could have invented the Transformer architecture earlier.
“Cloud7 could have invented”
Listen at 7:31
AI research is sufficiently verifiable for AI systems to match top human performance.
“I think that AI research is verifiable enough that AIs will soon be as good at it as the best humans.”
Listen at 7:50
Automated researchers will match the best human AI researchers around 2028.
“I'd guess 2028”
Listen at 8:26
AI systems can surpass the best humans in verifiable domains within years.
“they've gone from worse than most humans to better than the best humans within just a few years.”
Listen at 8:49
Verifiable training will make AI systems excellent at uncheckable research tasks.
“will make models excellent even at the parts of AI research that we can't directly check.”
Listen at 9:46
There is at least a 25% chance top-human AI research automation comes after 2032.
“there's at least a 25% chance that it doesn't happen until after 2032.”
Listen at 10:04
AI systems could match top human AI researchers by 2027.
“we hit this milestone in 2027.”
Listen at 10:13
Algorithmic progress will accelerate substantially after AI research automation.
“algorithmic progress will get a lot faster”
Listen at 11:18
Each automated researcher could perform roughly 50 times weekly human research work.
“each individual automated researcher will be cramming roughly 50 times more work every calendar week than a human researcher can do.”
Listen at 11:39
The top 50 researchers drive around 80% of frontier-lab algorithmic progress.
“around 80% of the algorithmic progress at a given frontier lab is driven by its 50 top researchers.”
Listen at 12:19
AI research is highly modular and has verifiable progress proxies.
“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.”
Listen at 13:17
AI companies must quadruple R&D compute and double workforce annually to maintain progress.
“AI companies have to quadruple their R&D compute and double their workforce every single year just to maintain the same rate of progress.”
Listen at 15:07
A software-only singularity could produce chip-limited superintelligence within months.
“within a matter of months, we end up with a superintelligence that scrapes against the physical limits of our current chips.”
Listen at 15:30
The Spark scenario would compress three to six years of progress into one year.
“We'd get something like three to six years of AI progress at the current rate in a single calendar year.”
Listen at 15:44
Post-Spark AI systems could outperform NVIDIA at designing AI accelerators.
“it'd be better than NVIDIA at coming up with the next generation of AI accelerators.”
Listen at 17:03
A superintelligent system could outperform Citadel and Jane Street financially.
“It would make more money than the current human versions of Citadel and Jane Street if it was given equal access to the market.”
Listen at 17:07
A superintelligent system could outperform humanity at finding room-temperature superconductors.
“you'd be better than the rest of humanity at finding a room temperature superconductor.”
Listen at 17:16
Scaling reinforcement learning will help AI systems adapt quickly to new contexts.
“scaling up RL training will give AIs the abilities necessary to ramp up quickly in new contexts”
Listen at 18:22
Recursive self-improvement could produce tens of thousands times more RL environments.
“During RSI... Automated engineers will build tens of thousands of times more RL environments than we have currently.”
Listen at 18:54
Future AI systems could greatly exceed humans in learning efficiency.
“future AIs could be far beyond humans at their learning efficiency.”
Listen at 20:46
Future AI systems could beat top human experts within months of deployment.
“they'll be able to beat top human experts within months of being deployed.”
Listen at 23:16
Post-Spark AI systems could reach top-human performance across domains within months.
“we should expect these future post-Spark AIs to be able to get to... top human performance in all domains within months of being developed.”
Listen at 24:48
Global AI compute has grown roughly threefold annually for five years.
“global AI compute has been growing roughly 3x a year. for the last half decade.”
Listen at 25:18
Post-Spark AI systems will excel at chip, fab, and data-center construction tasks.
“They'll be extremely good at designing new AI chips and laying out the designs of new fabs, even teleoperating robots to help build parts of fabs and data centers.”
Listen at 25:29
Every form of AI sample efficiency is measurable.
“every kind of sample efficiency is measurable.”
Listen at 26:35
AI progress generally can be directly optimized through training objectives.
“that means that AI progress in general is easy to train against”
Listen at 27:48
Important aspects of AI research taste are verifiable.
“even important parts of research tastes are verifiable.”
Listen at 29:49
Pre-training produces broad improvements in model intelligence and capabilities.
“pre-training leads to very general improvements in the model's intelligence and capabilities.”
Listen at 30:00
AI pre-training efficiency could improve by orders of magnitude.
“Orders of magnitude further improvements seem totally plausible.”
Listen at 31:57
AI capabilities could jump from GPT-3.5 to GPT-6 scale in 2028.
“we'll probably make something like a GPT 3.5 to GPT 6 size jump in 2028”
Listen at 33:44
Scaling in-context learning may solve continual learning limitations.
“scaling up in-context learning is enough.”
Listen at 35:13
There is a 25% chance top-human AI research automation occurs after 2032.
“That's why I put a 25% probability that we won't automate ILIA-level top human researchers until after 2032.”
Listen at 36:35
Pre-paradigm proxies can capture major performance improvements after paradigm shifts.
“Proxies made before a paradigm leap seem perfectly capable of capturing large step changes in performance.”
Listen at 37:49
Data improvements produced over three times more pre-training progress than model improvements.
“for pre-training, Over the last six years, more than three times the algorithmic progress has come from data improvements than model improvements.”
Listen at 38:57
Frontier AI labs have more than quadrupled compute annually.
“Frontier Labs have been more than quadrupling their compute every single year.”
Listen at 40:37
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
Some links are affiliate links — PodLume may earn a commission if you buy.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.
