AlphaProof: when reinforcement learning meets formal mathematics. ~ Thomas Hubert. youtu.be/TFBzP78Jp6A #AlphaProof #AI #Math
Thomas Hubert | AlphaProof: when reinforcement learning meets formal mathematics
YouTube video by Harvard CMSA
youtu.be
April 3, 2025 at 9:31 AM
AlphaProof: when reinforcement learning meets formal mathematics. ~ Thomas Hubert. youtu.be/TFBzP78Jp6A #AlphaProof #AI #Math
Calculating π is a deterministic, primitive recursive recipe—no branching, guaranteed convergence. Proof discovery is undecidable (Church-Turing). Enumerating strings is intractable; what systems like AlphaProof do is heuristic search space pruning. Conflating the two is a category error.
September 25, 2026 at 3:36 AM
Calculating π is a deterministic, primitive recursive recipe—no branching, guaranteed convergence. Proof discovery is undecidable (Church-Turing). Enumerating strings is intractable; what systems like AlphaProof do is heuristic search space pruning. Conflating the two is a category error.
It seems like a cheap shot that will mislead most people to link to AlphaProof then imply that the paper refutes the AlphaProof claims.
March 31, 2025 at 9:44 PM
It seems like a cheap shot that will mislead most people to link to AlphaProof then imply that the paper refutes the AlphaProof claims.
I keep reading interviews about how impressively smart and accomplished the AlphaProof team is. But I would be way more impressed if they were just like "we're super bad at math, who cares".
November 18, 2024 at 2:50 PM
I keep reading interviews about how impressively smart and accomplished the AlphaProof team is. But I would be way more impressed if they were just like "we're super bad at math, who cares".
The "LLMs are just fancy autocomplete" crowd has been wrong about basically everything, but they have been very successful at getting people to ignore/downplay AI progress.
I still meet smart people who say stuff like "AI can't do basic math," meanwhile AlphaProof won a silver medal at the IMO.
I still meet smart people who say stuff like "AI can't do basic math," meanwhile AlphaProof won a silver medal at the IMO.
November 19, 2024 at 5:13 PM
The "LLMs are just fancy autocomplete" crowd has been wrong about basically everything, but they have been very successful at getting people to ignore/downplay AI progress.
I still meet smart people who say stuff like "AI can't do basic math," meanwhile AlphaProof won a silver medal at the IMO.
I still meet smart people who say stuff like "AI can't do basic math," meanwhile AlphaProof won a silver medal at the IMO.
This lecture by Thomas Hubert of DeepMind gives some interesting details about how AlphaProof was trained for the 2024 International Mathematical Olympiad with medal wining performance, using Lean and reinforcement learning.
www.youtube.com/watch?v=TFBz...
www.youtube.com/watch?v=TFBz...
Thomas Hubert | AlphaProof: when reinforcement learning meets formal mathematics
YouTube video by Harvard CMSA
www.youtube.com
May 9, 2025 at 3:16 AM
This lecture by Thomas Hubert of DeepMind gives some interesting details about how AlphaProof was trained for the 2024 International Mathematical Olympiad with medal wining performance, using Lean and reinforcement learning.
www.youtube.com/watch?v=TFBz...
www.youtube.com/watch?v=TFBz...
Het is vooral Google Deepmind die hier custom modellen voor maakt (AlphaProof, zie github.com/google-deepm...). Zowel Anthropic als OpenAI claimen dit met een general purpose LLM gedaan te hebben.
GitHub - google-deepmind/alphaproof-nexus-results: Lean math proofs generated by AlphaProof Nexus and accompanying natural language prose proofs.
Lean math proofs generated by AlphaProof Nexus and accompanying natural language prose proofs. - google-deepmind/alphaproof-nexus-results
github.com
May 26, 2026 at 8:01 PM
Het is vooral Google Deepmind die hier custom modellen voor maakt (AlphaProof, zie github.com/google-deepm...). Zowel Anthropic als OpenAI claimen dit met een general purpose LLM gedaan te hebben.
We 💙 this talk by DeepMind's Thomas Hubert on AlphaProof combining Lean's formal verification with AlphaGo-style RL. Hubert believes formal mathematics is "inevitable" in math's future, just as software is in our world.
Watch: www.youtube.com/watch?v=TFBz...
#LeanLang #LeanProver #AlphaProof
Watch: www.youtube.com/watch?v=TFBz...
#LeanLang #LeanProver #AlphaProof
Thomas Hubert | AlphaProof: when reinforcement learning meets formal mathematics
YouTube video by Harvard CMSA
www.youtube.com
April 4, 2025 at 2:02 PM
We 💙 this talk by DeepMind's Thomas Hubert on AlphaProof combining Lean's formal verification with AlphaGo-style RL. Hubert believes formal mathematics is "inevitable" in math's future, just as software is in our world.
Watch: www.youtube.com/watch?v=TFBz...
#LeanLang #LeanProver #AlphaProof
Watch: www.youtube.com/watch?v=TFBz...
#LeanLang #LeanProver #AlphaProof
An AI model that is trained to use computational tools might accelerate mathematical discovery
go.nature.com/4pd11nA
go.nature.com/4pd11nA
Mathematicians put AI model AlphaProof to the test
Mathematicians use computational tools to prove theorems. An AI model that is trained to use these tools might accelerate mathematical discovery.
go.nature.com
November 17, 2025 at 9:43 AM
An AI model that is trained to use computational tools might accelerate mathematical discovery
go.nature.com/4pd11nA
go.nature.com/4pd11nA
It's so annoying that people (even within DeepMind) keep saying that OpenAI is ahead "in AI", when DeepMind has models like AlphaFold, AlphaProof, GenCast and GraphCast, FunSearch, etc
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
September 21, 2024 at 7:15 PM
It's so annoying that people (even within DeepMind) keep saying that OpenAI is ahead "in AI", when DeepMind has models like AlphaFold, AlphaProof, GenCast and GraphCast, FunSearch, etc
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
AI for Science has such a clearer potential for helping people, but VCs and the media are obsessed by GenAI
Spent last week @ias with DeepMind math teams. AlphaProof, AlphaGeometry, PDEs, and PAIRs teams were there with leadership. Impressed by the collaboration. Huge, open technical problems to be solved. Exciting times for mathematics & AI!
www.ias.edu/math/events/...
www.ias.edu/math/events/...
May 6, 2025 at 12:03 PM
Spent last week @ias with DeepMind math teams. AlphaProof, AlphaGeometry, PDEs, and PAIRs teams were there with leadership. Impressed by the collaboration. Huge, open technical problems to be solved. Exciting times for mathematics & AI!
www.ias.edu/math/events/...
www.ias.edu/math/events/...
My inevitable take on this: compare these abysmal scores detailed in the paper with the DeepMind AlphaProof model which writes Lean code instead of natural language; the system correctly formalized the solution to 3/6 of the 2024 IMO problems. LLMs need to learn formal like they learnt python.
April 1, 2025 at 10:34 PM
My inevitable take on this: compare these abysmal scores detailed in the paper with the DeepMind AlphaProof model which writes Lean code instead of natural language; the system correctly formalized the solution to 3/6 of the 2024 IMO problems. LLMs need to learn formal like they learnt python.
Yesterday was my last day at Google DeepMind. Interning there when the first GDM Nobel prize, AlphaProof, the Gemini releases and more all happened and having such amazing/ambitious colleagues was quite humbling and really exciting. I look forward to 2025!
December 21, 2024 at 10:51 AM
Yesterday was my last day at Google DeepMind. Interning there when the first GDM Nobel prize, AlphaProof, the Gemini releases and more all happened and having such amazing/ambitious colleagues was quite humbling and really exciting. I look forward to 2025!
LLMs are just fancy autocomplete. AlphaProof is not an LLM.
Straight from the blog post: "First, the problems were manually translated into formal mathematical language for our systems to understand."
Straight from the blog post: "First, the problems were manually translated into formal mathematical language for our systems to understand."
November 19, 2024 at 5:48 PM
LLMs are just fancy autocomplete. AlphaProof is not an LLM.
Straight from the blog post: "First, the problems were manually translated into formal mathematical language for our systems to understand."
Straight from the blog post: "First, the problems were manually translated into formal mathematical language for our systems to understand."
A paper in Nature describes an AI system that can prove complex mathematical theories. The system, Google DeepMind’s AlphaProof, has achieved a silver medal-worthy performance at the 2024 International Mathematics Olympiad. go.nature.com/4qUObvN 🧪
November 12, 2025 at 11:18 PM
A paper in Nature describes an AI system that can prove complex mathematical theories. The system, Google DeepMind’s AlphaProof, has achieved a silver medal-worthy performance at the 2024 International Mathematics Olympiad. go.nature.com/4qUObvN 🧪
This is a bit misleading. The first link is to AlphaProof/AlphaGeometry 2, which are not general LLMs like the ones shown in the image—they are neurosymbolic models trained specifically for Olympiad-type problems.
I don't think it comes as a surprise that general-purpose LLMs still struggle.
I don't think it comes as a surprise that general-purpose LLMs still struggle.
April 2, 2025 at 9:46 AM
This is a bit misleading. The first link is to AlphaProof/AlphaGeometry 2, which are not general LLMs like the ones shown in the image—they are neurosymbolic models trained specifically for Olympiad-type problems.
I don't think it comes as a surprise that general-purpose LLMs still struggle.
I don't think it comes as a surprise that general-purpose LLMs still struggle.
This approach (e.g. deep learning as heuristics or formalizers for a symbolic solver) can certainly be productive; e.g. AlphaProof relied on it for last year's silver. But I think it gets things precisely backwards about real intelligence.
July 21, 2025 at 10:20 PM
This approach (e.g. deep learning as heuristics or formalizers for a symbolic solver) can certainly be productive; e.g. AlphaProof relied on it for last year's silver. But I think it gets things precisely backwards about real intelligence.
Oh hell yeah. I took artificial neural networks in college. We've BEEN using some of this shit for a long time. The AlphaProof/AlphaGo stuff is absolutely crazy cool. It's the *specific* stand-alone inefficient as hell general-purpose LLMs and the hype train that drives me bonkers.
April 29, 2026 at 11:32 PM
Oh hell yeah. I took artificial neural networks in college. We've BEEN using some of this shit for a long time. The AlphaProof/AlphaGo stuff is absolutely crazy cool. It's the *specific* stand-alone inefficient as hell general-purpose LLMs and the hype train that drives me bonkers.
"Axiom’s approach involves combining large language models with a proprietary AI system called AxiomProver that is trained to reason through math problems to reach solutions that are provably correct. In 2024, Google demonstrated a similar idea with a system called AlphaProof."
A New AI Math Startup Just Cracked 4 Previously Unsolved Problems
Axiom says its AI found solutions to several long-standing math problems, a sign of the technology’s steadily advancing reasoning capabilities.
www.wired.com
February 6, 2026 at 12:42 AM
"Axiom’s approach involves combining large language models with a proprietary AI system called AxiomProver that is trained to reason through math problems to reach solutions that are provably correct. In 2024, Google demonstrated a similar idea with a system called AlphaProof."
Some neat results from the automated mathematical prover AlphaPtoof. I would love to see this applied to cryptographic protocols. rishimehta.xyz/2024/11/17/a...
AlphaProof’s Greatest Hits
Here I’ll try to explain the coolest ideas in each of AlphaProof’s IMO 2024 solutions. AlphaProof produces proofs in Lean, and each Lean proof is composed of a series of tactics. So I’ll pick out the ...
rishimehta.xyz
November 19, 2024 at 12:03 PM
Some neat results from the automated mathematical prover AlphaPtoof. I would love to see this applied to cryptographic protocols. rishimehta.xyz/2024/11/17/a...
AI just leveled up in math—and not just any math, but the legendary International Mathematical Olympiad (IMO)! 🤖✨
DeepMind’s AlphaProof and AlphaGeometry 2 tackled this year’s IMO problems, solving 4 out of 6—matching the silver medalist tier for the first time ever.
DeepMind’s AlphaProof and AlphaGeometry 2 tackled this year’s IMO problems, solving 4 out of 6—matching the silver medalist tier for the first time ever.
July 22, 2025 at 4:32 PM
AI just leveled up in math—and not just any math, but the legendary International Mathematical Olympiad (IMO)! 🤖✨
DeepMind’s AlphaProof and AlphaGeometry 2 tackled this year’s IMO problems, solving 4 out of 6—matching the silver medalist tier for the first time ever.
DeepMind’s AlphaProof and AlphaGeometry 2 tackled this year’s IMO problems, solving 4 out of 6—matching the silver medalist tier for the first time ever.
DeepMind’s latest: An AI for handling mathematical proofs. ~ Jacek Krywko. arstechnica.com/ai/2025/11/d... #AI #Math #LLMs #ITP #LeanProver #AlphaProof
DeepMind’s latest: An AI for handling mathematical proofs
AlphaProof can handle math challenges but needs a bit of help right now.
arstechnica.com
November 21, 2025 at 10:16 AM
DeepMind’s latest: An AI for handling mathematical proofs. ~ Jacek Krywko. arstechnica.com/ai/2025/11/d... #AI #Math #LLMs #ITP #LeanProver #AlphaProof
New models from Google DeepMind turned AI into a serious competitor in the International Mathematical Olympiad. When their model AlphaProof achieved a silver medal on the exam, it showcased AI’s fast-growing mathematical prowess. www.scientificamerican.com/article/ai-w...
AI Will Become Mathematicians’ ‘Co-Pilot’
Fields Medalist Terence Tao explains how proof checkers and AI programs are dramatically changing mathematics
www.scientificamerican.com
December 16, 2024 at 9:41 PM
New models from Google DeepMind turned AI into a serious competitor in the International Mathematical Olympiad. When their model AlphaProof achieved a silver medal on the exam, it showcased AI’s fast-growing mathematical prowess. www.scientificamerican.com/article/ai-w...
Hey, this is really cool. Still a ton of computation power used but at least someone is using an LLM in what seems like an appropriate way—as one component of a system that has other components that verify its output!
www.nytimes.com/2024/07/25/s...
www.nytimes.com/2024/07/25/s...
Move Over, Mathematicians, Here Comes AlphaProof
A.I. is getting good at math — and might soon make a worthy collaborator for humans.
www.nytimes.com
July 25, 2024 at 7:43 PM
Hey, this is really cool. Still a ton of computation power used but at least someone is using an LLM in what seems like an appropriate way—as one component of a system that has other components that verify its output!
www.nytimes.com/2024/07/25/s...
www.nytimes.com/2024/07/25/s...