<![CDATA[MKLab - ARTICLES]]> https://mklab.gr/ Tue, 27 Oct 2026 20:40:24 +0000 MyBB <![CDATA[The Classification of Finite Simple Groups, formalized in Lean]]> https://mklab.gr/showthread.php?tid=2354 Tue, 27 Oct 2026 22:04:30 +0200 mklabgr]]> https://mklab.gr/showthread.php?tid=2354
Axiom Math reports that it has now completed a formal computer-verified version of the entire classification proof, using AI agents and the proof-assistant language Lean. According to the project website, the formalization comprises approximately 43.1 million lines of Lean code, 2.05 million declarations, 100 milestones, and material drawn from 715 papers and books. Lean’s proof-checking kernel verifies the logical steps, allowing mathematicians to inspect a machine-readable version of the argument rather than relying exclusively on thousands of pages of conventional mathematical proofs.

The achievement is significant not simply because of its scale, but because the formalization process also exposed a problem in a published argument concerning the Unbalanced Group Theorem, prompting the team to develop an alternative approach for the relevant case. Nevertheless, formal verification does not mean that AI independently discovered the entire classification theorem: the underlying mathematics represents decades of human research, and the formalization follows the existing proof literature. The project provides a major opportunity for independent mathematical scrutiny and could help researchers build more reliable, reusable formal proofs. Its broader significance is that AI may increasingly assist mathematicians with large-scale research and rigorous verification, while human expertise remains essential for interpreting results and evaluating their mathematical significance.

ARTICLE]]>
Axiom Math reports that it has now completed a formal computer-verified version of the entire classification proof, using AI agents and the proof-assistant language Lean. According to the project website, the formalization comprises approximately 43.1 million lines of Lean code, 2.05 million declarations, 100 milestones, and material drawn from 715 papers and books. Lean’s proof-checking kernel verifies the logical steps, allowing mathematicians to inspect a machine-readable version of the argument rather than relying exclusively on thousands of pages of conventional mathematical proofs.

The achievement is significant not simply because of its scale, but because the formalization process also exposed a problem in a published argument concerning the Unbalanced Group Theorem, prompting the team to develop an alternative approach for the relevant case. Nevertheless, formal verification does not mean that AI independently discovered the entire classification theorem: the underlying mathematics represents decades of human research, and the formalization follows the existing proof literature. The project provides a major opportunity for independent mathematical scrutiny and could help researchers build more reliable, reusable formal proofs. Its broader significance is that AI may increasingly assist mathematicians with large-scale research and rigorous verification, while human expertise remains essential for interpreting results and evaluating their mathematical significance.

ARTICLE]]> <![CDATA[OpenAI just scooped mathematicians’ work on another million-dollar problem]]> https://mklab.gr/showthread.php?tid=2353 Tue, 27 Oct 2026 21:58:31 +0200 mklabgr]]> https://mklab.gr/showthread.php?tid=2353
The announcement also overshadowed the parallel work of mathematicians Levent Alpöge of Anthropic and Ethan Sussman of Northwestern University, who had been using AI to develop a proof based on earlier work by Tadeusz Balaban. Their approach employed multiple AI agents to generate extensive mathematical arguments, check their validity, fill gaps in existing research, and identify potentially novel ideas. The researchers believed they had made substantial progress shortly before OpenAI announced its result. OpenAI stated that its internal language model had obtained the result earlier in September and that the company subsequently worked to formalize it using Lean, a proof-assistant system designed to help verify mathematical correctness.

Beyond the mathematical breakthrough, the episode raises important questions about AI, scientific credit, transparency, and competition between research organizations. The article highlights concerns that commercial secrecy and the race for technological dominance may undermine established academic practices, particularly when researchers independently pursue the same problem and the origins of AI-generated ideas are difficult to establish. While experts see enormous potential for AI to accelerate mathematical research by handling lengthy and technically demanding proofs, the announcement still requires careful independent evaluation. The central message is that AI may be transforming mathematical discovery, but the scientific community must develop clearer standards for verification, authorship, attribution, and the publication of AI-assisted proofs.

ARTICLE]]>
The announcement also overshadowed the parallel work of mathematicians Levent Alpöge of Anthropic and Ethan Sussman of Northwestern University, who had been using AI to develop a proof based on earlier work by Tadeusz Balaban. Their approach employed multiple AI agents to generate extensive mathematical arguments, check their validity, fill gaps in existing research, and identify potentially novel ideas. The researchers believed they had made substantial progress shortly before OpenAI announced its result. OpenAI stated that its internal language model had obtained the result earlier in September and that the company subsequently worked to formalize it using Lean, a proof-assistant system designed to help verify mathematical correctness.

Beyond the mathematical breakthrough, the episode raises important questions about AI, scientific credit, transparency, and competition between research organizations. The article highlights concerns that commercial secrecy and the race for technological dominance may undermine established academic practices, particularly when researchers independently pursue the same problem and the origins of AI-generated ideas are difficult to establish. While experts see enormous potential for AI to accelerate mathematical research by handling lengthy and technically demanding proofs, the announcement still requires careful independent evaluation. The central message is that AI may be transforming mathematical discovery, but the scientific community must develop clearer standards for verification, authorship, attribution, and the publication of AI-assisted proofs.

ARTICLE]]> <![CDATA[Google DeepMind and Springer Nature to create AI peer review tool]]> https://mklab.gr/showthread.php?tid=2352 Tue, 27 Oct 2026 00:20:35 +0200 mklabgr]]> https://mklab.gr/showthread.php?tid=2352

Human expertise will remain central to the review process. The prototype, which is currently being tested, is intended to support rather than replace human reviewers. The project combines Springer Nature’s experience in academic publishing with Google DeepMind’s AI technology and builds on Google’s research tools, including ScholarPeer and Deep Review. The system uses Google’s open-source Agent Development Kit and operates within Springer Nature’s private Gemini Enterprise environment. According to the publisher, this arrangement is designed to protect the confidentiality of manuscripts and reviewer information during development and testing.

The partnership reflects a broader effort to integrate AI responsibly into scientific publishing. Springer Nature describes its approach as “human first, AI second”, emphasising that editorial responsibility and scientific judgement must remain with people. The expected benefits include a more efficient peer-review process, reduced pressure on reviewers and the preservation of editorial standards. However, the announcement concerns a prototype rather than a fully deployed product; its practical effectiveness and impact on the quality of scientific evaluation remain to be established through further testing. No timetable for wider implementation has been announced.


Key takeaway: AI is moving beyond assisting researchers in writing and analysing papers towards supporting the institutional processes that determine which scientific findings are published. The central challenge will be to improve efficiency without compromising scientific rigour, reviewer independence, confidentiality and human accountability.

Original article: Springer Nature and Google DeepMind collaborate to enhance the peer review experience]]>

Human expertise will remain central to the review process. The prototype, which is currently being tested, is intended to support rather than replace human reviewers. The project combines Springer Nature’s experience in academic publishing with Google DeepMind’s AI technology and builds on Google’s research tools, including ScholarPeer and Deep Review. The system uses Google’s open-source Agent Development Kit and operates within Springer Nature’s private Gemini Enterprise environment. According to the publisher, this arrangement is designed to protect the confidentiality of manuscripts and reviewer information during development and testing.

The partnership reflects a broader effort to integrate AI responsibly into scientific publishing. Springer Nature describes its approach as “human first, AI second”, emphasising that editorial responsibility and scientific judgement must remain with people. The expected benefits include a more efficient peer-review process, reduced pressure on reviewers and the preservation of editorial standards. However, the announcement concerns a prototype rather than a fully deployed product; its practical effectiveness and impact on the quality of scientific evaluation remain to be established through further testing. No timetable for wider implementation has been announced.


Key takeaway: AI is moving beyond assisting researchers in writing and analysing papers towards supporting the institutional processes that determine which scientific findings are published. The central challenge will be to improve efficiency without compromising scientific rigour, reviewer independence, confidentiality and human accountability.

Original article: Springer Nature and Google DeepMind collaborate to enhance the peer review experience]]> <![CDATA[A Mathematician Reads the AI Future]]> https://mklab.gr/showthread.php?tid=2343 Sun, 25 Oct 2026 17:33:46 +0200 mklabgr]]> https://mklab.gr/showthread.php?tid=2343 In The Atlantic article, the author examines how the rapid progress of artificial intelligence (AI) is transforming mathematics and challenging traditional ideas about mathematical discovery. In 2026, AI systems have reportedly generated solutions to hundreds of longstanding mathematical problems, raising expectations about how quickly new results can be obtained. These developments offer enormous opportunities to accelerate research, explore difficult conjectures, and tackle problems that have resisted human efforts for decades. However, they have also created uncertainty among mathematicians about the future of their profession and the role of human intelligence in a field traditionally associated with creativity, persistence, and intellectual achievement.
The article highlights a fundamental distinction between producing a mathematical proof and understanding the ideas behind it. Although AI can generate solutions at remarkable speed, its results may be difficult to verify, interpret, or connect to broader mathematical theories. For mathematicians, the process of solving a problem is valuable not only because it produces a theorem but also because it develops intuition, reveals unexpected connections, and creates new methods that can be applied elsewhere. The increasing automation of mathematical discovery therefore raises concerns about mathematical education, research careers, and opportunities for young researchers to develop their skills. If machines produce results faster than humans can understand them, the mathematical community risks accumulating proofs without gaining an equivalent level of insight.
Ultimately, the article raises a deeper question about whether mathematics is primarily the discovery of objective truths or also a distinctly human creative activity. AI may eventually outperform people in many aspects of problem-solving, but mathematical research involves more than obtaining correct answers: it also includes curiosity, collaboration, intellectual satisfaction, and the search for explanations that make complex ideas understandable. The challenge is to adapt mathematical practices so that AI becomes a tool for enhancing human understanding rather than simply replacing human effort. This may require changes in research incentives, mathematical education, and the way discoveries are evaluated, with greater emphasis on explanation, interpretation, and meaningful contributions to knowledge. The future of mathematics will depend not only on what AI can prove but also on how humans learn from, understand, and use those results.


ARTICLE
]]> In The Atlantic article, the author examines how the rapid progress of artificial intelligence (AI) is transforming mathematics and challenging traditional ideas about mathematical discovery. In 2026, AI systems have reportedly generated solutions to hundreds of longstanding mathematical problems, raising expectations about how quickly new results can be obtained. These developments offer enormous opportunities to accelerate research, explore difficult conjectures, and tackle problems that have resisted human efforts for decades. However, they have also created uncertainty among mathematicians about the future of their profession and the role of human intelligence in a field traditionally associated with creativity, persistence, and intellectual achievement.
The article highlights a fundamental distinction between producing a mathematical proof and understanding the ideas behind it. Although AI can generate solutions at remarkable speed, its results may be difficult to verify, interpret, or connect to broader mathematical theories. For mathematicians, the process of solving a problem is valuable not only because it produces a theorem but also because it develops intuition, reveals unexpected connections, and creates new methods that can be applied elsewhere. The increasing automation of mathematical discovery therefore raises concerns about mathematical education, research careers, and opportunities for young researchers to develop their skills. If machines produce results faster than humans can understand them, the mathematical community risks accumulating proofs without gaining an equivalent level of insight.
Ultimately, the article raises a deeper question about whether mathematics is primarily the discovery of objective truths or also a distinctly human creative activity. AI may eventually outperform people in many aspects of problem-solving, but mathematical research involves more than obtaining correct answers: it also includes curiosity, collaboration, intellectual satisfaction, and the search for explanations that make complex ideas understandable. The challenge is to adapt mathematical practices so that AI becomes a tool for enhancing human understanding rather than simply replacing human effort. This may require changes in research incentives, mathematical education, and the way discoveries are evaluated, with greater emphasis on explanation, interpretation, and meaningful contributions to knowledge. The future of mathematics will depend not only on what AI can prove but also on how humans learn from, understand, and use those results.


ARTICLE
]]> <![CDATA[Open Math OS]]> https://mklab.gr/showthread.php?tid=2341 Fri, 23 Oct 2026 18:20:41 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2341
The initiative's main platform, Open Math OS, integrates AI models, cloud computing, mathematical software, and collaboration tools into a single research environment. Researchers can use multiple AI agents to explore different approaches to a problem simultaneously through a feature called Ultramath. The platform also provides access to more than 100 mathematical and scientific computing tools, including Lean, Mathlib, SageMath, PARI/GP, SymPy, Python, and LaTeX. These resources enable mathematicians to test conjectures, perform symbolic calculations, write research papers, and formally verify proofs. Importantly, AI-generated solutions are treated as proposals that require human assessment; formal proof assistants such as Lean can help establish whether formalized arguments are mathematically valid.

A central principle of Open Math Model is that mathematicians should retain control over their research, data, and intellectual contributions. According to the initiative, projects, files, and conversations remain private by default, and researchers' work will be used for AI training only with their explicit consent under agreed terms. The project also advocates open development, appropriate licensing, transparent governance, and recognition of researchers' contributions. By combining AI-assisted reasoning with scientific computing and formal verification, Open Math Model aims to make mathematical research more efficient without replacing human judgment or compromising academic independence. However, these are the initiative's stated objectives; its long-term impact on mathematical discovery will depend on the reliability of its tools, independent evaluation, and their demonstrated usefulness in real research.


Key takeaways
  • AI-assisted mathematics: Multiple AI agents can investigate mathematical problems in parallel.
  • Rigorous verification: Lean and Mathlib help check formalized mathematical proofs.
  • Integrated research environment: AI models, computation, coding, and academic writing are brought together.
  • Researcher ownership: Data privacy, explicit consent for model training, and intellectual credit are core commitments.
  • Open scientific collaboration: The initiative aims to make advanced mathematical research tools more accessible.
  • Important distinction: Open Math Model is a research infrastructure initiative, not evidence that AI has independently solved major unsolved mathematical problems.

Official website: Open Math Model]]>
The initiative's main platform, Open Math OS, integrates AI models, cloud computing, mathematical software, and collaboration tools into a single research environment. Researchers can use multiple AI agents to explore different approaches to a problem simultaneously through a feature called Ultramath. The platform also provides access to more than 100 mathematical and scientific computing tools, including Lean, Mathlib, SageMath, PARI/GP, SymPy, Python, and LaTeX. These resources enable mathematicians to test conjectures, perform symbolic calculations, write research papers, and formally verify proofs. Importantly, AI-generated solutions are treated as proposals that require human assessment; formal proof assistants such as Lean can help establish whether formalized arguments are mathematically valid.

A central principle of Open Math Model is that mathematicians should retain control over their research, data, and intellectual contributions. According to the initiative, projects, files, and conversations remain private by default, and researchers' work will be used for AI training only with their explicit consent under agreed terms. The project also advocates open development, appropriate licensing, transparent governance, and recognition of researchers' contributions. By combining AI-assisted reasoning with scientific computing and formal verification, Open Math Model aims to make mathematical research more efficient without replacing human judgment or compromising academic independence. However, these are the initiative's stated objectives; its long-term impact on mathematical discovery will depend on the reliability of its tools, independent evaluation, and their demonstrated usefulness in real research.


Key takeaways
  • AI-assisted mathematics: Multiple AI agents can investigate mathematical problems in parallel.
  • Rigorous verification: Lean and Mathlib help check formalized mathematical proofs.
  • Integrated research environment: AI models, computation, coding, and academic writing are brought together.
  • Researcher ownership: Data privacy, explicit consent for model training, and intellectual credit are core commitments.
  • Open scientific collaboration: The initiative aims to make advanced mathematical research tools more accessible.
  • Important distinction: Open Math Model is a research infrastructure initiative, not evidence that AI has independently solved major unsolved mathematical problems.

Official website: Open Math Model]]> <![CDATA[AI is great at math, but does it know how to do research?]]> https://mklab.gr/showthread.php?tid=2336 Thu, 22 Oct 2026 23:13:34 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2336 Publication date: October 13, 2026 | Source: Princeton University, Department of Computer Science


1 — AI's growing role in mathematical research
Artificial intelligence is becoming a powerful tool for mathematical research, demonstrating an impressive ability to solve difficult problems and explore mathematical questions. According to a case study by theoretical computer scientists from Princeton University, UCLA and the University of Texas at Austin, AI systems helped improve the known bounds of the Grothendieck constant, a mathematical quantity introduced through the work of Alexander Grothendieck in 1953. The research illustrates how AI can contribute to problems that might otherwise require years of human effort. The researchers reported that AI performed much of the mathematical work, while humans primarily provided guidance when the systems encountered difficulties. Their findings suggest that AI could significantly accelerate discoveries in areas such as graph theory, theoretical computer science and quantum information.

2 — The limitations of AI in mathematical discovery
Despite its impressive computational and problem-solving capabilities, AI still faces important challenges when conducting original mathematical research. The article distinguishes between finding a solution and exercising the research judgment needed to determine whether a result is mathematically significant, sufficiently general or useful for future discoveries. In the Grothendieck constant project, the researchers combined a reasoning model with a coding agent that checked proposed mathematical work. The AI systems repeatedly attempted to solve the problem but eventually became stuck in a cycle of similar approaches. Human researchers intervened by reframing the question, enabling the process to continue and eventually produce improved bounds. This example highlights the importance of human insight, strategic decisions and the ability to recognize when a different approach is necessary. The underlying case study was published on arXiv as a preprint, meaning it had not yet undergone peer review at the time described in the article.


3 — The future of mathematics and human–AI collaboration
The increasing capabilities of AI raise important questions about the future of mathematical research, education and the role of mathematicians. Although AI can generate solutions and even lengthy formal proofs, a result is not necessarily a meaningful scientific contribution simply because its logical correctness can be verified. Mathematicians must also be able to understand, explain and communicate proofs so that other researchers can evaluate their significance and build upon them. AI also challenges traditional mathematical education: if machines can solve many introductory research problems, educators may need to rethink how students develop problem-solving skills and gain research experience. The article therefore presents AI most constructively as a research assistant rather than a complete replacement for human mathematicians. Its long-term value will depend not only on its ability to solve difficult problems, but also on whether its discoveries improve mathematical understanding and support the development of new theories.

Key Takeaways
  • AI can accelerate mathematical research: AI systems can explore difficult problems continuously and potentially reduce research timelines from years to weeks.
  • Solving problems is not the same as doing research: Successful mathematical research requires identifying important questions, evaluating results and deciding which directions are worth pursuing.
  • Human guidance remains valuable: Researchers can help AI overcome repetitive approaches by reframing questions and suggesting alternative strategies.
  • Verification is essential: Mathematical claims need rigorous checking, including computational validation or formal proof verification where appropriate.
  • Correctness is not the same as understanding: A formally verified proof may still be difficult for mathematicians to interpret, explain or connect to existing theories.
  • Mathematics education may need to adapt: Teachers and universities must reconsider how students develop independent reasoning, mathematical intuition and research skills in an era of increasingly capable AI.
  • Collaboration is a promising direction: Combining AI's speed and persistence with human creativity, judgment and understanding may lead to more productive mathematical discoveries.
  • Research status matters: The Grothendieck constant study was a preprint at the time of the article, so its claims should be distinguished from findings established through peer review.

Overall Conclusion
The central message is that AI is becoming increasingly capable of discovering mathematical results, but meaningful mathematical research involves more than producing answers. The greatest opportunities may arise from collaboration between AI systems and human mathematicians, ensuring that new results are not only correct but also understandable, significant and useful for future research.

ARTICLE]]> Publication date: October 13, 2026 | Source: Princeton University, Department of Computer Science


1 — AI's growing role in mathematical research
Artificial intelligence is becoming a powerful tool for mathematical research, demonstrating an impressive ability to solve difficult problems and explore mathematical questions. According to a case study by theoretical computer scientists from Princeton University, UCLA and the University of Texas at Austin, AI systems helped improve the known bounds of the Grothendieck constant, a mathematical quantity introduced through the work of Alexander Grothendieck in 1953. The research illustrates how AI can contribute to problems that might otherwise require years of human effort. The researchers reported that AI performed much of the mathematical work, while humans primarily provided guidance when the systems encountered difficulties. Their findings suggest that AI could significantly accelerate discoveries in areas such as graph theory, theoretical computer science and quantum information.

2 — The limitations of AI in mathematical discovery
Despite its impressive computational and problem-solving capabilities, AI still faces important challenges when conducting original mathematical research. The article distinguishes between finding a solution and exercising the research judgment needed to determine whether a result is mathematically significant, sufficiently general or useful for future discoveries. In the Grothendieck constant project, the researchers combined a reasoning model with a coding agent that checked proposed mathematical work. The AI systems repeatedly attempted to solve the problem but eventually became stuck in a cycle of similar approaches. Human researchers intervened by reframing the question, enabling the process to continue and eventually produce improved bounds. This example highlights the importance of human insight, strategic decisions and the ability to recognize when a different approach is necessary. The underlying case study was published on arXiv as a preprint, meaning it had not yet undergone peer review at the time described in the article.


3 — The future of mathematics and human–AI collaboration
The increasing capabilities of AI raise important questions about the future of mathematical research, education and the role of mathematicians. Although AI can generate solutions and even lengthy formal proofs, a result is not necessarily a meaningful scientific contribution simply because its logical correctness can be verified. Mathematicians must also be able to understand, explain and communicate proofs so that other researchers can evaluate their significance and build upon them. AI also challenges traditional mathematical education: if machines can solve many introductory research problems, educators may need to rethink how students develop problem-solving skills and gain research experience. The article therefore presents AI most constructively as a research assistant rather than a complete replacement for human mathematicians. Its long-term value will depend not only on its ability to solve difficult problems, but also on whether its discoveries improve mathematical understanding and support the development of new theories.

Key Takeaways
  • AI can accelerate mathematical research: AI systems can explore difficult problems continuously and potentially reduce research timelines from years to weeks.
  • Solving problems is not the same as doing research: Successful mathematical research requires identifying important questions, evaluating results and deciding which directions are worth pursuing.
  • Human guidance remains valuable: Researchers can help AI overcome repetitive approaches by reframing questions and suggesting alternative strategies.
  • Verification is essential: Mathematical claims need rigorous checking, including computational validation or formal proof verification where appropriate.
  • Correctness is not the same as understanding: A formally verified proof may still be difficult for mathematicians to interpret, explain or connect to existing theories.
  • Mathematics education may need to adapt: Teachers and universities must reconsider how students develop independent reasoning, mathematical intuition and research skills in an era of increasingly capable AI.
  • Collaboration is a promising direction: Combining AI's speed and persistence with human creativity, judgment and understanding may lead to more productive mathematical discoveries.
  • Research status matters: The Grothendieck constant study was a preprint at the time of the article, so its claims should be distinguished from findings established through peer review.

Overall Conclusion
The central message is that AI is becoming increasingly capable of discovering mathematical results, but meaningful mathematical research involves more than producing answers. The greatest opportunities may arise from collaboration between AI systems and human mathematicians, ensuring that new results are not only correct but also understandable, significant and useful for future research.

ARTICLE]]> <![CDATA[Top mathematicians grade AI’s ability to solve toughest problems]]> https://mklab.gr/showthread.php?tid=2335 Thu, 22 Oct 2026 23:01:29 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2335 Source: Harvard Faculty of Arts and Sciences · Published October 19, 2026 · Author: Kermit Pattison

1 — The First Proof project and AI’s mathematical progress
The article examines the rapid development of artificial intelligence (AI) in advanced mathematical research, focusing on the First Proof project, an initiative involving leading mathematicians from universities such as Harvard, Stanford, and the University of Texas at Austin. The project aims to evaluate AI systems objectively by challenging them with difficult research problems that human mathematicians had solved but had not yet published. In its third round of testing, AI systems collectively solved all 10 mathematical problems, demonstrating substantial progress in their ability to produce sophisticated mathematical proofs. The problems covered diverse fields, including topology, statistics, logarithmic geometry, and symplectic geometry. The results suggest that AI is becoming increasingly capable of contributing to research-level mathematics rather than merely solving routine calculations or familiar textbook exercises.


2 — Impressive results but persistent limitations
Despite these achievements, the experiment revealed significant weaknesses in the quality of AI-generated mathematical solutions. The best-performing system, ProofCouncil, which combines several advanced AI models, solved nine out of ten problems, while OpenAI’s GPT-6 Astra solved seven. However, the systems often struggled to explain their reasoning clearly, organize proofs effectively, use mathematical notation appropriately, and provide accurate citations to existing research. Some solutions were unnecessarily complicated or difficult for human mathematicians to verify. In certain cases, the systems failed to acknowledge relevant published work, raising concerns about plagiarism and proper attribution. Researchers also identified errors arising from outdated versions of mathematical papers. These findings demonstrate that producing a potentially correct answer is not sufficient: mathematical research requires rigorous verification, transparent reasoning, clear exposition, and appropriate recognition of previous contributions.


3 — Implications for mathematical research and the future of AI
The findings highlight both the opportunities and challenges associated with integrating AI into mathematical research. AI systems could help researchers solve difficult problems more efficiently, explore alternative proofs, and accelerate certain stages of mathematical discovery. Nevertheless, human mathematicians remain essential for checking correctness, assessing originality, identifying errors, and ensuring that results are communicated reliably. Another important concern is the performance gap between advanced commercial AI systems and open-weight models, whose underlying model parameters are publicly available. In the latest experiment, every submission from the tested open-weight model contained major errors, raising questions about access to powerful mathematical AI tools and the concentration of technological capabilities among a small number of companies. The First Proof organizers plan further evaluations to understand these limitations better. Overall, the article argues for a balanced assessment of AI: its mathematical capabilities are advancing rapidly, but reliability, transparency, and human oversight remain crucial.

Key takeaways
  • Significant progress: AI systems collectively solved all 10 problems in the third round of the First Proof evaluation.
  • Best-performing system: ProofCouncil solved 9 out of 10 problems, while GPT-6 Astra solved 7.
  • Research-level mathematics: AI can increasingly tackle challenging problems across several advanced mathematical disciplines.
  • Correctness is not enough: A proof must also be logically rigorous, clearly written, and independently verifiable.
  • Weak exposition: AI-generated proofs can be excessively long, poorly organized, or difficult to understand.
  • Citation and originality concerns: AI systems may fail to acknowledge existing mathematical research adequately.
  • Human oversight remains essential: Expert mathematicians are needed to check proofs, detect errors, and evaluate originality.
  • Unequal access to AI capabilities: The experiment revealed a substantial performance gap between leading commercial systems and the tested open-weight model.
  • Future research: Continued independent testing is necessary to establish the strengths and limitations of mathematical AI.

Overall conclusion
The article presents AI as an increasingly powerful tool for advanced mathematical problem-solving, while emphasizing that its achievements should not be confused with fully reliable or independent mathematical research. The next major challenge is not simply enabling AI to solve more problems, but ensuring that its solutions are correct, understandable, reproducible, and properly documented.

Original article: Top mathematicians grade AI’s ability to solve toughest problems — Harvard FAS]]> Source: Harvard Faculty of Arts and Sciences · Published October 19, 2026 · Author: Kermit Pattison

1 — The First Proof project and AI’s mathematical progress
The article examines the rapid development of artificial intelligence (AI) in advanced mathematical research, focusing on the First Proof project, an initiative involving leading mathematicians from universities such as Harvard, Stanford, and the University of Texas at Austin. The project aims to evaluate AI systems objectively by challenging them with difficult research problems that human mathematicians had solved but had not yet published. In its third round of testing, AI systems collectively solved all 10 mathematical problems, demonstrating substantial progress in their ability to produce sophisticated mathematical proofs. The problems covered diverse fields, including topology, statistics, logarithmic geometry, and symplectic geometry. The results suggest that AI is becoming increasingly capable of contributing to research-level mathematics rather than merely solving routine calculations or familiar textbook exercises.


2 — Impressive results but persistent limitations
Despite these achievements, the experiment revealed significant weaknesses in the quality of AI-generated mathematical solutions. The best-performing system, ProofCouncil, which combines several advanced AI models, solved nine out of ten problems, while OpenAI’s GPT-6 Astra solved seven. However, the systems often struggled to explain their reasoning clearly, organize proofs effectively, use mathematical notation appropriately, and provide accurate citations to existing research. Some solutions were unnecessarily complicated or difficult for human mathematicians to verify. In certain cases, the systems failed to acknowledge relevant published work, raising concerns about plagiarism and proper attribution. Researchers also identified errors arising from outdated versions of mathematical papers. These findings demonstrate that producing a potentially correct answer is not sufficient: mathematical research requires rigorous verification, transparent reasoning, clear exposition, and appropriate recognition of previous contributions.


3 — Implications for mathematical research and the future of AI
The findings highlight both the opportunities and challenges associated with integrating AI into mathematical research. AI systems could help researchers solve difficult problems more efficiently, explore alternative proofs, and accelerate certain stages of mathematical discovery. Nevertheless, human mathematicians remain essential for checking correctness, assessing originality, identifying errors, and ensuring that results are communicated reliably. Another important concern is the performance gap between advanced commercial AI systems and open-weight models, whose underlying model parameters are publicly available. In the latest experiment, every submission from the tested open-weight model contained major errors, raising questions about access to powerful mathematical AI tools and the concentration of technological capabilities among a small number of companies. The First Proof organizers plan further evaluations to understand these limitations better. Overall, the article argues for a balanced assessment of AI: its mathematical capabilities are advancing rapidly, but reliability, transparency, and human oversight remain crucial.

Key takeaways
  • Significant progress: AI systems collectively solved all 10 problems in the third round of the First Proof evaluation.
  • Best-performing system: ProofCouncil solved 9 out of 10 problems, while GPT-6 Astra solved 7.
  • Research-level mathematics: AI can increasingly tackle challenging problems across several advanced mathematical disciplines.
  • Correctness is not enough: A proof must also be logically rigorous, clearly written, and independently verifiable.
  • Weak exposition: AI-generated proofs can be excessively long, poorly organized, or difficult to understand.
  • Citation and originality concerns: AI systems may fail to acknowledge existing mathematical research adequately.
  • Human oversight remains essential: Expert mathematicians are needed to check proofs, detect errors, and evaluate originality.
  • Unequal access to AI capabilities: The experiment revealed a substantial performance gap between leading commercial systems and the tested open-weight model.
  • Future research: Continued independent testing is necessary to establish the strengths and limitations of mathematical AI.

Overall conclusion
The article presents AI as an increasingly powerful tool for advanced mathematical problem-solving, while emphasizing that its achievements should not be confused with fully reliable or independent mathematical research. The next major challenge is not simply enabling AI to solve more problems, but ensuring that its solutions are correct, understandable, reproducible, and properly documented.

Original article: Top mathematicians grade AI’s ability to solve toughest problems — Harvard FAS]]> <![CDATA[AI Proved the Quasi-Riemann Hypothesis.]]> https://mklab.gr/showthread.php?tid=2330 Thu, 22 Oct 2026 15:32:14 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2330
The article, by Konstantin Kakaes in Quanta Magazine (October 19, 2026), reports on a claimed AI-generated proof of the quasi-Riemann hypothesis, released by OpenAI as one of hundreds of results announced on October 6, 2026. The Riemann hypothesis concerns the zeta function $\zeta(s) = 1 + \frac{1}{2^s} + \frac{1}{3^s} + \cdots$, extended to complex inputs $s$. It asserts that every nontrivial zero lies on the line $\operatorname{Re}(s) = \tfrac{1}{2}$. The link to prime numbers goes back to Euler's product formula $\zeta(s) = \prod_{p \text{ prime}} \frac{1}{1 - p^{-s}}$, and the location of the zeros governs how the primes thin out. The quasi-Riemann hypothesis is a weaker statement. It only requires all nontrivial zeros to lie in a strip narrower than the classically known one, $0 < \operatorname{Re}(s) < 1$. Even this weaker version has many of the consequences of the full conjecture, which one expert called "staggering".

OpenAI released two versions of the proof. The first, dated September 30, is 199 pages long and puts the zeros in $\tfrac{1}{8} < \operatorname{Re}(s) < \tfrac{7}{8}$. It has been formally verified in Lean. The second, about a week later, is only 49 pages and proves the weaker bound $\tfrac{1}{12} < \operatorname{Re}(s) < \tfrac{11}{12}$. It is not yet formally verified, but it is the only one of OpenAI's 722 papers from October 6 to have been edited by a human. Mathematicians say both papers are poorly written by journal standards, so they have focused on the shorter one. Maksym Radziwiłł says he can sketch its argument in half a page, and write-ups by other experts run 7 to 13 pages. Peter Sarnak has repurposed a course at Princeton to study the paper. The experts largely agree the result is correct, although Sarnak will not call it a theorem until humans have checked every detail.

The proof's novelty is a synthesis of existing ideas rather than new concepts. It follows Sarnak's long-standing view that the zeta function should be studied as one member of a family of functions. The AI found a family that earlier work, including Henryk Iwaniec's 2002 attempt, had narrowly missed. It also used metaplectic forms, which Alexander Dunn has studied for five years. A formula Dunn and Radziwiłł proved in 2021 turned out to be a crucial ingredient. Number theorists had believed their tools were too weak for this problem, so the result overturns that expectation. Experts are skeptical that these techniques will push the bound down to $\tfrac{1}{2}$ and prove the full hypothesis. Will Sawin expects the full conjecture to feel as distant as before. Thomas Bloom notes that the same would have been said about quasi-RH weeks earlier.

Key takeaways
  • Claim: OpenAI's model proved the quasi-Riemann hypothesis, which confines the nontrivial zeros of $\zeta(s)$ to a strip narrower than $0 < \operatorname{Re}(s) < 1$.
  • Two versions: a 199-page proof with bound $\tfrac{1}{8}$ to $\tfrac{7}{8}$ (Lean-verified) and a 49-page proof with bound $\tfrac{1}{12}$ to $\tfrac{11}{12}$ (not yet formally verified, human-edited).
  • Expert verdict: specialists largely believe the proof is correct. Some, like Sarnak, want full human verification first.
  • Why it matters: most consequences of the full Riemann hypothesis already follow from quasi-RH, with results across number theory, geometry and theoretical computer science.
  • Method: the proof is a surprising synthesis of known ideas. It uses a previously overlooked family of functions and metaplectic forms.
  • Human work reused: a 2021 formula by Dunn and Radziwiłł became a key ingredient.
  • Quality caveat: the papers are poorly written, but mathematicians have already condensed the argument into short write-ups.
  • Outlook: the full Riemann hypothesis, which needs the bound $\tfrac{1}{2}$, is still seen as out of reach, though Bloom says nothing would surprise him now.
  • Context: the first 12 trillion nontrivial zeros have been checked numerically on the line $\operatorname{Re}(s) = \tfrac{1}{2}$.

Sources: Quanta Magazine, "AI Proved the Quasi-Riemann Hypothesis. What Happens Now?"]]>
The article, by Konstantin Kakaes in Quanta Magazine (October 19, 2026), reports on a claimed AI-generated proof of the quasi-Riemann hypothesis, released by OpenAI as one of hundreds of results announced on October 6, 2026. The Riemann hypothesis concerns the zeta function $\zeta(s) = 1 + \frac{1}{2^s} + \frac{1}{3^s} + \cdots$, extended to complex inputs $s$. It asserts that every nontrivial zero lies on the line $\operatorname{Re}(s) = \tfrac{1}{2}$. The link to prime numbers goes back to Euler's product formula $\zeta(s) = \prod_{p \text{ prime}} \frac{1}{1 - p^{-s}}$, and the location of the zeros governs how the primes thin out. The quasi-Riemann hypothesis is a weaker statement. It only requires all nontrivial zeros to lie in a strip narrower than the classically known one, $0 < \operatorname{Re}(s) < 1$. Even this weaker version has many of the consequences of the full conjecture, which one expert called "staggering".

OpenAI released two versions of the proof. The first, dated September 30, is 199 pages long and puts the zeros in $\tfrac{1}{8} < \operatorname{Re}(s) < \tfrac{7}{8}$. It has been formally verified in Lean. The second, about a week later, is only 49 pages and proves the weaker bound $\tfrac{1}{12} < \operatorname{Re}(s) < \tfrac{11}{12}$. It is not yet formally verified, but it is the only one of OpenAI's 722 papers from October 6 to have been edited by a human. Mathematicians say both papers are poorly written by journal standards, so they have focused on the shorter one. Maksym Radziwiłł says he can sketch its argument in half a page, and write-ups by other experts run 7 to 13 pages. Peter Sarnak has repurposed a course at Princeton to study the paper. The experts largely agree the result is correct, although Sarnak will not call it a theorem until humans have checked every detail.

The proof's novelty is a synthesis of existing ideas rather than new concepts. It follows Sarnak's long-standing view that the zeta function should be studied as one member of a family of functions. The AI found a family that earlier work, including Henryk Iwaniec's 2002 attempt, had narrowly missed. It also used metaplectic forms, which Alexander Dunn has studied for five years. A formula Dunn and Radziwiłł proved in 2021 turned out to be a crucial ingredient. Number theorists had believed their tools were too weak for this problem, so the result overturns that expectation. Experts are skeptical that these techniques will push the bound down to $\tfrac{1}{2}$ and prove the full hypothesis. Will Sawin expects the full conjecture to feel as distant as before. Thomas Bloom notes that the same would have been said about quasi-RH weeks earlier.

Key takeaways
  • Claim: OpenAI's model proved the quasi-Riemann hypothesis, which confines the nontrivial zeros of $\zeta(s)$ to a strip narrower than $0 < \operatorname{Re}(s) < 1$.
  • Two versions: a 199-page proof with bound $\tfrac{1}{8}$ to $\tfrac{7}{8}$ (Lean-verified) and a 49-page proof with bound $\tfrac{1}{12}$ to $\tfrac{11}{12}$ (not yet formally verified, human-edited).
  • Expert verdict: specialists largely believe the proof is correct. Some, like Sarnak, want full human verification first.
  • Why it matters: most consequences of the full Riemann hypothesis already follow from quasi-RH, with results across number theory, geometry and theoretical computer science.
  • Method: the proof is a surprising synthesis of known ideas. It uses a previously overlooked family of functions and metaplectic forms.
  • Human work reused: a 2021 formula by Dunn and Radziwiłł became a key ingredient.
  • Quality caveat: the papers are poorly written, but mathematicians have already condensed the argument into short write-ups.
  • Outlook: the full Riemann hypothesis, which needs the bound $\tfrac{1}{2}$, is still seen as out of reach, though Bloom says nothing would surprise him now.
  • Context: the first 12 trillion nontrivial zeros have been checked numerically on the line $\operatorname{Re}(s) = \tfrac{1}{2}$.

Sources: Quanta Magazine, "AI Proved the Quasi-Riemann Hypothesis. What Happens Now?"]]> <![CDATA[The end of mathematicians?]]> https://mklab.gr/showthread.php?tid=2328 Sun, 18 Oct 2026 18:05:43 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2328
1 — The rapid rise of AI in mathematics. 
In his October 3, 2026, article, The End of Mathematicians?, Rowland Manthorpe explores how advances in artificial intelligence are challenging the future of mathematical research. Through an interview with Tim Gowers, a Fields Medal-winning mathematician at Cambridge and the Collège de France, the article examines the extraordinary speed at which AI capabilities have progressed. According to Gowers, AI has advanced from solving mathematical Olympiad problems to tackling research-level problems and challenges traditionally associated with leading mathematicians. This rapid development has forced mathematicians to reconsider their assumptions about human intellectual superiority and the future of their profession.

2 — The threat to mathematical research and human motivation. 
The central concern is not simply that AI might solve problems faster than humans, but that it could fundamentally change what makes mathematics meaningful. Gowers explains that the pleasure of mathematical research often comes from struggling with a difficult problem, developing new ideas, and finally discovering a solution after months or years of effort. If AI increasingly performs these tasks, mathematicians could lose an important source of intellectual satisfaction and professional identity. There is also a potential long-term consequence: fewer students may pursue mathematics PhDs if they believe machines will outperform them. This could create a paradox in which AI generates an unprecedented wealth of mathematical discoveries, but the human expertise needed to understand, evaluate, and use those discoveries becomes harder to sustain.


3 — The changing role of mathematicians. 
Despite these concerns, Gowers does not conclude that mathematicians will necessarily become obsolete. Instead, he suggests that their role may shift from discovering mathematical results to understanding, interpreting, connecting, and applying them. Mathematicians could become more like scholars who explain complex theories, assess the significance of AI-generated proofs, and help other disciplines use mathematical knowledge effectively. This transition, however, raises questions about professional status, university funding, doctoral education, and the motivation of future researchers. The article ultimately presents an uncertain future: AI could unlock mathematical discoveries that humans might otherwise never achieve, but society must find ways to preserve human understanding and sustain the mathematical community in an era of increasingly capable machines.

Key takeaways
1. AI is advancing at an extraordinary pace. Its capabilities are moving beyond competition-level mathematics toward major research problems.
2. Mathematical research is being redefined. Tasks once considered uniquely human, such as solving difficult research problems, may increasingly be automated.
3. The process matters as much as the answer. For mathematicians, the intellectual struggle and satisfaction of discovery are central to the appeal of the subject.
4. The future of mathematics education is uncertain. If AI can outperform researchers, students may question whether years of advanced mathematical training are worthwhile.
5. Mathematicians may become interpreters and curators of knowledge. Understanding proofs, connecting theories, evaluating results, and identifying applications could become increasingly important.
6. AI could accelerate scientific progress. Physicists, engineers, economists, and other professionals may benefit from access to a much larger body of mathematical results.
7. Human expertise will remain valuable if it evolves. Society needs people who can assess the significance and reliability of mathematical knowledge, rather than merely produce solutions.
8. The institutional consequences remain unclear. Universities, research departments, and doctoral programmes may need to rethink their purpose and structure.
9. The article does not claim that mathematicians are already obsolete. It presents a serious possibility and a set of concerns, not a settled prediction.

ARTICLE]]>
1 — The rapid rise of AI in mathematics. 
In his October 3, 2026, article, The End of Mathematicians?, Rowland Manthorpe explores how advances in artificial intelligence are challenging the future of mathematical research. Through an interview with Tim Gowers, a Fields Medal-winning mathematician at Cambridge and the Collège de France, the article examines the extraordinary speed at which AI capabilities have progressed. According to Gowers, AI has advanced from solving mathematical Olympiad problems to tackling research-level problems and challenges traditionally associated with leading mathematicians. This rapid development has forced mathematicians to reconsider their assumptions about human intellectual superiority and the future of their profession.

2 — The threat to mathematical research and human motivation. 
The central concern is not simply that AI might solve problems faster than humans, but that it could fundamentally change what makes mathematics meaningful. Gowers explains that the pleasure of mathematical research often comes from struggling with a difficult problem, developing new ideas, and finally discovering a solution after months or years of effort. If AI increasingly performs these tasks, mathematicians could lose an important source of intellectual satisfaction and professional identity. There is also a potential long-term consequence: fewer students may pursue mathematics PhDs if they believe machines will outperform them. This could create a paradox in which AI generates an unprecedented wealth of mathematical discoveries, but the human expertise needed to understand, evaluate, and use those discoveries becomes harder to sustain.


3 — The changing role of mathematicians. 
Despite these concerns, Gowers does not conclude that mathematicians will necessarily become obsolete. Instead, he suggests that their role may shift from discovering mathematical results to understanding, interpreting, connecting, and applying them. Mathematicians could become more like scholars who explain complex theories, assess the significance of AI-generated proofs, and help other disciplines use mathematical knowledge effectively. This transition, however, raises questions about professional status, university funding, doctoral education, and the motivation of future researchers. The article ultimately presents an uncertain future: AI could unlock mathematical discoveries that humans might otherwise never achieve, but society must find ways to preserve human understanding and sustain the mathematical community in an era of increasingly capable machines.

Key takeaways
1. AI is advancing at an extraordinary pace. Its capabilities are moving beyond competition-level mathematics toward major research problems.
2. Mathematical research is being redefined. Tasks once considered uniquely human, such as solving difficult research problems, may increasingly be automated.
3. The process matters as much as the answer. For mathematicians, the intellectual struggle and satisfaction of discovery are central to the appeal of the subject.
4. The future of mathematics education is uncertain. If AI can outperform researchers, students may question whether years of advanced mathematical training are worthwhile.
5. Mathematicians may become interpreters and curators of knowledge. Understanding proofs, connecting theories, evaluating results, and identifying applications could become increasingly important.
6. AI could accelerate scientific progress. Physicists, engineers, economists, and other professionals may benefit from access to a much larger body of mathematical results.
7. Human expertise will remain valuable if it evolves. Society needs people who can assess the significance and reliability of mathematical knowledge, rather than merely produce solutions.
8. The institutional consequences remain unclear. Universities, research departments, and doctoral programmes may need to rethink their purpose and structure.
9. The article does not claim that mathematicians are already obsolete. It presents a serious possibility and a set of concerns, not a settled prediction.

ARTICLE]]> <![CDATA[Responsible Release of AI-Generated Mathematics]]> https://mklab.gr/showthread.php?tid=2324 Fri, 02 Oct 2026 03:08:46 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2324 “Responsible Release of AI-Generated Mathematics”, published on September 29, 2026, argues that the rapid emergence of AI systems capable of producing advanced mathematical results requires the mathematical community to establish new norms for responsible dissemination. The authors emphasize that mathematics traditionally depends on human understanding, verification, accountability, peer review, and communication. They are particularly concerned that AI can now generate mathematical arguments that the humans prompting the systems may not themselves understand or be able to verify. The article is informed by more than 600 responses from the mathematical community and proposes principles intended for AI laboratories whose systems could significantly affect mathematical research.

The central distinction is between mathematical papers that are fully understood by a human mathematician and those whose AI-generated content is not yet understood. In the first case, the authors recommend following traditional academic norms: producing a preprint, submitting the work for peer review, and presenting it through seminars and conferences. For results that remain unexplained to their human users, the article proposes an initial release process requiring AI labs to improve the exposition, search the literature and provide proper citations, deposit results in independent scholarly repositories, and avoid using mathematical discoveries primarily as marketing tools. Labs should also disclose information about the model, prompts, summarized reasoning process, computation time, computational cost, and the circumstances under which AI was used. 

The article also recommends that AI-generated proofs be formalized whenever possible, using standards familiar to the mathematical community, and that releases explain not only successful solutions but also the broader experimental context. For example, if a laboratory presents many successful results, it should disclose how many comparable problems the models attempted and failed to solve and how the problems were selected. Crucially, the authors argue that AI laboratories have a responsibility to support the subsequent development of human understanding. Depending on the complexity and significance of the results, this could involve funding conferences, summer schools, workshops, working groups, postdoctoral or student positions, books, or extensive expository writing. Decisions about distributing such support, however, should be made by independent nonprofit institutions, rather than by the AI laboratories themselves. 

Finally, the article addresses access to AI systems. The authors warn that proprietary frontier models could create a two-tier mathematical ecosystem in which AI laboratories advance beyond the wider research community, while unequal access to publicly available models could further amplify existing inequalities based on wealth and technological resources. They therefore advocate broad and equitable access to publicly available AI models, arguing that mathematical progress depends heavily on collective verification, shared intuition, and conceptual synthesis. Overall, the article presents responsible AI-generated mathematics as requiring not merely the publication of AI-produced results, but mechanisms that preserve transparency, accountability, reproducibility, human comprehension, community participation, and equitable access. 

Key Takeaways
  1. Human understanding remains central to responsible mathematical research, even when AI produces the underlying results.
  2. AI laboratories should release significant mathematical results responsibly and promptly, rather than keeping important discoveries proprietary.
  3. AI-generated papers should follow appropriate academic and scholarly standards, including citations, clear exposition, repositories, and peer review where applicable.
  4. Labs should disclose the model, prompts, summarized reasoning, computation time, computational cost, and AI's role in producing each result.
  5. Formal verification and formalization of AI-generated proofs should be pursued whenever reasonably possible.
  6. AI laboratories should provide financial and institutional support for human understanding of mathematical results that humans cannot initially comprehend.
  7. Such support could include conferences, workshops, working groups, students, postdocs, books, and expository articles.
  8. The mathematical community should not become divided between researchers with access to proprietary frontier models and those without it.
  9. The article advocates broad and equitable access to publicly available mathematical AI systems.
  10. The overarching goal is to ensure that AI accelerates mathematical discovery while preserving verification, transparency, accountability, shared understanding, and collective mathematical progress.

ARTICLE]]> “Responsible Release of AI-Generated Mathematics”, published on September 29, 2026, argues that the rapid emergence of AI systems capable of producing advanced mathematical results requires the mathematical community to establish new norms for responsible dissemination. The authors emphasize that mathematics traditionally depends on human understanding, verification, accountability, peer review, and communication. They are particularly concerned that AI can now generate mathematical arguments that the humans prompting the systems may not themselves understand or be able to verify. The article is informed by more than 600 responses from the mathematical community and proposes principles intended for AI laboratories whose systems could significantly affect mathematical research.

The central distinction is between mathematical papers that are fully understood by a human mathematician and those whose AI-generated content is not yet understood. In the first case, the authors recommend following traditional academic norms: producing a preprint, submitting the work for peer review, and presenting it through seminars and conferences. For results that remain unexplained to their human users, the article proposes an initial release process requiring AI labs to improve the exposition, search the literature and provide proper citations, deposit results in independent scholarly repositories, and avoid using mathematical discoveries primarily as marketing tools. Labs should also disclose information about the model, prompts, summarized reasoning process, computation time, computational cost, and the circumstances under which AI was used. 

The article also recommends that AI-generated proofs be formalized whenever possible, using standards familiar to the mathematical community, and that releases explain not only successful solutions but also the broader experimental context. For example, if a laboratory presents many successful results, it should disclose how many comparable problems the models attempted and failed to solve and how the problems were selected. Crucially, the authors argue that AI laboratories have a responsibility to support the subsequent development of human understanding. Depending on the complexity and significance of the results, this could involve funding conferences, summer schools, workshops, working groups, postdoctoral or student positions, books, or extensive expository writing. Decisions about distributing such support, however, should be made by independent nonprofit institutions, rather than by the AI laboratories themselves. 

Finally, the article addresses access to AI systems. The authors warn that proprietary frontier models could create a two-tier mathematical ecosystem in which AI laboratories advance beyond the wider research community, while unequal access to publicly available models could further amplify existing inequalities based on wealth and technological resources. They therefore advocate broad and equitable access to publicly available AI models, arguing that mathematical progress depends heavily on collective verification, shared intuition, and conceptual synthesis. Overall, the article presents responsible AI-generated mathematics as requiring not merely the publication of AI-produced results, but mechanisms that preserve transparency, accountability, reproducibility, human comprehension, community participation, and equitable access. 

Key Takeaways
  1. Human understanding remains central to responsible mathematical research, even when AI produces the underlying results.
  2. AI laboratories should release significant mathematical results responsibly and promptly, rather than keeping important discoveries proprietary.
  3. AI-generated papers should follow appropriate academic and scholarly standards, including citations, clear exposition, repositories, and peer review where applicable.
  4. Labs should disclose the model, prompts, summarized reasoning, computation time, computational cost, and AI's role in producing each result.
  5. Formal verification and formalization of AI-generated proofs should be pursued whenever reasonably possible.
  6. AI laboratories should provide financial and institutional support for human understanding of mathematical results that humans cannot initially comprehend.
  7. Such support could include conferences, workshops, working groups, students, postdocs, books, and expository articles.
  8. The mathematical community should not become divided between researchers with access to proprietary frontier models and those without it.
  9. The article advocates broad and equitable access to publicly available mathematical AI systems.
  10. The overarching goal is to ensure that AI accelerates mathematical discovery while preserving verification, transparency, accountability, shared understanding, and collective mathematical progress.

ARTICLE]]> <![CDATA[COMPUTING MACHINERY AND INTELLIGENCE [A. M. TURING]]]> https://mklab.gr/showthread.php?tid=2321 Thu, 01 Oct 2026 22:25:22 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2321 Alan Turing’s “Computing Machinery and Intelligence” (1950) is a landmark paper that examines the question of whether machines can think. Instead of trying to define precisely what “thinking” means, Turing proposes the Imitation Game, which later became known as the Turing Test. In the original game, a human interrogator communicates with a man and a woman and attempts to identify them. Turing replaces one participant with a machine and asks whether the interrogator can distinguish the machine from the human through written communication. This approach transforms an abstract philosophical question into a practical test of intelligent behaviour.

Turing then considers a series of objections to the possibility of machine intelligence, including mathematical, theological, philosophical, and psychological arguments. One objection claims that machines simply follow instructions and therefore cannot genuinely think or be creative. Another argues that intelligence requires consciousness and subjective experience. Turing questions whether such concepts can be objectively established even in other human beings and suggests that observable behaviour provides a more useful basis for scientific investigation. He also discusses mathematical limitations associated with Gödel’s incompleteness results, arguing that such limitations do not demonstrate that machines are incapable of intelligent behaviour.

A major part of the paper concerns the possibility of creating learning machines. Turing suggests that instead of attempting to program every possible form of intelligent behaviour, researchers could construct machines that learn from experience, education, experimentation, and feedback. He compares this process with the development of a human child, proposing that a relatively simple machine could gradually acquire increasingly sophisticated abilities. He also considers the possibility of machines modifying their own instructions and developing behaviour that was not explicitly specified by their programmers. These ideas were remarkably forward-looking and anticipate important principles underlying modern machine learning and artificial intelligence.

Turing concludes that the question of whether machines can think should be approached through experimentation and technological development rather than by relying exclusively on philosophical definitions. He predicts that computers would eventually become capable of producing responses sufficiently similar to those of humans that distinguishing between them would become difficult. Although some of his predictions were tied to the technological circumstances of his time, the paper established many of the central questions of artificial intelligence, including machine intelligence, learning, consciousness, creativity, and the relationship between intelligent behaviour and thought. Its enduring significance lies in replacing the vague question “Can machines think?” with a concrete research problem concerning whether machines can exhibit behaviour indistinguishable from human intelligence.


ARTICLE]]> Alan Turing’s “Computing Machinery and Intelligence” (1950) is a landmark paper that examines the question of whether machines can think. Instead of trying to define precisely what “thinking” means, Turing proposes the Imitation Game, which later became known as the Turing Test. In the original game, a human interrogator communicates with a man and a woman and attempts to identify them. Turing replaces one participant with a machine and asks whether the interrogator can distinguish the machine from the human through written communication. This approach transforms an abstract philosophical question into a practical test of intelligent behaviour.

Turing then considers a series of objections to the possibility of machine intelligence, including mathematical, theological, philosophical, and psychological arguments. One objection claims that machines simply follow instructions and therefore cannot genuinely think or be creative. Another argues that intelligence requires consciousness and subjective experience. Turing questions whether such concepts can be objectively established even in other human beings and suggests that observable behaviour provides a more useful basis for scientific investigation. He also discusses mathematical limitations associated with Gödel’s incompleteness results, arguing that such limitations do not demonstrate that machines are incapable of intelligent behaviour.

A major part of the paper concerns the possibility of creating learning machines. Turing suggests that instead of attempting to program every possible form of intelligent behaviour, researchers could construct machines that learn from experience, education, experimentation, and feedback. He compares this process with the development of a human child, proposing that a relatively simple machine could gradually acquire increasingly sophisticated abilities. He also considers the possibility of machines modifying their own instructions and developing behaviour that was not explicitly specified by their programmers. These ideas were remarkably forward-looking and anticipate important principles underlying modern machine learning and artificial intelligence.

Turing concludes that the question of whether machines can think should be approached through experimentation and technological development rather than by relying exclusively on philosophical definitions. He predicts that computers would eventually become capable of producing responses sufficiently similar to those of humans that distinguishing between them would become difficult. Although some of his predictions were tied to the technological circumstances of his time, the paper established many of the central questions of artificial intelligence, including machine intelligence, learning, consciousness, creativity, and the relationship between intelligent behaviour and thought. Its enduring significance lies in replacing the vague question “Can machines think?” with a concrete research problem concerning whether machines can exhibit behaviour indistinguishable from human intelligence.


ARTICLE]]> <![CDATA[The Work of Math in the Age of Artificial Reproduction]]> https://mklab.gr/showthread.php?tid=2315 Thu, 01 Oct 2026 19:31:45 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2315 Navier–Stokes Millennium Prize Problem, one of the most famous unresolved questions in mathematics. The problem concerns whether the Navier–Stokes equations, which describe the motion of fluids, can develop a finite-time singularity or “blow-up” from initially smooth conditions. The equations are extraordinarily successful in practical applications such as weather modeling and aircraft design, yet mathematicians had never established whether their solutions always remain smooth. Rockmore explains that mathematicians Tristan Buckmaster and Levent Alpöge had independently developed important work on related Euler equations, building on earlier results by Diego Córdoba and Luis Martínez-Zoroa.

The article then describes how AI became involved in the mathematical breakthrough. OpenAI reportedly deployed its latest model against several Millennium Problems, first using roughly one hundred AI agents on the Euler equations and subsequently scaling the effort to approximately 10,000 agents for Navier–Stokes. Over about eighty-eight hours, the agents exchanged around 2.7 million messages, producing approximately 130 billion tokens of mathematical work. Human mathematicians supervised the process and used additional models to consolidate promising approaches. Although the resulting proof still requires thorough human verification, Rockmore reports that initial assessments suggested the problem had been resolved. The episode therefore represents not simply AI assisting a mathematician with calculations, but a fundamentally different model of mathematical research involving large-scale machine-generated exploration.

However, Rockmore is interested in much more than whether the proof is correct. He highlights the controversy over priority and credit, particularly because Buckmaster and Alpöge had already produced important results and published their work shortly before OpenAI announced its achievement. The episode raises broader questions about how mathematical credit should work when discoveries emerge from a combination of human researchers, previous literature, AI systems, and enormous computational resources. Rockmore argues that mathematics is particularly valuable to AI companies as a demonstration of apparent machine intelligence: unlike many real-world problems, a mathematical proof can ultimately be checked for correctness. Consequently, solving a celebrated mathematical problem becomes a powerful public demonstration of what AI systems can supposedly accomplish, as well as a way of showcasing enormous computational power.

Finally, Rockmore connects the episode to Walter Benjamin’s famous essay “The Work of Art in the Age of Mechanical Reproduction.” Benjamin argued that technologies such as photography and film changed the status of artistic works by separating them from their unique origins and “aura.” Rockmore asks whether AI is doing something analogous to human thought and mathematical creativity: reproducing the products of thinking while potentially disconnecting those products from the individual thinkers and historical processes that produced them. He acknowledges the enormous potential of AI to accelerate mathematical research and even suggests that traditional educational and research structures may eventually change dramatically. At the same time, he worries that dependence on common AI systems could make human thought increasingly standardized and concentrated around technologies controlled by a small number of companies. His central question is therefore not simply “Can AI solve mathematics?”, but whether mathematics—and human thinking more generally—loses something important when the process of discovery becomes increasingly automated.

ARTICLE]]> Navier–Stokes Millennium Prize Problem, one of the most famous unresolved questions in mathematics. The problem concerns whether the Navier–Stokes equations, which describe the motion of fluids, can develop a finite-time singularity or “blow-up” from initially smooth conditions. The equations are extraordinarily successful in practical applications such as weather modeling and aircraft design, yet mathematicians had never established whether their solutions always remain smooth. Rockmore explains that mathematicians Tristan Buckmaster and Levent Alpöge had independently developed important work on related Euler equations, building on earlier results by Diego Córdoba and Luis Martínez-Zoroa.

The article then describes how AI became involved in the mathematical breakthrough. OpenAI reportedly deployed its latest model against several Millennium Problems, first using roughly one hundred AI agents on the Euler equations and subsequently scaling the effort to approximately 10,000 agents for Navier–Stokes. Over about eighty-eight hours, the agents exchanged around 2.7 million messages, producing approximately 130 billion tokens of mathematical work. Human mathematicians supervised the process and used additional models to consolidate promising approaches. Although the resulting proof still requires thorough human verification, Rockmore reports that initial assessments suggested the problem had been resolved. The episode therefore represents not simply AI assisting a mathematician with calculations, but a fundamentally different model of mathematical research involving large-scale machine-generated exploration.

However, Rockmore is interested in much more than whether the proof is correct. He highlights the controversy over priority and credit, particularly because Buckmaster and Alpöge had already produced important results and published their work shortly before OpenAI announced its achievement. The episode raises broader questions about how mathematical credit should work when discoveries emerge from a combination of human researchers, previous literature, AI systems, and enormous computational resources. Rockmore argues that mathematics is particularly valuable to AI companies as a demonstration of apparent machine intelligence: unlike many real-world problems, a mathematical proof can ultimately be checked for correctness. Consequently, solving a celebrated mathematical problem becomes a powerful public demonstration of what AI systems can supposedly accomplish, as well as a way of showcasing enormous computational power.

Finally, Rockmore connects the episode to Walter Benjamin’s famous essay “The Work of Art in the Age of Mechanical Reproduction.” Benjamin argued that technologies such as photography and film changed the status of artistic works by separating them from their unique origins and “aura.” Rockmore asks whether AI is doing something analogous to human thought and mathematical creativity: reproducing the products of thinking while potentially disconnecting those products from the individual thinkers and historical processes that produced them. He acknowledges the enormous potential of AI to accelerate mathematical research and even suggests that traditional educational and research structures may eventually change dramatically. At the same time, he worries that dependence on common AI systems could make human thought increasingly standardized and concentrated around technologies controlled by a small number of companies. His central question is therefore not simply “Can AI solve mathematics?”, but whether mathematics—and human thinking more generally—loses something important when the process of discovery becomes increasingly automated.

ARTICLE]]> <![CDATA[Data Ethics and Mathematicians]]> https://mklab.gr/showthread.php?tid=2314 Thu, 01 Oct 2026 19:26:10 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2314 “Data Ethics and Mathematicians”, by Mason A. Porter, argues that data ethics should become an integral part of mathematics, statistics, and computational science. Mathematical work increasingly involves human, social, and other sensitive data, meaning that mathematical scientists can have significant effects on individuals and society. Porter emphasizes that mathematics and mathematics education do not exist in a vacuum: choices about what researchers study, what data they use, and how they apply mathematical methods can have ethical consequences. He therefore argues that students and researchers should be trained to approach these issues with reflection, humility, and ethical awareness, learning from disciplines such as medicine and the social sciences that have longer traditions of research ethics.

A major issue discussed is the tension between open science, reproducibility, and privacy. Researchers should, whenever possible, make their data, source code, methods, and data-processing procedures available so that others can reproduce and evaluate their results. However, this ideal can conflict with privacy regulations, nondisclosure agreements, licensing restrictions, or the potential harm caused by releasing sensitive information. The article stresses that apparently anonymous data may not actually be anonymous: combining different datasets can make it possible to re-identify individuals. Consequently, researchers must carefully consider anonymization, sampling, data cleaning, licensing, and methods such as differential privacy, rather than treating data publication as automatically desirable. 

Porter also examines the rapidly growing ethical challenges associated with artificial intelligence and large language models (LLMs). Generative AI raises questions about data ownership, copyright, training data, authorship, and responsible use in research and education. At the same time, the article emphasizes that ethical research requires more than simply complying with laws or institutional regulations. Researchers should consider principles such as respect for persons, beneficence, justice, and respect for law and the public interest, while also addressing informed consent, privacy, informational risk, and uncertainty. Ethical decision-making should therefore be viewed as a continuous process rather than as a simple requirement that is either satisfied or not satisfied. 

In conclusion, the article argues that the increasing use of digital data to study people and build technologies makes ethical responsibility an essential component of modern mathematical practice. Porter believes that mathematics and computational science should develop stronger ethics education, drawing on established practices in other disciplines and incorporating ethical questions into curricula. Mathematicians should think carefully about the consequences of their research, openly discuss difficult ethical choices, and recognize that technical correctness alone does not guarantee responsible research. 

The central message is that mathematical scientists who work with human data have a professional responsibility to consider not only whether something can be done mathematically or computationally, but also whether and how it ought to be done.

ARTICLE]]> “Data Ethics and Mathematicians”, by Mason A. Porter, argues that data ethics should become an integral part of mathematics, statistics, and computational science. Mathematical work increasingly involves human, social, and other sensitive data, meaning that mathematical scientists can have significant effects on individuals and society. Porter emphasizes that mathematics and mathematics education do not exist in a vacuum: choices about what researchers study, what data they use, and how they apply mathematical methods can have ethical consequences. He therefore argues that students and researchers should be trained to approach these issues with reflection, humility, and ethical awareness, learning from disciplines such as medicine and the social sciences that have longer traditions of research ethics.

A major issue discussed is the tension between open science, reproducibility, and privacy. Researchers should, whenever possible, make their data, source code, methods, and data-processing procedures available so that others can reproduce and evaluate their results. However, this ideal can conflict with privacy regulations, nondisclosure agreements, licensing restrictions, or the potential harm caused by releasing sensitive information. The article stresses that apparently anonymous data may not actually be anonymous: combining different datasets can make it possible to re-identify individuals. Consequently, researchers must carefully consider anonymization, sampling, data cleaning, licensing, and methods such as differential privacy, rather than treating data publication as automatically desirable. 

Porter also examines the rapidly growing ethical challenges associated with artificial intelligence and large language models (LLMs). Generative AI raises questions about data ownership, copyright, training data, authorship, and responsible use in research and education. At the same time, the article emphasizes that ethical research requires more than simply complying with laws or institutional regulations. Researchers should consider principles such as respect for persons, beneficence, justice, and respect for law and the public interest, while also addressing informed consent, privacy, informational risk, and uncertainty. Ethical decision-making should therefore be viewed as a continuous process rather than as a simple requirement that is either satisfied or not satisfied. 

In conclusion, the article argues that the increasing use of digital data to study people and build technologies makes ethical responsibility an essential component of modern mathematical practice. Porter believes that mathematics and computational science should develop stronger ethics education, drawing on established practices in other disciplines and incorporating ethical questions into curricula. Mathematicians should think carefully about the consequences of their research, openly discuss difficult ethical choices, and recognize that technical correctness alone does not guarantee responsible research. 

The central message is that mathematical scientists who work with human data have a professional responsibility to consider not only whether something can be done mathematically or computationally, but also whether and how it ought to be done.

ARTICLE]]> <![CDATA[How to spot AI writing]]> https://mklab.gr/showthread.php?tid=2307 Wed, 30 Sep 2026 23:51:47 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2307 The widespread use of artificial intelligence (AI) has transformed written communication, with large language models (LLMs) increasingly producing emails, essays, scientific papers, journalism and other forms of content. Although AI can imitate different literary styles and generate large quantities of text rapidly, its writing can exhibit distinctive linguistic patterns. Commonly suggested indicators include em dashes, sophisticated vocabulary and expressions such as “delve” and “rich tapestry.” However, the article stresses that there is no single linguistic feature that reliably proves that a text was written by AI. Identifying AI-generated prose therefore requires examining several characteristics together rather than focusing on individual words or punctuation marks.
To investigate these characteristics, The Economist compared human-written articles with texts generated by ChatGPT, Claude, Gemini and Grok. The models were asked to reproduce versions of the magazine’s articles without access to the internet, using AI-generated summaries as prompts. The resulting dataset contained 55,940 sentences and 1.2 million words. The researchers also compared the AI-generated material with journalism from CNN, The New York Times and The Washington Post, as well as bestselling fiction published between 1950 and 2022. The analysis found differences between AI and human writing in vocabulary, punctuation, sentence length and paragraph structure. At the same time, these characteristics can change considerably when AI models are updated.
One notable characteristic is vocabulary. Contrary to the popular belief that AI consistently overuses words such as “delve” and “tapestry,” contemporary models often favour longer and more sophisticated words, including “significant,” “increasingly” and “consequences,” as well as technical terminology such as “methodology” and “interdependence.” AI-generated texts also make frequent use of nominalisations, turning verbs into nouns, and words with Latin-derived suffixes. The article also examines punctuation and challenges the idea that AI universally overuses em dashes. ChatGPT's newer versions use fewer dashes, whereas Claude continues to use them relatively frequently. More persistent differences include the comparatively limited use of commas, semicolons and parentheses, alongside longer sentences and fewer quotations from experts.
The article identifies further patterns in sentence structure and rhetoric. AI-generated prose tends to contain longer sentences, less variation in sentence length and fewer short, emphatic statements. Models also frequently use rhetorical constructions such as “not X but Y,” “not only but also,” and the rule of three, in which three related ideas are grouped together for emphasis. ChatGPT and Claude use these constructions particularly often compared with human writing. AI-detection tools can help identify machine-generated text, but they are imperfect and can produce false positives. The article's central conclusion is that AI writing may display a combination of formulaic structures, sophisticated vocabulary and limited stylistic variation, but these characteristics are changing rapidly. As AI models become increasingly capable of producing human-like prose, reliably distinguishing AI-generated writing from human writing is becoming more difficult.

Key Takeaways
  1. AI writing is increasingly widespread, appearing in journalism, education, science and everyday communication.
  2. No single feature reliably identifies AI writing; several linguistic characteristics need to be considered together.
  3. Sophisticated vocabulary and technical terminology are common characteristics of AI-generated prose.
  4. Nominalisations and longer, more formal words can occur more frequently in AI-generated writing.
  5. Punctuation patterns vary between AI models, so the presence of em dashes alone is not a reliable indicator.
  6. Longer sentences and limited variation in sentence length are recurring characteristics of AI-generated prose.
  7. AI models frequently employ rhetorical patterns such as “not X but Y” and the rule of three.
  8. AI-detection software is imperfect and can produce false positives.
  9. AI writing changes as models are updated, meaning that detection techniques based on older models may become less effective.
  10. The increasing ability of AI systems to produce human-like prose makes reliable identification progressively more challenging.

ARTICLE]]> The widespread use of artificial intelligence (AI) has transformed written communication, with large language models (LLMs) increasingly producing emails, essays, scientific papers, journalism and other forms of content. Although AI can imitate different literary styles and generate large quantities of text rapidly, its writing can exhibit distinctive linguistic patterns. Commonly suggested indicators include em dashes, sophisticated vocabulary and expressions such as “delve” and “rich tapestry.” However, the article stresses that there is no single linguistic feature that reliably proves that a text was written by AI. Identifying AI-generated prose therefore requires examining several characteristics together rather than focusing on individual words or punctuation marks.
To investigate these characteristics, The Economist compared human-written articles with texts generated by ChatGPT, Claude, Gemini and Grok. The models were asked to reproduce versions of the magazine’s articles without access to the internet, using AI-generated summaries as prompts. The resulting dataset contained 55,940 sentences and 1.2 million words. The researchers also compared the AI-generated material with journalism from CNN, The New York Times and The Washington Post, as well as bestselling fiction published between 1950 and 2022. The analysis found differences between AI and human writing in vocabulary, punctuation, sentence length and paragraph structure. At the same time, these characteristics can change considerably when AI models are updated.
One notable characteristic is vocabulary. Contrary to the popular belief that AI consistently overuses words such as “delve” and “tapestry,” contemporary models often favour longer and more sophisticated words, including “significant,” “increasingly” and “consequences,” as well as technical terminology such as “methodology” and “interdependence.” AI-generated texts also make frequent use of nominalisations, turning verbs into nouns, and words with Latin-derived suffixes. The article also examines punctuation and challenges the idea that AI universally overuses em dashes. ChatGPT's newer versions use fewer dashes, whereas Claude continues to use them relatively frequently. More persistent differences include the comparatively limited use of commas, semicolons and parentheses, alongside longer sentences and fewer quotations from experts.
The article identifies further patterns in sentence structure and rhetoric. AI-generated prose tends to contain longer sentences, less variation in sentence length and fewer short, emphatic statements. Models also frequently use rhetorical constructions such as “not X but Y,” “not only but also,” and the rule of three, in which three related ideas are grouped together for emphasis. ChatGPT and Claude use these constructions particularly often compared with human writing. AI-detection tools can help identify machine-generated text, but they are imperfect and can produce false positives. The article's central conclusion is that AI writing may display a combination of formulaic structures, sophisticated vocabulary and limited stylistic variation, but these characteristics are changing rapidly. As AI models become increasingly capable of producing human-like prose, reliably distinguishing AI-generated writing from human writing is becoming more difficult.

Key Takeaways
  1. AI writing is increasingly widespread, appearing in journalism, education, science and everyday communication.
  2. No single feature reliably identifies AI writing; several linguistic characteristics need to be considered together.
  3. Sophisticated vocabulary and technical terminology are common characteristics of AI-generated prose.
  4. Nominalisations and longer, more formal words can occur more frequently in AI-generated writing.
  5. Punctuation patterns vary between AI models, so the presence of em dashes alone is not a reliable indicator.
  6. Longer sentences and limited variation in sentence length are recurring characteristics of AI-generated prose.
  7. AI models frequently employ rhetorical patterns such as “not X but Y” and the rule of three.
  8. AI-detection software is imperfect and can produce false positives.
  9. AI writing changes as models are updated, meaning that detection techniques based on older models may become less effective.
  10. The increasing ability of AI systems to produce human-like prose makes reliable identification progressively more challenging.

ARTICLE]]> <![CDATA[The End of Mathematics]]> https://mklab.gr/showthread.php?tid=2305 Wed, 30 Sep 2026 16:00:15 +0300 mklabgr]]> https://mklab.gr/showthread.php?tid=2305 “The End of Mathematics” imagines a future in which AI becomes robustly superhuman at mathematics, while emphasizing that this scenario is a thought experiment rather than a prediction. The central concern is not that AI will simply solve mathematics, but that mathematical progress and human understanding could stall because of the way the mathematical profession responds to AI. Litt begins by observing a rapid increase in mathematical output, including a substantial rise in papers, while questioning how much of this material is genuinely interesting, correct, or meaningfully read. At the same time, community institutions such as MathOverflow appear to be declining, with both questions and answers decreasing. He also highlights the emergence of duplicated AI-assisted discoveries, such as multiple groups independently producing similar proofs, suggesting that increasingly capable models may cause researchers to spend large amounts of effort rediscovering results rather than generating genuinely new mathematical directions. 

For 2027, Litt imagines that existing academic incentives could become badly misaligned with mathematical value. If careers continue to reward papers, theorem-proving, and resolved conjectures, researchers could use AI systems to generate and verify large numbers of short papers, effectively treating conjectures as a “slot machine.” The important problem, in his view, is that such productivity may produce little mathematical expertise or understanding: researchers could become increasingly detached from the mathematics underlying their own publications. AI could also make it dangerous to discuss unfinished research, because models might reconstruct a paper from a few key ideas before the human researcher has completed it. Litt therefore worries about a breakdown of the traditional collaborative culture of mathematics, in which researchers discuss ideas, share work in progress, and gradually develop understanding together. 

In 2028, Litt sees an important positive development in autoformalization: AI could become sufficiently effective at translating mathematical arguments into formal systems that it discovers and repairs gaps or errors in the existing literature. However, this introduces another form of disconnection. A formalization may differ subtly from the original English mathematical statement in ways that are not obvious to readers, making it easier to trust the formal proof while losing confidence in the underlying mathematical ideas. From 2029 onward, Litt imagines AI systems performing essentially all major research functions—theory-building, conjecturing, proving, and iterative problem-solving—while human mathematicians increasingly act as managers of computational agents. The resulting question is how the next generation will be trained and who will actually understand the mathematics being produced. If academic institutions continue rewarding paper production rather than deep engagement, they could inadvertently reward researchers who have little connection to mathematical understanding. 

Litt ultimately presents the scenario as a warning about institutional incentives, rather than an argument that mathematics itself is doomed. The mathematical profession already contains weaknesses, and he argues that the arrival of highly capable AI may expose or intensify them. Existing institutions aim to promote high-quality science, human capital, and human understanding, but the mechanisms used to achieve these goals may not remain effective when AI can cheaply produce mathematical outputs. Nevertheless, Litt ends on an explicitly optimistic note, saying that he expects mathematics to survive and flourish and that humans have an opportunity to learn and understand extraordinary things. His broader hope is that considering these problems now will help mathematics—and potentially other disciplines—adapt their institutions so that technological progress does not come at the expense of human understanding.

Key Takeaways
  1. AI-superhuman mathematics is treated as a premise, not a prediction; Litt uses it to explore a possible future in which mathematical progress nevertheless stalls.
  2. Mathematical output is rapidly increasing, but its correctness, significance, and actual human engagement are uncertain.
  3. AI can produce duplicated discoveries, potentially making large amounts of mathematical labor redundant.
  4. Academic incentives may become misaligned when papers and solved conjectures become extremely cheap to produce.
  5. Human mathematical understanding is the central concern: producing correct results is not necessarily equivalent to understanding them.
  6. AI could weaken mathematical communities by reducing discussion, collaboration, and willingness to share work in progress.
  7. Autoformalization could repair mathematical literature, but it may simultaneously increase the gap between formal correctness and human understanding of mathematical ideas.
  8. From 2029 onward, AI could potentially perform most functions currently performed by mathematical researchers, leaving humans primarily directing computational agents.
  9. Mathematics education and research institutions may need to adapt if they are to continue producing mathematicians who deeply understand mathematics.
  10. Litt's conclusion is optimistic: the danger lies primarily in failing to adapt institutions and incentives, rather than in mathematics itself inevitably coming to an end.

ARTICLE]]> “The End of Mathematics” imagines a future in which AI becomes robustly superhuman at mathematics, while emphasizing that this scenario is a thought experiment rather than a prediction. The central concern is not that AI will simply solve mathematics, but that mathematical progress and human understanding could stall because of the way the mathematical profession responds to AI. Litt begins by observing a rapid increase in mathematical output, including a substantial rise in papers, while questioning how much of this material is genuinely interesting, correct, or meaningfully read. At the same time, community institutions such as MathOverflow appear to be declining, with both questions and answers decreasing. He also highlights the emergence of duplicated AI-assisted discoveries, such as multiple groups independently producing similar proofs, suggesting that increasingly capable models may cause researchers to spend large amounts of effort rediscovering results rather than generating genuinely new mathematical directions. 

For 2027, Litt imagines that existing academic incentives could become badly misaligned with mathematical value. If careers continue to reward papers, theorem-proving, and resolved conjectures, researchers could use AI systems to generate and verify large numbers of short papers, effectively treating conjectures as a “slot machine.” The important problem, in his view, is that such productivity may produce little mathematical expertise or understanding: researchers could become increasingly detached from the mathematics underlying their own publications. AI could also make it dangerous to discuss unfinished research, because models might reconstruct a paper from a few key ideas before the human researcher has completed it. Litt therefore worries about a breakdown of the traditional collaborative culture of mathematics, in which researchers discuss ideas, share work in progress, and gradually develop understanding together. 

In 2028, Litt sees an important positive development in autoformalization: AI could become sufficiently effective at translating mathematical arguments into formal systems that it discovers and repairs gaps or errors in the existing literature. However, this introduces another form of disconnection. A formalization may differ subtly from the original English mathematical statement in ways that are not obvious to readers, making it easier to trust the formal proof while losing confidence in the underlying mathematical ideas. From 2029 onward, Litt imagines AI systems performing essentially all major research functions—theory-building, conjecturing, proving, and iterative problem-solving—while human mathematicians increasingly act as managers of computational agents. The resulting question is how the next generation will be trained and who will actually understand the mathematics being produced. If academic institutions continue rewarding paper production rather than deep engagement, they could inadvertently reward researchers who have little connection to mathematical understanding. 

Litt ultimately presents the scenario as a warning about institutional incentives, rather than an argument that mathematics itself is doomed. The mathematical profession already contains weaknesses, and he argues that the arrival of highly capable AI may expose or intensify them. Existing institutions aim to promote high-quality science, human capital, and human understanding, but the mechanisms used to achieve these goals may not remain effective when AI can cheaply produce mathematical outputs. Nevertheless, Litt ends on an explicitly optimistic note, saying that he expects mathematics to survive and flourish and that humans have an opportunity to learn and understand extraordinary things. His broader hope is that considering these problems now will help mathematics—and potentially other disciplines—adapt their institutions so that technological progress does not come at the expense of human understanding.

Key Takeaways
  1. AI-superhuman mathematics is treated as a premise, not a prediction; Litt uses it to explore a possible future in which mathematical progress nevertheless stalls.
  2. Mathematical output is rapidly increasing, but its correctness, significance, and actual human engagement are uncertain.
  3. AI can produce duplicated discoveries, potentially making large amounts of mathematical labor redundant.
  4. Academic incentives may become misaligned when papers and solved conjectures become extremely cheap to produce.
  5. Human mathematical understanding is the central concern: producing correct results is not necessarily equivalent to understanding them.
  6. AI could weaken mathematical communities by reducing discussion, collaboration, and willingness to share work in progress.
  7. Autoformalization could repair mathematical literature, but it may simultaneously increase the gap between formal correctness and human understanding of mathematical ideas.
  8. From 2029 onward, AI could potentially perform most functions currently performed by mathematical researchers, leaving humans primarily directing computational agents.
  9. Mathematics education and research institutions may need to adapt if they are to continue producing mathematicians who deeply understand mathematics.
  10. Litt's conclusion is optimistic: the danger lies primarily in failing to adapt institutions and incentives, rather than in mathematics itself inevitably coming to an end.

ARTICLE]]>