Terence Tao is challenging what he calls an “excessive emphasis on automated solving of open problems by AI.” In his newly posted Math 2.0 announcement, the mathematician describes an alternative in which AI tools sustain the mathematical community and strengthen the role of mathematical understanding in real-world applications.
His proposal asks whether AI leaves mathematics better equipped to explain results, develop researchers, and guide the use of mathematical ideas outside the discipline.
Tao published the post on October 10, 2026, sharing slides from his public lecture at Caltech the previous evening. For the AI industry, his choice of objective carries the most weight: automated breakthroughs should contribute to a healthy mathematical system, without becoming the sole measure of its success.
The short announcement establishes that direction. It should not be stretched into a detailed policy platform or treated as a transcript of everything he said.
The Lecture Changed, but the Event Listing Did Not
The lecture’s two titles reflect a change in its subject. Caltech’s event listing calls the October 9 talk “Machine Assisted Proof.” Its abstract describes a long history of computer assistance and says interactive theorem provers, machine learning, and generative AI are enabling machines to participate in more creative and substantive ways.
Tao explains that the lecture had been planned more than a year earlier, initially for May. After its postponement to October, he says, recent events significantly changed the talk, while the old title and abstract remained on Caltech’s website.
The event description introduces the expanding capabilities of mathematical tools. Tao’s subsequent announcement raises a more pointed concern about the purposes those tools serve. The post does not specify which “recent events” prompted the changes; assigning the lecture to a particular company announcement or controversy would go beyond its account.
Caltech identifies Tao as a UCLA mathematics professor and 2006 Fields Medal recipient. He is intervening as a mathematician concerned with the direction of AI-assisted research. His argument deserves attention on those terms, without recasting him as an AI company executive or making unsupported claims about industry reactions.
Solving an Open Problem Is Only One Kind of Progress
Tao’s criticism concerns excessive emphasis on automated open-problem solving. It does not establish that he opposes solving open problems with AI, or that successful automated proofs lack mathematical value. Valuing an achievement is different from organizing the entire field around it.
An open problem provides a clear target, and a successful solution can establish something previously unknown. Yet the existence of a solution leaves other questions unanswered: why the result holds, how the method connects to other work, which assumptions matter, or whether the approach helps researchers make further progress.
Tao’s living AI summary characterizes current AI tools as powerful but unreliable, with uneven capabilities. It argues for interactive use instead of accepting a single output without scrutiny.
The summary also explains why mathematics is attractive for AI development: proofs offer opportunities for checking outputs. That makes mathematical success easier to establish than success in settings where errors are difficult to verify.
Verification has a defined scope. A formally checked proof establishes a conclusion within its formal assumptions. It does not, by itself, explain why that theorem is useful, whether its assumptions suit an application, or how a reader should understand the proof’s central idea.
For evaluation, the risk is straightforward. If the only rewarded outcome is “problem solved,” work that improves explanation, collaboration, or the reuse of mathematical knowledge risks becoming secondary. Tao’s stated objection concerns that imbalance, not a prohibition on automation.
Math 2.0 Connects Community Health With Real-World Use
Tao describes the historical state of mathematics as “Math 1.0,” then contrasts it with the current emphasis on automated solutions and his proposed future. His announcement identifies two responsibilities for that future: sustaining the internal mathematical community and sustaining mathematical understanding’s role in advancing and aligning applications to the real world.
Inside mathematics, a result becomes more useful when people can examine it, explain it, teach it, and build on it. Outside mathematics, applying a result requires judgment about what the mathematical objects represent and where the model’s assumptions stop matching reality. These responsibilities are connected.
A proof about an idealized model, for example, may be correct while the model remains unsuitable for a particular practical decision. This is an illustrative distinction, not an example reported from Tao’s lecture. Correctness and appropriate application require different kinds of scrutiny.
The wording about “aligning” applications also needs to be read in context. Tao’s post places alignment within the relationship between mathematical understanding and real-world applications. It does not announce a new technical method for aligning general-purpose AI systems.
His framework expands the question researchers and developers should ask. Alongside “Can the system solve this?”, there is “Does its use improve the community’s ability to understand and apply mathematics?” The second question is harder to reduce to a single score, yet central to the future Tao describes.
Labs Could Evaluate What Their Outputs Let People Do
AI labs could assess more than whether a system reaches a correct result. The following are editorial applications of Tao’s stated goals, not a checklist he publishes in the announcement.
One approach would examine whether researchers can understand and extend a generated argument. Does the system identify the assumptions it uses? Can it distinguish established steps from claims that still need checking? Does its explanation help a mathematician recognize which parts of an argument might transfer to another problem?
These questions should supplement rigorous correctness checks. A persuasive explanation can still be wrong, while a correct proof can still be difficult to use.
Evaluation could also account for the work transferred to human researchers. If an AI system produces a promising argument but leaves substantial checking, clarification, and repair to others, that effort belongs in the assessment. The comparison needs to cover the complete research workflow as well as the machine’s output.
Difficult mathematical benchmarks still have a place, with limits worth recognizing. Solving a selected problem can demonstrate a capability without demonstrating that the system strengthens research practice more broadly.
For developers, this suggests a product priority: tools that help people inspect reasoning, correct mistakes, and retain control of an investigation. Such tools may produce less dramatic demonstrations than an announced breakthrough, but they fit Tao’s goal of sustaining mathematical capacity.
Funders Could Support the Work Around the Breakthrough
A funding program organized entirely around automated solutions would reward different work from one designed to strengthen mathematical understanding and participation.
Sources
- Math 2.0 announcementterrytao.wordpress.com
- Caltech’s event listingcaltech.edu
- Tao’s living AI summaryteorth.github.io
