When AI Meets Proof
AI for DC
Hi
Welcome (back) to The Prompt. We’ve sure had an eventful week. How about you?
AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you:
Fields medalist Terence Tao and our Chief Research Officer Mark Chen on how AI is changing math
Software demand is booming in the age of Codex
Applications for AI in newsrooms
If you find them helpful, make sure you’re signed up for the next issue.
[News] When AI meets proof
Math and theoretical physics are fast becoming the first scientific domains where AI’s productivity gains are unmistakable. That was the clear takeaway yesterday when the OpenAI Forum co-hosted a full-day gathering of top researchers at UCLA’s Institute for Pure and Applied Mathematics (IPAM) to discuss how ChatGPT is changing the practice of discovery in math, particle physics, cosmology, and adjacent fields.
The line-up underscored how quickly this conversation is moving into the scientific mainstream: renowned mathematician Terence Tao and OpenAI Chief Research Officer Mark Chen were joined by OpenAI’s VP of Science Kevin Weil and such luminaries as Caltech mathematician Sergei Gukov, UCSB physicist Nathaniel Craig, Stanford’s Eva Silverstein, Lance Dixon of SLAC National Accelerator Laboratory, UCLA’s Zvi Bern, Wahid Bhimji of Lawrence Berkeley National Laboratory and NERSC, University of Wisconsin physicist Kyle Cranmer, and OpenAI’s Alex Lupsasca for talks, panels, and public discussion.
In a fireside conversation bringing Tao and Chen together again after their first OpenAI Forum discussion in late 2024:
Tao said AI was already valuable for the many proofs, lemmas, and exploratory paths that can slow human researchers, even if it is not yet operating at the deepest frontier of mathematical insight.
Chen emphasized that models are able to work autonomously for longer and longer stretches, with lower error rates and less scaffolding, making it possible to tackle larger and more ambitious problems.
Tao made the point that the real payoff will come from redesigning workflows around AI, not simply inserting AI into old ones, just as cities eventually came to be built around the new technology of cars. In math and physics, that points toward new divisions of labor across problem generation, strategy selection, execution, verification, and communication.
What makes math and theoretical physics stand out when it comes to AI is that both fields have unusually strong norms of verification, and both can generate, test, and refine ideas at scale. We already told you that 2026 is the year when AI was going to revolutionize science. Well, it’s also the year when it might start changing how mathematicians and physicists explore conjectures, validate intermediate results, and collaborate across disciplines.
What happened at IPAM suggests that this shift is moving from theory to practice. – Chris Nicholson, Editorial Lead
Editor’s note: In a new example of AI tools speeding up science, physicists using GPT-5.2 Pro to study how gravity behaves at the tiniest scales discovered that a type of particle interaction they once thought impossible can actually happen under certain conditions. Specifically, when particles that carry gravity (called gravitons) interact in particular ways, the effects don’t cancel out as scientists had long assumed – they can produce real, measurable outcomes. This is another example of how AI can help scientists navigate super complicated mathematical spaces, generate hypotheses, and test proofs faster, potentially speeding up discovery in fundamental physics and other research fields.
[Insight] AI’s great software expansion
Generative AI is one of the fastest-adopted technologies we’ve ever seen, and it already exceeds human performance across several important domains. Its impacts on how we work will be both significant and uneven across industries – new roles will emerge, others will change, and some will go away altogether. One area where we already see these dynamics taking shape is software.
We told you we’d update you on how AI is affecting software developers given how AI coding tools are often framed as a way to replace them. What we’re seeing following the introduction of OpenAI’s Codex and similar tools is that they’re lowering the cost of producing software and making it possible for many more people to build applications.
As a result, we’re starting to see demand for software emerging in places we didn’t expect. Some people call it “vibe coding,” but it’s a hugely democratizing phenomenon with people and small businesses now able to build custom software tools. Some data points and developments to chew on:
Weekly active Codex users have more than tripled since the start of the year, with usage up 5x to over 1.6M weekly active users.
The Codex MacOS app has been downloaded more than 1M times in the month since launch.
Adoption spans both professional developers and first-time coders.
Among those who can now build software with AI: A teacher can create a grading workflow, a shop owner can build a tailored inventory tracker, or a small marketing team can create custom measurement tools. In software, this may look less like one-for-one substitution and more like what happened when spreadsheets like Microsoft Excel spread through the corporations – they didn’t eliminate data analysts, but dramatically expanded how much analysis the economy could produce.
The same dynamic seems to be unfolding in software. Rather than spending most of their time debugging or writing basic code, many of my technical colleagues at OpenAI are focused on orchestrating higher-level tasks: designing complex systems, specifying new features, and directing AI tools to implement them. The result is that each developer can produce far more than before. This increase in productivity could translate into more software development in the future, not less. – Adam Cohen, Head of Economic Policy, OpenAI
[Event] How AI is helping in newsrooms
On March 12, the OpenAI Forum is hosting a discussion among journalists, editors, and technologists about how AI is being used in newsrooms today and can help strengthen the high-quality local and national journalism that is critical for a healthy democracy. Speakers include Jim VandeHei and Mike Allen of Axios, the Boston Globe’s Shira T. Center, and Sarabeth Berman of the American Journalism Project.
After the Forum discussion, an OpenAI Academy workshop will feature use cases and workflows from veteran journalists and entrepreneurs who have integrated AI in their newsrooms.
8:00 AM – 11:00 AM EST on Mar 12
[About] OpenAI Academy
The Academy is OpenAI’s free online and in-person AI literacy trainings for beginners through experts.
OpenAI has called for a nationwide AI education strategy – rooted in local communities in partnership with American companies – to help our current workforce and students become AI-ready, bolster the economy, and secure America’s continued leadership on innovation.
2:00 PM – 2:30 PM EST on Mar 13
[Disclosure]
Graphics created by Base Three using ChatGPT.
Interesting perspective.
In my own research, I have been exploring a related directionthe idea that physical systems can be understood as environment-dependent structures, where particle behavior emerges from interactions between waves, frequencies, motion, and a closed environmental framework.
If this perspective is correct, AI could become a powerful tool not only for analyzing data, but for discovering hidden physical relationships that are difficult to derive analytically.
For example, machine learning can analyze complex spectral patterns and dynamic interactions between fields and particles, potentially revealing new emergent laws that traditional analytical approaches might overlook.
In other words, AI may not just assist mathematics it may help us discover new physical models of reality by recognizing patternsinsystems where wave interactionenvironmental constraints, and motion shape observable behavior.
The intersectionofAIphysicsand mathematical structure discovery may be one of the most important scientific directions of the coming decades.
Was anyone else confused that an article with the subtitle "AI for DC" didn't mention the Pentagon deal?
I saw that an immediately jumped in hoping for some insights and clarity on the topic and just kept scrolling around the article.
It seems a little tone deaf to me.