New MIT Research Reveals AI Is On Track to Upend Multiple Fields of Theoretical Science

Last week, I joined a campus talk at MIT hosted by Neil Thompson, a leading researcher specializing in computing, artificial intelligence, and the trajectory of global scientific progress. In just 30 data-packed minutes, Thompson laid out a rigorous, evidence-backed case that a wave of AI-driven disruption is on the verge of reshaping a slew of established scientific disciplines entirely.

As working mathematicians are already reeling from AI’s rapid automation of core research workflows in their field, Thompson’s latest co-authored study identifies several high-impact domains next in line for that same transformation: probability and statistics theory, economic theory, high-energy physics, gravitational physics and cosmology, and quantum physics. This projection stems from staggering performance gains recorded by frontier AI models over a narrow 24-month window: between mid-2024 and mid-2026, the share of complex theoretical problems across these fields that top-tier AI models could solve correctly skyrocketed from just 2% to roughly 35%, with no visible sign of this improvement curve slowing down. “Within a handful of years, we expect the landscape of theoretical inquiry across all these fields to be completely redefined by AI’s capacity to tackle core open research questions independently,” Thompson shared during his presentation.

Thompson’s lab broadly focuses on mapping how advances in computer science shift societal systems and alter the fundamental mechanics of human innovation. For this new study, which he led alongside PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun, the team designed a strict, bias-free testing framework to avoid the common pitfall of benchmark data leaking into AI training sets. They curated a full bank of research problems pulled from preprints posted online between December 2025 and May 2026, then ran a suite of OpenAI’s most advanced models — o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5 — against the full problem set.

Critically, every verified, peer-reviewed solution to these problems was published after the official training cutoff date for GPT-5.5 and all earlier models tested. This deliberate design guaranteed none of the AI systems could have encountered the problems or their answers during training, forcing them to derive full, rigorous solutions from scratch. The post-cutoff published papers also served as official ground truth for grading model outputs, and the researchers fully disabled internet access for every test run to eliminate any possibility of the models looking up external solutions mid-task.

The study’s scope is intentionally limited: the team only selected fields where problems can be solved via pure logical deduction, with no requirement for physical simulations, lab trials, or experimental data collection. Even with that narrow framing, the findings make clear that the seismic disruption already rattling the mathematics community is set to ripple out to every other math-heavy, theory-driven corner of scientific research.

Back in September, OpenAI sent shockwaves through the global math world when it announced that an unreleased internal model had cracked the Navier-Stokes equations, a decades-old legendary grand challenge focused on describing the behavior of fluid flow. The announcement quickly spiraled into a fierce academic controversy, after a New York University mathematician (who had been collaborating with another researcher affiliated with Anthropic) publicly accused OpenAI of scooping their team’s breakthrough after learning of their unpublished progress. In the months following that reveal, OpenAI and other leading AI labs have unveiled hundreds more verified solutions to long-standing unsolved math problems, triggering a widespread reckoning that many researchers describe as a field-wide crisis. A cohort of prominent mathematicians has even called for targeted guardrails and limits on unregulated AI use in research, arguing that the iterative, creative process of working through problems manually is often as educationally and scientifically valuable as arriving at the final correct answer.

Thompson notes that, just like the ongoing shifts in mathematics, coming AI breakthroughs in other theoretical fields will bring a complicated mix of positive and negative impacts. He points out that foundational theoretical leaps are almost always the precursor to real-world, practical scientific innovations, and that widespread AI adoption could dramatically accelerate that pipeline of progress. A growing number of top-tier AI researchers have already launched new dedicated labs focused exclusively on AI for scientific discovery, and every major generative AI firm is ramping up in-house scientific R&D programs to build tools that can support both theoretical analysis and even run physical lab experiments alongside human researchers.

Even as AI grows capable of solving core open theoretical problems, human experts will not become obsolete, the team emphasizes. “We are still going to need people to generate the big, creative new ideas that connect disparate lines of research and point AI toward meaningful, high-impact problems,” Manso explains. He adds that even the most advanced AI models are prone to subtle, hard-to-spot errors on extremely complex tasks — a point backed up by OpenAI’s own recent move to retract several purported math solutions from its public dataset after post-publication review revealed flaws in their reasoning. “We will also need robust, human-led verification systems to check every output,” Manso adds. “It’s trivially easy for a model to spit out a result and claim it’s the correct answer, but once you walk through the full line of reasoning step by step, you will often find gaps or errors no one spotted at first glance.”

---

This piece is the latest edition of Will Knight’s AI Lab newsletter. You can access the full archive of past issues here.