PickAI moves from math into the rest of science
MIT researchers tell Wired that AI may soon be the default way to tackle theoretical problems in economics, physics and beyond.
After making rapid progress in mathematics, AI may be about to become the standard tool for theoretical work across a range of sciences, according to MIT researchers interviewed by Wired.
The study
Neil Thompson, who studies progress in computing at MIT, worked with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun to measure how well AI models solve theory problems in fields including probability, statistical theory, economic theory, high-energy physics, gravitation and cosmology, and quantum physics.
The team took problems from research papers posted between December 2025 and May 2026 and tested a series of OpenAI models, from o1 through several GPT-5 versions. Importantly, the published solutions appeared after the models' training cutoffs, so the models could not simply have memorized the answers.
The results show steep progress. Between mid-2024 and mid-2026, the best models went from solving about 2 percent of these theoretical problems to about 35 percent. Thompson told Wired he sees no signs of the improvement slowing and expects many theory questions to be transformed within a few years.
Why it matters for builders
A jump from 2 to 35 percent in two years is a reminder that capability curves for reasoning tasks can be steep. Products that seemed unrealistic a year ago, such as assistants that propose derivations, check proofs or explore model assumptions, may now be viable for at least some problems.
For developers building research tools, the opportunity is less about replacing scientists and more about workflow: helping researchers set up problems precisely, verify AI-generated steps, track sources, and compare approaches. Verification is the key feature. A model that is right a third of the time is useful only if the tool makes it easy to tell which third.
It also suggests a useful method for evaluating models in your own domain: collect real problems whose answers were published after a model's training cutoff, and track how successive models perform. That gives you a more honest benchmark than public leaderboards.
Research assistants such as Elicit, Consensus and NotebookLM are listed in our AI Research & Search category.
Source: Wired. This article is PickAI’s own summary and analysis based on the original reporting.
