Early experiments in accelerating science with GPT-5
Science moves the world forward, but turning ideas into validated results still takes too long. Recent polling shows a clear mandate: most people think breakthroughs reach them too slowly and want better ways to speed discovery. If AI can compress the time from hypothesis to tested result, the benefits ripple across health, energy, and security.
This article summarizes early experiments using GPT-5 with researchers across institutions including Vanderbilt, UC Berkeley, Columbia, Oxford, Cambridge, Lawrence Livermore National Laboratory, and The Jackson Laboratory. The work spans math, physics, biology, computer science, astronomy, and materials science-and documents both progress and limits.
What we studied
Across dozens of case studies, GPT-5 helped experts synthesize prior work, explore tricky computations, propose experiments, and in some cases outline novel proofs. These are not autopilot results. They are human-AI collaborations where domain experts set the agenda and validate every step.
What is OpenAI for Science?
The mission: accelerate discovery by pairing frontier models with the right tools and workflows. The approach rests on two beliefs:
- Specialized scientific tools (simulators, databases, CAS) are vital for precision.
- Scaling foundation models adds cross-domain reasoning: connecting ideas, sketching proofs, proposing mechanisms, and reading literature conceptually, not just by keyword.
Use specialized tools when they exist. Use general reasoning when the path is unclear. Both approaches reinforce each other.
How scientists are working with GPT-5 today
- Experts define problems, critique ideas, and validate results. GPT-5 adds breadth, speed, and parallel exploration.
- Effective use is a skill: ask crisp questions, break problems into steps, push back, and verify independently.
- Productive sessions look like dialogue-iterate until a direction survives scrutiny or gets discarded.
Emerging capabilities
- Conceptual literature search: find deeper links across fields, languages, and less-visible sources.
- Fast proof ideation in math and TCS: viable outlines in minutes, with quick feedback loops.
- Physics and computational domains: propose simplifying transformations and structural analogies.
- Biology and empirical sciences: suggest mechanisms and concrete experiments for lab validation.
GPT-5 does not run projects end-to-end. In the hands of experts, it trims cycles and expands the search space around the right answers.
Case studies at a glance
Biology: mechanistic insight in immunology
After months puzzling over T-cell behavior under transient 2DG exposure, GPT-5 analyzed an unpublished flow cytometry figure and proposed a mechanism in minutes: disrupted N-linked glycosylation during priming, driven by memory T cells. It suggested a mannose rescue experiment (restoring N-glycosylation without restoring glycolysis) that matched prior unpublished results. It also predicted improved CAR-T killing efficiency after transient 2DG during generation, which aligned with lab data.
Mathematics: progress on Erdős problems and beyond
While working on a decades-old Erdős problem, researchers were stuck on the final step. GPT-5 suggested how a single "out-of-place" number would force contradictions across the set, unlocking the complete proof. In separate work, the model helped clean up entries in the Erdős problem database-surfacing missed solutions, partial progress, and even a misprint-while also proposing a density estimate later tightened by the authors.
Algorithms and optimization
Given a recent theorem on when gradient descent iterates form a convex curve over time, GPT-5 proposed a sharper step-size bound and a cleaner proof that a researcher then verified by hand. In online algorithms, it suggested a geometric construction that led to stronger lower bounds after human refinement. It also produced clear counterexamples showing failure modes in a common decision-making method used in robotics and routing.
Physics and astronomy
After a warm-up on a simpler case, GPT-5 reconstructed the hidden SL(2,ℝ) symmetry algebra for waves in the Kerr black hole spacetime-matching recent human results. In fusion and plasma work, it helped build and analyze a reduced reaction-diffusion model, ran parameter sweeps, and proposed a physical explanation for an optimal "ridge" of profiles. In cosmology, it sanity-checked derivations, translated between model parameterizations, and pointed to matching literature.
Deep literature search and attribution
For a new convex geometry theorem, GPT-5 identified links to density estimation, learning theory, and multi-objective optimization, surfacing specific references across languages. A cautionary note: in work on clique-avoiding codes, the model reformulated the problem and pointed to a classical theorem yielding an optimal bound-but initially failed to attribute a prior publication with essentially the same proof. Human checks on sourcing remain essential.
New results generated with AI assistance
- Number theory: completed the proof of an Erdős conjecture by isolating the effect of one nonconforming integer on the entire set's structure.
- Online algorithms: a cleaner, stronger lower bound for convex body chasing after refining a model-suggested construction.
- Graph theory: short, self-contained proofs for two inequalities in trees (including a previously conjectured one), later checked and adopted by the authors.
- Network science: a proof that a hidden growth parameter is identifiable from the long-run leaf fraction of the final tree.
How to apply this in your research
- Scaffold the problem: define constraints, success criteria, and known obstacles before asking for ideas.
- Use warm-ups: start with a simplified instance, then scale to the full problem.
- Force alternatives: request multiple distinct approaches, counterexamples, and failure modes.
- Verify everything: check algebra, run simulations, and trace proofs line-by-line.
- Track sources: ask for citations, then independently confirm them. Be strict about attribution.
- Pair with tools: connect the model to CAS, simulators, protein databases, and code executors where appropriate.
- Time-box exploration: short, focused sprints (10-30 minutes) often surface the next testable step.
Limitations to keep in mind
- Citations and proofs can look convincing and still be wrong or misattributed.
- Results can depend on the prompt, scaffolding, or warm-up examples.
- Domain subtleties are easy to miss without expert oversight.
- The model can push down unproductive paths unless steered.
These are active research problems. Expect to validate and iterate.
What's next
These early studies show GPT-5 assisting with theorem proofs, rediscovering structures, surfacing cross-field links, and proposing mechanisms and experiments. It is not autonomous. But in expert hands-and with more time and compute for extended reasoning-we should see deeper results and faster movement from idea to verified outcome.
References and resources
Upgrade your AI research workflow
If you want structured practice in prompting and tooling for scientific work, see these resources:
Your membership also unlocks:
