From Proofs to Plasma: GPT-5 Speeds Discovery

Early experiments in accelerating science with GPT-5

AI's promise in research is simple: shorten the path from idea to tested result. Recent polling shows the need is real-60% of people in the U.S. feel breakthroughs reach them too slowly, 73% want better ways to speed discovery, and 69% see scientific leadership as a national priority.

A new paper, "Early science acceleration experiments with GPT-5," shares case studies across math, physics, biology, computer science, astronomy, and materials science. In each, expert teams used GPT-5 to synthesize results, run conceptual literature searches, explore hard computations, and propose novel arguments-and they also documented failures. The point isn't hype. It's a clear read on what these systems can and can't do today.

Why this matters

• Shortened cycles: researchers move from idea to validation faster.
• Broader search: conceptual literature review surfaces links across fields and languages.
• Human-AI teams: scientists set the agenda; models provide speed, breadth, and alternative routes.
• No autonomy: GPT-5 does not run projects on its own-expert oversight is still essential.

What is OpenAI for Science?

The mission is to accelerate discovery by pairing frontier models with the right tools, workflows, and collaborations across universities, national labs, and industry. Two beliefs guide the work: specialized scientific tools (simulators, protein databases, CAS) are non-negotiable for precision, and scaling foundation models keeps unlocking new reasoning abilities-connecting ideas, sketching proofs, proposing mechanisms, and scanning literature conceptually rather than by keyword. Used together, they compound.

How scientists are working with GPT-5 today

Productive work looks like dialogue. Researchers pose questions, push back, decompose problems, and validate. GPT-5 explores branches in parallel, drafts outlines, critiques gaps, and suggests tests. Skill matters-experts learn how to prompt, when to stop, and what to verify independently.

Highlights from early case studies

• Biology (immunology, CAR-T): From an unpublished flow cytometry chart, GPT-5 proposed disrupted N-linked glycosylation during priming as the driver of a lasting T-cell shift after 2DG exposure, predicted memory T cells as the actors, and suggested a mannose rescue experiment. The lab had already run it; the prediction matched. It also forecast improved killing for CAR-T cells after transient 2DG pulsing, aligning with unpublished data. For background on CAR-T, see the National Cancer Institute overview: NCI: CAR-T cells.
• Mathematics: Mehtaab Sawhney and Mark Sellke were stuck on the last step of a decades-old Erdős problem. GPT-5 contributed the missing idea: how a single "out-of-pattern" number forces contradictions across almost all others-closing the proof.
• Algorithms & optimization: Sébastien Bubeck and Christian Coester used GPT-5 to expose a clear counterexample to a widely trusted decision method and to tighten a recent convex optimization theorem with a sharper step-size bound and cleaner proof structure.
• Physics (black holes): After a warm-up on a simpler system, GPT-5 reconstructed the hidden SL(2,ℝ) symmetry algebra of the Kerr wave equation, matching human results. Context on the Kerr solution: Kerr metric (Wikipedia).
• Deep literature search: Given a new convex geometry theorem, GPT-5 mapped concrete links to density estimation, learning theory, and multi-objective optimization, surfacing specific references-including non-English sources researchers had missed.
• Erdős database cleanup: GPT-5 located existing solutions for problems still marked "open," identified strong partial results, and even caught a misprint. It also suggested a density estimate that, after human refinement, completed Erdős Problem #848.
• Cautionary tale (clique-avoiding codes): GPT-5 reframed the problem via quadratic equations over a finite field and pointed to a classical theorem, yielding the optimal lower bound. The same argument existed in earlier literature, which the model did not cite until asked explicitly. Correctness ≠ attribution-humans must verify both.
• Working style (combinatorics): Tim Gowers used GPT-5 as a fast critic to stress-test constructions, spot missing cases, and propose counterexamples. Useful, but not yet at a level for co-authorship.
• Cosmology: GPT-5 helped sanity-check derivations, translate between parameterizations of dark energy, and flag algebraic slips-reducing the gap from a notebook idea to something testable.
• Fusion and plasma physics: Teams used GPT-5 to build a reduced reaction-diffusion model of burn propagation, run parameter sweeps, find a ridge of optimal density profiles, and propose an energy-balance explanation that guides simple engineering rules. Oversight corrected occasional unstable runs and overconfident takes.

New scientific results obtained with AI assistance

• Erdős number theory: The final insight needed to finish Erdős Problem #848 came from GPT-5's "one misfit number constrains the rest" idea, which the authors proved out.
• Online algorithms: A GPT-5-suggested geometric construction led to a cleaner, stronger lower bound for convex body chasing.
• Graph theory inequalities: With custom math scaffolding, GPT-5 produced short, self-contained proofs of two inequalities in trees (one previously conjectured). Humans checked and adopted the argument.
• Identifiability in growing networks: Focusing on the long-run fraction of leaves unlocked a direct, provable way to recover a hidden attachment parameter from a single final tree snapshot.

What GPT-5 can do today

• Shorten portions of the workflow: proof sketches, sanity checks, symmetry discovery, and conceptual literature search.
• Suggest mechanisms and experiments that experts can validate-useful in biology, high-energy density physics, and materials science.
• Offer alternative proof routes and counterexamples in math and theoretical CS, where fast feedback loops exist.

Limitations to keep in mind

• Attribution gaps: can reproduce known arguments without citing sources unless prompted.
• Hallucinations: plausible citations, mechanisms, or proofs that don't hold up.
• Scaffolding sensitivity: warm-ups and problem framing can change outcomes.
• Domain blind spots: misses subtleties or follows dead-ends without intervention.

Bottom line: expert oversight is required. Treat outputs as drafts to verify, not results to trust by default.

Practical guidance for research leaders

• Define where the model fits: conceptual search, early proof sketches, sanity checks, and hypothesis generation.
• Use tools together: pair GPT-5 with simulators, CAS, and domain databases for precision and speed.
• Time-box deep reasoning: let the model think longer on difficult steps; compare multiple runs.
• Instrument for verification: unit tests for math proofs, replication scripts for simulations, blinded checks for analysis.
• Track provenance: log prompts, seeds, references, and versioned datasets for reproducibility and credit.
• Audit attribution: ask explicitly for sources; cross-check citations before adoption.
• Set review gates: require human sign-off before experiments, submissions, or lab deployments.

What's next

These studies show GPT-5 already helps experts prove theorems, recover hidden structures, connect literatures, and generate testable mechanisms. The model is not autonomous, but with more time and compute, we expect deeper results-think minutes to hours, hours to days-with tight human loops.

If you're training your team to work productively with AI in research workflows, see curated options by role here: AI courses by job.