AI's scientific crisis: 49 out of 50 in physics and 722 initial manuscripts

An AI system receives fifty unpublished theoretical physics problems. It solves forty-nine. The most striking part is not just the score: when researchers revisit some answers initially marked wrong, they discover mistakes in their own reference solutions.

At the same time, OpenAI releases hundreds of mathematical manuscripts, researchers use Gemini agents to extend analytical results relevant to plasma confinement, and Anthropic announces a Claude-assisted discovery of an unusual enzyme system. The question is no longer merely whether AI can help write scientific papers. It is what happens when producing a potential scientific contribution becomes much cheaper than reviewing it.

In La crisis científica que creó OpenAI, Freddy Vega brings these developments together into an optimistic argument: science may be entering an age of abundance. It is worth examining that perspective while carefully separating measurements, inferences and predictions.

Central thesis: there is real evidence that AI is changing the cost and speed of specific research tasks. That is not the same as proving that AI can independently choose the most important scientific questions, establish that every result is valid, or turn theoretical work into useful technology on demand.

1. A physics test where some human answer keys were wrong

The TheoPhysEval study, released as a preprint on October 9, 2026, assembled 50 unpublished theoretical physics problems with verifiable answers. According to its authors, a world expert could generally solve a typical task in about a day or less, while specialists from adjacent fields would need substantially longer.

The best agent, GPT-6 Astra in the Codex harness, answered 49 of the 50 correctly (98%) when given realistic research tools: literature search and code for numerical or symbolic calculations. The reported average cost was less than US$3 per problem.

The surprising episode came during grading. Physicist and coauthor Suvrat Raju explains that the model initially appeared to score in the high eighties, but challenged several reference answers. Upon revisiting the calculations, experts found subtle errors and the score rose to 98%.

Two assumptions are now under pressure: that human-produced answer keys are necessarily infallible, and that AI can only reproduce problems it has seen before.

Important limits remain:

These were not fifty world-changing discoveries. The team deliberately chose concrete, checkable problems rather than asking the AI to set a long-term research agenda.
The score describes a particular evaluation. It is not a measured success rate for arbitrary questions across all scientific disciplines.
The harness matters. The result belongs to a model equipped with research and computation tools, not a tool-free chat response.
Human work did not vanish. Constructing the benchmark, investigating discrepancies, and assessing the answers required expert judgment.

The useful conclusion is not that scientists are obsolete. It is that AI agents can now function as technically capable counterparts that challenge specialists’ calculations.

2. OpenAI and the mathematical proof-review bottleneck

On October 6, 2026, OpenAI announced a release of mathematical results generated by an internal frontier model. The original release contained 722 manuscripts organized into 372 families. But the public repository currently lists 719 manuscripts: on October 7, OpenAI withdrew three papers after a sign error invalidated a central argument and two dependent results. Fourteen other manuscripts were also revised. This is a concrete example of why checking the output matters.

A crucial distinction: The original 722 documents did not constitute 722 distinct theorems independently accepted by mathematicians. A family can contain a primary result, follow-up arguments and alternative proofs. Verification status also varies. Some manuscripts have Lean formalizations; others still need checking.

The research pipeline starts to look different:

BEFORE
question → months/years of exploration → proof
         → review → shared knowledge

WITH AI AGENTS
thousands of attempts → many candidates and manuscripts
         → formal checking + expert review
         → selection, explanation and integration

The bottleneck shifts toward detecting mistakes, establishing novelty, checking assumptions and explaining why findings matter.

In his Math 2.0 lecture, mathematician Terence Tao argues that mathematics is more than closing open problems. It is also about developing understanding, connections and communities able to apply knowledge.

The danger is not merely an incorrect paper. It is an ecosystem that rewards the production of thousands of candidate proofs while failing to support the people and tools needed to audit them.

We previously explored in Capital de Tokens why solving is not the same as understanding. The distinction is now relevant far beyond pure mathematics.

3. Stellarators: AI agents proposing useful mathematics for fusion

A stellarator aims to confine extremely hot plasma with complex three-dimensional magnetic fields. Understanding the equilibrium between pressure and magnetic forces supports the design and validation of simulations.

In Analytic non-axisymmetric 3D MHD equilibria without stellarator symmetry, posted on October 5, 2026, Opal Issan, Jonathan Citrin, Matt Landreman and Brendan Tracey generalized two previous families of analytical solutions by removing a symmetry restriction.

The study’s AI-tools disclosure describes work with Gemini 4 agents in Google Antigravity. The agents proposed promising candidates within hours. The researchers checked the derivations, developed additional verification and retained scientific responsibility for the paper.

This is a tangible example of researcher + agents + independent verification. But the result concerns analytical solutions and reference cases for simulation codes. It does not mean that commercial fusion reactors have been completed or that plasma confinement is a solved engineering problem.

Materials, stability, manufacturing, cost and scale still separate a mathematical advance from abundant electricity.

4. Biology: Claude spots an unusual enzyme system, but its function is unresolved

In September, Anthropic reported that Claude helped identify a system involving a reverse transcriptase, repeated DNA sequences and an accessory protein. Anthropic calls the system ART and notes features reminiscent of CRISPR-associated repeat arrays.

The compelling part is the search process: analyzing large genomic datasets, recognizing unusual patterns and proposing a testable biological hypothesis.

Precision matters here too. Anthropic says the system’s function is still unknown. It has not demonstrated that ART is an established replacement for CRISPR-based genome editing in medicine. This is a promising experimental lead, not a treatment.

Biology highlights a different opportunity from physics: when the search space is vast, agents may help determine which hypotheses deserve laboratory experiments.

5. Plausible applications versus speculative promises

The video looks ahead to extraordinary outcomes: hurricane forecasts weeks in advance, widespread fusion energy, almost unhackable software, highly capable on-device AI and dramatically longer battery life.

How do those claims compare with what the evidence establishes today?

Possibility	What current evidence supports
Assistants accelerating technical scientific calculations	Demonstrated in particular physics and mathematics evaluations
Agents generating testable scientific hypotheses	Documented examples in biology and physics
Better simulation tools for fusion	A plausible research benefit, not proof that commercial fusion is solved
Formal verification reducing software bugs	Useful for correctly specified components; it cannot guarantee that entire systems are impossible to compromise
Perfect hurricane forecasts or turbulence-free flights	Speculation; better models do not erase measurement uncertainty or physical limits
Revolutionary batteries or fully local frontier AI	Also dependent on hardware, memory, power, manufacturing and future measured advances

The operative word is possibility, not certainty. New scientific capabilities expand the set of plausible applications, but each practical leap requires additional evidence.

6. A new discipline: reviewing agent-produced science

To turn abundance into reliable knowledge rather than noise, AI-assisted laboratories need at least six practices.

Provenance: preserve relevant prompts, model versions, tools, data and code.
Reproducibility: make it possible for independent teams to repeat the experiment or calculations.
Independent verification: separate the proposing agent from the reviewer; use Lean or other formal proof systems where appropriate.
Adversarial checking: actively search for counterexamples, hidden assumptions and failure conditions.
Novelty assessment: compare claims with existing literature and recognize earlier contributions.
Human understanding: explain not only whether something works, but why, when and for what purpose.

Not every result requires a formal proof, and a formal proof is not evidence that a model accurately represents physical reality. Experimental science additionally requires replication, measurements and physical evidence.

This is reminiscent of engineering with coding agents: success comes from a pipeline of specification → implementation → tests → review → integration, not merely generated output. Science needs an even stronger version of that pipeline because empirical truth is not decided by a compiler.

7. The human side of the crisis

If an AI can propose in hours what previously took months, academic careers, doctoral training and the economics of specialized work may change.

The concerns are substantial: overwhelmed reviewers, low-quality publications, disputed attribution and weaker incentives for young researchers to invest years in academic careers.

But there is an opportunity too: less wealthy research teams may explore more ideas; smaller laboratories may evaluate more hypotheses; and scientists across disciplines may collaborate with an assistant competent in tools they could never personally master.

None of that automatically ensures an equitable distribution of benefits. Access to computation, instruments, datasets, funding and independent review could concentrate advantages unless institutions address those constraints.

Conclusion: not the end of science, but the end of one bottleneck

The 98% score on fifty theoretical physics tasks is striking. The original 722 mathematics manuscripts—719 remain after three withdrawals— pose a challenge of scale. The examples from stellarator physics and biology show AI taking part in genuinely exploratory work, not just drafting text.

Yet the greatest advance would not be another thousand papers. It would be a process that makes machine-assisted discoveries auditable, reproducible, understandable and safe to apply.

In the emerging research landscape, asking good questions may remain profoundly human; performing calculations may become cheaper; and verification, interpretation and the judgment to choose what deserves study may matter more than ever.