AI as a scientist needs an evaluation with a known answer key

The problem with a solved problem

The Lex Fridman podcast episode on an AI crisis in math starts from a startling set of claims: machines have recently claimed to solve hundreds of hard mathematical problems, including Navier-Stokes, the quasi-Riemann hypothesis, and the unique games conjecture. The episode discusses an AI-generated proof for Navier-Stokes that constructs a counterexample showing finite-time blowup in three dimensions, which the episode notes is physically counterintuitive. It also discusses a formal Lean proof of Fermat Last Theorem consisting of 13 million lines of code generated by AI with minimal human intervention. These are claims discussed on the episode, not settled verdicts in this story, and that distinction is exactly the point. m:…8662

The episode argues that the current academic system for verifying mathematical proofs is under severe strain due to the volume of submissions, independent of AI. When a proof runs to millions of lines, or when a claimed solution arrives faster than referees can check it, the bottleneck is no longer finding an answer. It is knowing whether the answer counts. The panel also names two traps in thinking about this shift. One is the doorman fallacy, the mistake of reducing a complex job to its most visible task and assuming automation of that task replaces the entire role. The other is the solved axiom, the idea discussed on the episode that any mathematical problem solvable by a human interacting with AI could also have been solved by AI alone. Both ideas are presented as arguments on the episode, and both point to the same gap: a solved claim without a trustworthy check is not yet science. m:…8662

A sandbox that knows the answer

The Latent Space episode on an evaluation for AI scientists offers a concrete answer to that gap. Ian Quigley, co-founder of Leash Bio, described a virtual sandbox environment that allows AI agents to perform medicinal chemistry tasks with instant feedback. The sandbox uses pre-measured DNA-encoded library data, so agents test structure-activity hypotheses against protein targets where the data already exist. That pre-measurement is what makes the evaluation work: the system can run thousands of automated rollouts and score whether an agent keeps exploring a dead end or moves on. Quigley noted that more performant models are significantly better at abandoning unproductive chemical neighborhoods than lower-performing models, which turns a vague idea like scientific judgment into something measurable. m:…1104

The episode also gives the kind of model-level detail that a sandbox makes possible. Astra 6 is currently the top-performing model in the Leash Bio chemistry environment, according to Quigley, while Anthropic Sonnet 5 is described as both expensive and low-performing for these chemistry tasks, and DeepSeek V4 Flash demonstrated relatively strong performance. Frontier models previously struggled to convert SMILES strings effectively, which required internal supervised fine-tuning projects. None of that is a general ranking of these models. It is a ranking inside one environment, on one kind of task, with ground truth available in advance, and that narrowness is a feature. A scientist evaluation that can say where a model fails is more useful than a leaderboard that only says a model is smart. m:…1104

Hypotheses that update themselves

A third item, in a preprint posted this week, tries to build the discovery process itself into the model. Principia, described in the preprint, is an LLM-based framework that constructs and refines scientific principles during inference. It forms comparative causal hypotheses from related inputs and observed outcomes, synthesizes recurring relationships across those hypotheses into inductive principles, and revises hypotheses based on new observations obtained through its principles. The pipeline is hypotheses to principles, repeated as new results arrive, rather than a single answer generated once and left alone. Because this rests on preprint results, it should be read as a proposed method with reported results, not as peer-reviewed evidence that the method works generally. m:…7287

In a preprint, Principia is reported to outperform baselines relying on internal knowledge, external sources, or experience summaries. It is tested on molecular optimization tasks, protein engineering tasks, gene perturbation tasks, and molecular property prediction tasks, and its code is available on GitHub. Those test areas overlap with the chemistry sandbox idea in an instructive way: both treat scientific progress as a loop of hypothesis, observation, and revision, rather than a single flash of insight. The difference is that the Leash Bio sandbox can check the loop against data measured before the agent ever ran, while a preprint report still needs independent replication. m:…7287

What a good scientist test looks like

Read together, the three items sketch a standard for the next generation of scientific AI claims. First, say what the answer key is before the run starts, whether it is pre-measured chemical data or a formal proof checker that can verify each step. Second, report where the model gave up on a bad path, not only where it succeeded, because abandoning a dead end is a large part of what a scientist does. Third, keep the distinction between a claim discussed on a podcast, a result measured in a sandbox, and a result reported in a preprint, because each carries a different weight. The math episode also notes that P versus NP remains the hardest problem in complexity theory and is unlikely to be solved in the next year or two, which shows an answer key does not make every problem easy, only checkable. m:…8662 m:…1104 m:…7287