I write about scientific AI, computational biology, single-cell and multi-omics research, and the challenges of using increasingly capable AI systems in scientific work. Many of these notes begin as LinkedIn posts and are archived here in a more permanent form.
Frontier models already contain enormous amounts of scientific knowledge. The harder problem may increasingly be deciding which knowledge is relevant, what deserves attention, and what should be checked next.
In a PBMC annotation experiment, AI systems had the biological knowledge needed to identify suspicious labels, but often failed to retrieve and apply it until explicitly prompted.
Simple models are not only baselines. Their successes and failures can reveal which biological structure or information a more complex model truly needs to capture.
As AI accelerates scientific generation, human verification can become the limiting step. The challenge is not only keeping humans in the loop, but deciding where human judgment must remain essential.
AI may already possess the relevant biological knowledge. The harder problem is often deciding what matters, what looks unusual, and what should be checked next.
An algorithm can produce a technically valid and visually convincing result while the biological interpretation built on top of it is still unsupported.
AI performs especially well in coding partly because code has cheap and immediate verifiers. Scientific reasoning often lacks equally fast and reliable feedback.
Scientific AI should be evaluated not only by whether an agent completes a task, but by whether assumptions, interpretations, and reasoning remain defensible throughout the workflow.