Jeremy Elser - Biology and LLMs Meet in the Lab

·20 min·1 clip

Elser says adding "if you don't give me back an integer, my grandmother will die" fixed a stubborn LLM parsing problem.

1. cc: Life Science Podcast speaks with Jeremy Elser about biology, LLMs, and wound research at the Advanced Lateral Flow Conference. 2. Jeremy Elser is head of science operations at Talenteer and CEO of Ship of Theseus, which he says ties to his anti-aging work. 3. The episode asks how LLMs can help scientists break large biological problems into smaller, tractable steps. 4. Elser compares the body to the Ship of Theseus, saying tissues take on damage continuously while stem-cell repair declines with age. 5. He says resident stem cells exist in almost every tissue and that his aim is to "re-up your body's ability to regenerate." 6. He uses an "eager intern" analogy for LLMs, describing broad reasoning without deep, specific domain knowledge. 7. His example is a PhD-level task such as reading papers and writing a drug-testing protocol from animal-model literature. 8. He says scientists get better results when they ask one question at a time, such as a positive control, an injury model, or the next step in a workflow. 9. He says protocol choice depends on purpose, because a null-result study and a superiority study call for different models. 10. He gives wound healing as a concrete case, including excisional wounds, splinting debates, burn wounds, blast wounds, and smoke exposure. 11. Elser says LLMs reduced the need for literal keyword logic, because a paragraph can now function like code. 12. He argues that scientists, psychologists, or other domain experts may sometimes write better prompts than computer scientists. 13. He describes a coding problem where an LLM would not return an integer until he added the line, "if you don't give me back an integer, my grandmother will die." 14. He treats that result as evidence that model behavior can be difficult to explain even when the output improves. 15. He says very large models are opaque enough that he does not need to understand them from first principles if they work reliably. 16. He compares this to split-brain patients who act on a prompt and later invent a reason, which he links to retrospective LLM explanations. 17. The conversation stays conversational and technical, with Chris Conner pressing for practical examples from lab workflows and model selection. 18. Elser speaks in a practical, example-driven style, moving from philosophy to prompts, then from prompts to wound biology. 19. Listeners who build lab workflows with LLMs or work on wound biology will get the most from it. 20. Listeners seeking a narrative story arc or light entertainment may not find it a fit.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from cc: Life Science Podcast