factual

Cell modeling mirrors LLM pretraining then RL fine-tuning

NewLimit's modeling structure parallels LLM training: a base model predicts how TF perturbations change every gene the cell expresses (analogous to pretraining a general world representation), built on protein foundation models, then separate prediction 'heads' impose value judgments like 'this transcriptome looks younger' (analogous to RL on a specific objective), and additional heads could target arbitrary desired cell states.

factualpending

Speaker

Jacob Kimmel

Evidence Quote

“you could have multiple heads on a language model. One of those for us is actually just predicting every gene the cell is expressing... Then we also have something more like value judgments.”

Source

Evolution designed us to die fast; we can change that — Jacob Kimmel— Dwarkesh Patel Podcast
Created: 6/13/2026, 3:28:32 AM

Your notes

Loading notes...