Dwarkesh Patel: uncertainty

Transcript context

…Absolutely, we actually do both today. We can train these models where the inputs are a notion of what that cell looked like at the starting place, here's what a generic old cell looked like, and then representations of the transcription factors themselves. We derive those from protein foundation models. They're language models trained on protein sequences. It turns out that gives you a really good base level understanding of biology. The model's starting from a pretty smart place. Then you can predict a number of different targets from some learned embedding, the same way you could have multiple heads on a language model. One of those for us is actually just predicting every gene the cell is expressing. Can I just recapitulate the entire state and guess what effect these transcription factors will have on every given gene? You can think about that as an objective rather than a value judgment on the cell. I'm not asking whether or not I want this particular transcriptome. I'm just asking what it will look like. Then we also have something more like value judgments. I believe that that transcriptome looks like a younger cell. I'm going to select on that and train ahead to predict it where I can denoise across genes and then select for younger cells. But you could do that for arbitrary numbers of additional heads. What are some other states you might want? Do I want to polarize T cells to a less inflammatory state in somebody with an autoimmune disease? Do I want to make liver cells more functional in a patient who's suffering from certain types of metabolic syndrome, maybe even orthogonal to the way that they age? Do I want to go in and change the way a neuron is functioning to a different state to treat a particular type of neurodegenerative disease? These are all questions you can ask. They're not the ones we're going after, but that is the more general, broader vision. This is so similar to, in LLMs, you first have imitation learning with pre-training that builds a general-purpose representation of the world. Then you do RL about a particular objective in math or coding or whatever that you care about. You are describing an extremely similar procedure where first you just learn to predict perturbations in genes to broad effects on the cell. That's the pre-training, just learning how cells work. Then there's another afterward layer of these value judgments of, “How would we have to perturb it to have effect X?” That actually seems very similar to “How do we get the base model to answer this math problem or answer this coding problem? ” I don't know if people usually put it this way, but it actually just seems extremely similar. That makes me more optimistic on this. LLMs work and RL works. Yeah, they do. I think the conceptual analogy is very apt. We don't actually use RL at the moment, so I don't want to overstate the level of sophistication we've got. But I think the general problem reduces down in a similar way. You can think about your earlier question of what does the general model look like that enables you to actually have compounding returns in drug discovery. You might have something like this base model, which as you said, just predicts this object function of, “How are these perturbations hitting these targets going to change which genes are turned on and off in this cell?” Then there's an entirely other task, which is, well, which genes do you want to turn on and off? What state do I want the cell to adopt? Our lens on that is that across many different diseases people have, age is one of the strongest predictors of how they're going to progress, whether that disease arises. In many, many circumstances you have evidence in humans where you can say, “Ah, if I could make the cell younger, maybe that's not a perfect fix, but that's going to dramatically benefit not only patients who have a diagnosed disease, but it might actually help most of us stay healthier longer, even subclinically before anyone would formally say that we're sick.” Now that's another more general function. The same way that in LLMs, you might have to create these particular RLHF environments, you need to have places where you can state a value function of the particular task that you're trying to optimize for. In drug discovery, you would then need to know, “Well, what are the cell states I want to engineer for?” That's kind of the next generation of what a target might be. Beyond just which genes do I want to move up and down, and which gene perturbations do I put in, you then need to know what cell state am I engineering for? What do I want this T cell to do?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.