Plan A and Plan B in Physical Capabilities
world models vs data plays; a pattern across physical AI capabilities, such as biology and robotics
23.09.26
As everybody knows, "data is the fossil fuel of AI". We've been warned it's going to run out for years and yet we keep miraculously finding more.
However, even though we have quickly found/hired/made trillions of tokens of coding RL environments, knowledge work and computer use, capabilities on the horizon that promise incredible benefits, such as biology and robotics, seem harder. In particular, there's a conversation between two leading approaches to robotics that I think might analogise well to biology.
Plan A: Just Go And Collect The Data
The approach to coding that ended up working was just going and getting the data; buying up all the dead companies' codebases, and igniting an industry of RL vendors that sell coding environments to teach specific capabilities. In pretraining, you similarly just go and get the data, buying up old books and hard drives.
In robotics, there is no internet pretraining. There is, at least from the academic literature, thousands of hours of data for simple tasks like picking things up with Kuka arms. For humanoid data on real tasks, you're gonna have to strap some cameras onto some people.
Indeed, this is what companies have been doing; the leading Silicon Valley robotics teams such as pi, Sunday and 1X all appear to have massively scaled their data collection efforts; it looks like (from Sunday's hiring post and Tesla's hiring post) the going rate for labelled humanoid data in the US is about $15-30/hour. Let's say the industry currently has on the order of 100k hours, and it needs on the order of 100 million hours to create a robust generalist model, which is napkin math based on the tokens/sec of video-action tokenizers like FAST, and the 10T tokens required to train a GPT-4 scale language model (this might be a totally incorrect analogy!). Unfortunately there is no easy scaling here; every extra hour requires hiring one extra human and taping a GoPro to their head. So this dataset costs on the order of $1 billion - not out of the question, considering the capital market for this kind of thing, and Mercor does report currently having $2B in ARR - but that's just to get to the first general capabilities that might be economically useful. The first economically useful LLM, GPT-4, cost about $100 million, and now can be obtained for under $5 million. This seems like a pretty big slog for just the first step in that process.
To be clear - there's many reasons why this might work. We are developing better and more sample-efficient ML techniques all the time, and as RSI kicks in in 2027/28, potentially we will be able to get general humanoid policies out of orders of magnitude less data (we only need one or two!). But, ceteris paribus, it seems like it won't work out of the box.
Plan B: Train a World Model
Ok, so we don't have embodied data that can be used to train a forward policy, but maybe we can train a model to produce that instead. This is the thrust of world model research: we have a lot of general videos taken from the real world, and if we create a model that can produce extremely realistic videos of the real world, it'll need to internalise physics and once it has, we can train robot models inside of that model rather than having to go and get data ourselves.
This has higher technical risk, since it’s not been proven to work yet, but it seems to be much, much cheaper if it does. And indeed, most frontier teams appear to be taking both of these bets simultaneously, in a whatever-gets-us-a-generalist-policy-fastest strategy. NVIDIA Cosmos, JEPA/AMI, and World Labs are the leading world model-only teams I’ve heard of.
Plan A and Plan B in Biology
In robotics, we’re trying to get a generalist policy, to use to sell labor. In biology, we’re trying to get a model that causally understands the human body, to use to make drugs (among other things!).
In bio, there’s even less data than in robotics, and it’s even harder to come by. To get the correct next token for the sentence “Is tirzepatide an effective drug for obesity? Assistant: “ costs about $1.1B, and about eight years. In fact, there’s only about 5,000 such tokens in the world. If we want to RL a model to be good at designing new drugs, we’re going to need to wait a very long time for the first gradient update!
I’m being a bit facetious here, and obviously there’s more than one bit of information for each clinical trial, but if we want more bits of information, that’s where we’ll have to get them. “Just going and getting the data” here seems extremely difficult, although teams such as Isomorphic and others do appear to be trying.
If running a clinical trial is too slow, what if we trained a model to predict the output of a clinical trial, and trained on that! The world model equivalent for AI is a physiological world model, or virtual tissue, or virtual cell. The level of abstraction necessary is not entirely clear - biology has shown itself not to be easily compressible, but not to be completely chaotic either. It is a very high-dimensional latent system, but it seems, in theory, predictable.
Unlike robotics, in this case it seems much more clear-cut to me that the world model is the only plausible approach. In this case, we just need to collect as much data as possible of all kinds, in vitro out of the wet lab (for cells) and in vivo out of humans, to train our biology models on. Unlike typical science, the exact question the data answers is not important; the important part is that it captures the latent properties of the system, perturbationally. I expect this to be the dominant story of AI and biology over the next few years.