Alex Rives, head of science at Biohub, expects the recipe behind today's chatbots to carry over to biology, up to a point. He thinks pretraining, drawing out what a model has learned, and even reinforcement learning with verifiable rewards should all work for cells and proteins. He made the case on The Cognitive Revolution, in an episode published October 20, 2026. Where the two fields part is the source of truth. In math or code, a computer can check whether a model got the answer right. In biology, Rives said, the check has to happen in a laboratory.
Rives previously led ESM, a project that trains language models on protein sequences rather than on text. In the episode he spoke with host Nathan Labenz.
Labenz's question
Labenz framed the question in terms of how language models were built. For a time, he said, most of the computing went into "simple pre-training on data of like kind of unknown quality and relevance." Pretraining means letting a model learn the patterns in a huge pile of raw data, for example by predicting the next word. Then developers added a little instruction-following training at the end, and "all of a sudden you have like a pretty helpful assistant." Could biology work the same way, he asked: train on large amounts of raw, mediocre data, build up internal representations, and then draw out what you want at the end?
Principles that travel
Rives said he thinks the principles in biology "are going to be shared with all the other domains where we're seeing progress in AI." In his view biology is fortunate to come later, because researchers can see how those principles played out in other fields and pick out the parts that generalize.
The core principle he named is self-supervision, or representation learning: a model learns structure from unlabeled data instead of from answers people have written down. He traced it to the hope of deep learning pioneers Geoffrey Hinton, Yann LeCun and Yoshua Bengio, "the idea that you didn't have to know all the relevant variables." A neural network could learn them, along with the right settings, without anyone encoding them by hand. Rives said that idea has been proven out over decades, and he called scaling laws its clearest current expression. Scaling laws are the observed regularities in how a model improves as it gets more data, parameters and computing power.
He described two stages. First the model learns from statistical patterns and builds a model of the underlying system that produces them. Then comes elicitation, which brings out the part of that model relevant to a particular problem. Rives said this is already happening with proteins, and he expects it to happen for cellular biology too.
His own earlier work shows the split. In a 2021 paper, Rives and colleagues trained models on up to 250 million protein sequences. The task was to fill in amino acids that had been hidden, using the rest of the sequence as context. The researchers then tested whether the learned features could predict structural properties, such as which parts of a protein form helices or which amino acids sit in contact. Those structural labels were never part of the pretraining task. The features also supported predictions of how mutations affect a protein.
Reinforcement learning with a lab in the loop
Rives also expects one of the newest steps in the language-model recipe to reach biology: "it's going to be possible to have RLVR for biology," he said. RLVR stands for reinforcement learning with verifiable rewards. In the Tülu 3 report, Lambert and colleagues describe how it works for language models. The model is rewarded when a fixed, automatic checker confirms its answer, and that checker takes the place of a second model trained to judge responses. For math, the checker compares the model's final answer with a reference answer. For instruction following, it checks the output against explicit constraints. Because the check is an ordinary program, it can grade answers throughout training with no person or lab involved.
Rives wants more than a single check. He wants to build what he called "empirical AI": systems that ask a question, perform the experiment, learn from it, improve their understanding and propose a better experiment, then repeat the cycle so that they keep improving. He thinks all of that is possible in biology, but it will need to be built differently, with models that can accurately represent the many scales and the complexity of living systems.
The biggest challenge, he said, is that biology is empirical. RLVR has moved fastest in fields where answers can be verified quickly by computer, such as mathematics, coding and reasoning tasks. In biology, he said, researchers will have to build systems that allow "verification in the laboratory." In that loop, the reward would come from an experiment rather than from a program, and an experiment is far slower to run than a program that grades an answer.
One shared space, or separate tools?
Labenz then asked whether language and biology would eventually be trained together. He compared it to text and images. When he gives a model a picture along with instructions, he said, the model both sees the image and understands his request. That creates "a new kind of magic" that a language model passing work to a separate image generator through a tool call could not quite achieve. Would biology models stay at arm's length like that, or would biological encoders one day feed into the same reasoning space as language?
"I think that the truth is, I'm not sure," Rives said. He called language "a substrate for reasoning," and said it also lets AI tap into the model of biology that humans have built through centuries of empirical investigation, all of it recorded in text and learnable from it. The problem, he said, is that "the things that we most want to know about biology, they're not represented in text" and cannot be derived from first principles or from existing knowledge. They are empirical facts that only experiments can produce. In his account, that difference is what makes predictive models of biology necessary.
A latent space is the internal space where a model represents what it knows. On whether biology models will be fused with language models into one such space, Rives said he does not hold a strong view either way, and he can see advantages to fusion. It will be hard at first, he said, because fusing the spaces takes a lot of bridging data that links the two, and little of it exists now. Asked whether the field could get there eventually, he answered: "Probably."
Concepts that have no names yet
Rives closed this part of the conversation with proteins. Billions of proteins have been sequenced across life on Earth, he said, and scientists do not know the function of the vast majority of them, so there is no vocabulary for much of that biology. A protein language model trained on those sequences still develops internal features, including features that represent protein biology no one yet understands. Rives described them as "novel concepts that we don't have words and names for yet," and said a great deal can be learned just from the structure of those spaces.