"All bio foundation models lead to target prioritization"

So Anthropic is doing drug design, and my feed is now full of jaded bio folks declaring what a terrible idea this is. "Drug design is hard" / "don't compete with your customers" / "if they were serious they'd be acquiring Cradle, Boltz" / etc I disagree I am SO EXCITED. 1/2

I generally like this person's takes, but there are simple tests in the theoretical biophysics literature to test whether a trajectory is Markovian. My strong suspicion is this cell state trajectory will fail those tests and this whole virtual cell argument will fall apart

Link to paper: journals.plos.org/plosone/ar…

This work was done in collaboration with Robert Shafer, Chiara Sabatti, Kaiming Tao, and other amazing researchers at @Stanford.

In summary, we found that finetuning was superior to RAG/prompting in a run of the mill biomedical research use case, if you have access to sufficiently capable/large LLMs. But even with small models, you can achieve remarkably good recall with clever prompting. Internally, we think of this behavior as “Thinking fast vs Thinking slow” – Advanced prompting allows even a small model to reduce false negatives at inference time (fast thinking), while finetuning allows you to improve performance overall by updating the model’s weights (slow thinking).

For this task, we found that RAG had similar performance as QSP (if not worse). A simple way to understand this result is that RAG replaces the full text article with a set of evidence snippets that may or may not contain the information (e.g. sequences, genes, drugs, etc) we are interested in. Since many articles do not contain the correct answer, RAG is more susceptible to false positives than base models with access to full text articles.

Finetuning (FT) on an expert-curated instruction set improved performance across the board for GPT-4o; but for the Llama3.1 models, finetuning did not improve recall. For Llama3.1-70B, finetuning improved accuracy, precision, and F1 reaching statistical significance, while for Llama3.1-8B, finetuning only improved precision. Interestingly, a question-specific prompting (QSP) approach, analogous to the one used by Cao et al Annals of Internal Medicine 2025 for systematic reviews, had a complementary effect. For every model, this approach improved recall without any finetuning, with increasing improvements as model size/capability decreased! Unfortunately, these complementary benefits of finetuning and prompting were not additive – FT QSP did not improved performance over FT.

To start, we recapitulate the performance improvement as model size/capability increases: GPT-4o > Llama3.1-70B > Llama3.1-8B. However, vanilla models generally found it harder to get good recall, i.e. they predicted more false negatives than false positives.

As tokenmaxxing frontier models becomes increasingly infeasible for regular research, exploiting abilities of smaller/open source models becomes an attractive option. We set out to benchmark LLM performance on a practical problem of extracting structured information from full-text biomedical research papers to populate the Stanford HIV Drug Resistance database.

Can finetuning LLMs outperform RAG or advanced prompting techniques for a niche biomedical knowledge-work use case? (Link to paper in comments)

This is such an amazing story and finally provides a clear path for personalized medicine 🖤

I asked codex to solve all my problems and it gave me the single cell expression levels of all protein coding genes in my body

Google did a prospective clinical study of their AMIE medical LLM chatbot in the clinic! They used AIME to conduct clinical history taking and present potential diagnoses for patients to discuss with their provider at urgent care appointments at Beth Israel Deaconess Medical Center. Blinded assessment of AMIE and primary care provider differential diagnosis (DDx) and management (Mx) plans suggested similar overall DDx and Mx plan quality, without significant differences for DDx (p = 0.6) and appropriateness and safety of Mx (p = 0.1 and 1.0, respectively). "PCPs outperformed AMIE in the practicality (p = 0.003) and cost effectiveness (p = 0.004) of Mx. "

Larger 2-state folders fold faster! journals.aps.org/prl/abstrac…

There's something very chilling about this coming from an authoritative figure like Demis. Why do a small group of people get to decide the balance of benefit vs harm for humanity?

“Even though you do everything right you don’t always get what you hope for” One of my favorite shows lately (and definitely one of the only live action shows I like) is The Pitt. In one of the episodes, Dennis Whitaker (4th year med student) loses a patient. He blames himself for his death - patient had unstable angina due to CAD masquerading as GI pain from gallstones The attending (Dr Robby) reassures him: "We did an EKG. We did a troponin. He had a HEART score of three" → "No doctor on the planet could have caught this." My doctor friends love this show because it provides a realistic version of the problems they deal with at work. Inspired by it, I created this game (pittgym.lovable.app) that uses transcripts from the show to simulate an emergency room - you pick a case from one of the episodes, then make decisions to inquire patient history, hypothesize differential diagnoses, order labs, prescriptions, perform procedures, and make a final disposition. It uses a senior attending (AI Dr Robby) that provides hints based on your state in the game, and a final debrief to help you understand how you did. Check it out!

1/🧵 We are very excited to release our new paper! From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence arxiv.org/abs/2601.03220 with amazing team @ShikaiQiu @yidingjiang @Pavel_Izmailov @zicokolter @andrewgwils

Finally had a chance to listen through this pod with Sutton, which was interesting and amusing. As background, Sutton's "The Bitter Lesson" has become a bit of biblical text in frontier LLM circles. Researchers routinely talk about and ask whether this or that approach or idea is sufficiently "bitter lesson pilled" (meaning arranged so that it benefits from added computation for free) as a proxy for whether it's going to work or worth even pursuing. The underlying assumption being that LLMs are of course highly "bitter lesson pilled" indeed, just look at LLM scaling laws where if you put compute on the x-axis, number go up and to the right. So it's amusing to see that Sutton, the author of the post, is not so sure that LLMs are "bitter lesson pilled" at all. They are trained on giant datasets of fundamentally human data, which is both 1) human generated and 2) finite. What do you do when you run out? How do you prevent a human bias? So there you have it, bitter lesson pilled LLM researchers taken down by the author of the bitter lesson - rough! In some sense, Dwarkesh (who represents the LLM researchers viewpoint in the pod) and Sutton are slightly speaking past each other because Sutton has a very different architecture in mind and LLMs break a lot of its principles. He calls himself a "classicist" and evokes the original concept of Alan Turing of building a "child machine" - a system capable of learning through experience by dynamically interacting with the world. There's no giant pretraining stage of imitating internet webpages. There's also no supervised finetuning, which he points out is absent in the animal kingdom (it's a subtle point but Sutton is right in the strong sense: animals may of course observe demonstrations, but their actions are not directly forced/"teleoperated" by other animals). Another important note he makes is that even if you just treat pretraining as an initialization of a prior before you finetune with reinforcement learning, Sutton sees the approach as tainted with human bias and fundamentally off course, a bit like when AlphaZero (which has never seen human games of Go) beats AlphaGo (which initializes from them). In Sutton's world view, all there is is an interaction with a world via reinforcement learning, where the reward functions are partially environment specific, but also intrinsically motivated, e.g. "fun", "curiosity", and related to the quality of the prediction in your world model. And the agent is always learning at test time by default, it's not trained once and then deployed thereafter. Overall, Sutton is a lot more interested in what we have common with the animal kingdom instead of what differentiates us. "If we understood a squirrel, we'd be almost done". As for my take... First, I should say that I think Sutton was a great guest for the pod and I like that the AI field maintains entropy of thought and that not everyone is exploiting the next local iteration LLMs. AI has gone through too many discrete transitions of the dominant approach to lose that. And I also think that his criticism of LLMs as not bitter lesson pilled is not inadequate. Frontier LLMs are now highly complex artifacts with a lot of humanness involved at all the stages - the foundation (the pretraining data) is all human text, the finetuning data is human and curated, the reinforcement learning environment mixture is tuned by human engineers. We do not in fact have an actual, single, clean, actually bitter lesson pilled, "turn the crank" algorithm that you could unleash upon the world and see it learn automatically from experience alone. Does such an algorithm even exist? Finding it would of course be a huge AI breakthrough. Two "example proofs" are commonly offered to argue that such a thing is possible. The first example is the success of AlphaZero learning to play Go completely from scratch with no human supervision whatsoever. But the game of Go is clearly such a simple, closed, environment that it's difficult to see the analogous formulation in the messiness of reality. I love Go, but algorithmically and categorically, it is essentially a harder version of tic tac toe. The second example is that of animals, like squirrels. And here, personally, I am also quite hesitant whether it's appropriate because animals arise by a very different computational process and via different constraints than what we have practically available to us in the industry. Animal brains are nowhere near the blank slate they appear to be at birth. First, a lot of what is commonly attributed to "learning" is imo a lot more "maturation". And second, even that which clearly is "learning" and not maturation is a lot more "finetuning" on top of something clearly powerful and preexisting. Example. A baby zebra is born and within a few dozen minutes it can run around the savannah and follow its mother. This is a highly complex sensory-motor task and there is no way in my mind that this is achieved from scratch, tabula rasa. The brains of animals and the billions of parameters within have a powerful initialization encoded in the ATCGs of their DNA, trained via the "outer loop" optimization in the course of evolution. If the baby zebra spasmed its muscles around at random as a reinforcement learning policy would have you do at initialization, it wouldn't get very far at all. Similarly, our AIs now also have neural networks with billions of parameters. These parameters need their own rich, high information density supervision signal. We are not going to re-run evolution. But we do have mountains of internet documents. Yes it is basically supervised learning that is ~absent in the animal kingdom. But it is a way to practically gather enough soft constraints over billions of parameters, to try to get to a point where you're not starting from scratch. TLDR: Pretraining is our crappy evolution. It is one candidate solution to the cold start problem, to be followed later by finetuning on tasks that look more correct, e.g. within the reinforcement learning framework, as state of the art frontier LLM labs now do pervasively. I still think it is worth to be inspired by animals. I think there are multiple powerful ideas that LLM agents are algorithmically missing that can still be adapted from animal intelligence. And I still think the bitter lesson is correct, but I see it more as something platonic to pursue, not necessarily to reach, in our real world and practically speaking. And I say both of these with double digit percent uncertainty and cheer the work of those who disagree, especially those a lot more ambitious bitter lesson wise. So that brings us to where we are. Stated plainly, today's frontier LLM research is not about building animals. It is about summoning ghosts. You can think of ghosts as a fundamentally different kind of point in the space of possible intelligences. They are muddled by humanity. Thoroughly engineered by it. They are these imperfect replicas, a kind of statistical distillation of humanity's documents with some sprinkle on top. They are not platonically bitter lesson pilled, but they are perhaps "practically" bitter lesson pilled, at least compared to a lot of what came before. It seems possibly to me that over time, we can further finetune our ghosts more and more in the direction of animals; That it's not so much a fundamental incompatibility but a matter of initialization in the intelligence space. But it's also quite possible that they diverge even further and end up permanently different, un-animal-like, but still incredibly helpful and properly world-altering. It's possible that ghosts:animals :: planes:birds. Anyway, in summary, overall and actionably, I think this pod is solid "real talk" from Sutton to the frontier LLM researchers, who might be gear shifted a little too much in the exploit mode. Probably we are still not sufficiently bitter lesson pilled and there is a very good chance of more powerful ideas and paradigms, other than exhaustive benchbuilding and benchmaxxing. And animals might be a good source of inspiration. Intrinsic motivation, fun, curiosity, empowerment, multi-agent self-play, culture. Use your imagination.

Lift your weights, eat your veggies, and build something beautiful