A Scaling Paradigm for AI in Life Science

There is a growing conviction today that scaling training-time compute with larger datasets, bigger models, and more compute will not produce the kind of artificial scientific intelligence where you achieve in science the groundbreaking capabilities you have today in code.

What made code and math scale was a tight, cheap, repeatable verification loop, and science resists this in both of its modes: theory can't be verified until hindsight separates the real breakthroughs from dead ends, while applied science can't be replayed the way a codebase can due to its resource heavy and time consuming nature.

Science seems like the natural next avenue for scaling foundation models, but it has a completely different nature, governing mechanics, and scaling laws compared to software engineering or mathematics.

The bet today is on scaling pre-training with domain data, scaling reinforcement learning across environments, model distillation, continual learning, and rethinking inference-time compute and architecture. All these buckets are being molded to embody the mental model that scientific intelligence requires.

Life sciences, and particularly human therapeutics, is the most lucrative field of science which the frontier AI community is racing to build capabilities in. There are two ways AI can play a role here. First, creating models or agents for biologists and chemists. Although this can make a good business, it's a classical SaaS play, and by the SaaS physical boundaries it conforms to by nature, it doesn't bring game changing value to the model providers. This is primarily the beta play in our discussion.

The alpha, and arguably the reason why the AI community is investing massively in life sciences, resides in assets. Developing drugs is an extremely risky and complicated process, but a single successful drug can throw off billions in annual revenue.

For a model provider that builds powerful models, owns the data flywheel, and compounds intelligence, discovering even a handful of drugs on that scale is a different order of upside than selling tokens, the kind of upside that dwarfs the entire SaaS business it might otherwise settle for.

01The Verification Conundrum

Today, reaching this vision is still bottlenecked by the complexity of verification, a problem derived from inherent architecture limitations of today's models. Drug development resists verification on three fronts at once: it's expensive, so you can't brute-force it with volume; it runs in many years, not seconds; and it's not replayable, you can't rerun the same program under the same conditions to isolate what caused what change.

Today's technology is incapable of reaching a state where it can predict the outcome of a drug program end-to-end. This is because of sample inefficiency and the inability to serve at such an extremely complicated and long-horizon. Reaching this state means taking a drug program, modeling it end-to-end, investigating what caused its success or failure, and investigating every other factor that could have produced a different outcome. This includes modeling the entire space of drugs and targets, the formulation, the ADMET properties, PK, product attributes, the FDA process, the regulatory environment, the factory, the economic situation, the political situation, stakeholder priorities, and many others. Each one of these factors is its own world model and it affects the rest. To arrive at a general intelligence model that predicts the program, you would have to run this investigation for every single program out there, which is impossible to do even once, and then create billions of new programs to train on.

If we narrow down the process into reasonable practice, the binding constraints remain the physical, chemical, biological and regulatory factors that could determine whether a molecule survives; the broader context is real but secondary to the core verification problem.

The AI in science community is therefore developing predictive and generative tools that solve different parts of the process. This includes analyzing drug-target interactions, predicting drug properties, modeling how it behaves in the body, how it's manufactured, or predicting whether it will pass regulatory screening.

The fundamental question to ask ourselves here is how close are we to real world verification? How do we verify a prediction made today, when the real world feedback will come after 3 years, and how do we isolate the actual effect from the broader context that is shaping the real-world outcome?

Because full end-to-end verification remains out of reach on the timescales that matter, the workable path is to create many short, probeable checkpoints whose outputs can be composed into longer decisions.

02Navigating the fuzziness by being close to the verifier

The question is what role should AI play in the drug development process. What agents need to look like, where should humans steer them, and how should agents work with input from the real world. On the scientific side, we should understand what is realistic given the physical limitations of LLMs and today's models, how can deep learning models be leveraged, and what kind of training and verification mechanisms do we need to balance short-term rewards with long horizon goals.

The problem should be approached from a practical angle, derived backward from the shape of the problem: agents deployed at silos where we could be very close to verification. This modular approach respects the idiosyncratic nature of each of the process fragments, and the way they behave in the system. The philosophy is to batch long-horizon thinking into multiple connected short-horizon intelligence buckets with a transcendent ontology that connects them in a closed loop system. In each bucket, we can reasonably define a sub-system, predict and generate over nested input, then produce refined and curated nested output we probe for verification.

The ontology functions as the composition layer: each verified short-horizon output becomes structured context that the next bucket conditions on, so sequential decisions accumulate without requiring a single unverifiable leap across years.

The goal is to build a contextual embodiment of the drug system without having to model the whole process and limit the thread of context derivation to checkpoints where you can probe for verification.

This is done in three ways:

• BuildImbuing sequential context into predictions through chains that extend down to the molecular mechanisms.
• ValidateProbing low-confidence links for real-world verification.
• InferRelative rather than absolute ranking in a well defined optimization design space.

These three moves keep the reasoning thread short enough to verify while still carrying sequential context forward, so the lag between prediction and real-world outcome no longer blocks progress.

03The New Scaling Paradigm

Realizing this vision requires frontier breakthroughs in AI on the following layers: representation, execution, learning, and improvement.

04The Tyranny of Coupling

Drug companies are wasting billions of dollars every year on failed programs. A drug's fate is set by the interactions it lives inside: the physics within the molecule, and the physics between the molecule and its environment. These two scales are interconnected, and the coupling is what kills programs. You push potency and pay for it in colloidal stability. You fix stability and the viscosity moves on you. You solve viscosity and bioavailability collapses in a trial two years later, and by then the program has spent millions on its way toward death. A solution that attacks any one of these from a narrow angle can't see the others, because the failure was never in the property it was modeling. It was in the coupling it left out.

That is why attacking any single property in isolation fails, and why the short-horizon, relative, closed-loop approach described above is required.

Automation is also a wrong angle to approach this topic from, because it supposes you can make development more efficient by fitting an AI seam into the current workflows and artefacts. But fragmented AI deployment across functions only accelerates the individual steps. You still hit a dead end, because the endpoints get absorbed into the same fuzziness and the underlying bottleneck never moves. Executing faster in siloes doesn't touch it, and as we argued above, layering one supermodel over the process doesn't either. Both fail for the same reason: the problem has a very long context length, coupled factors resolving over a horizon where the task is unverifiable and not replayable. That is the shape LLMs are furthest from, and may be physically incapable of serving.

So the approach has to be derived backward from that endpoint, not assembled forward from siloes that don't connect, or from scaling laws borrowed from a domain where they don't belong. What is required is a system that moves dynamically with the problem, pursuing relative convergence on targets while verifying the right task at the right time. What that gives the drug developer is control over outcomes. They see where the risk lives, they act on it in time, and they reach past the boundary of the problem as they first defined it to create the breakthrough drugs that were never in view.