LLMs Can Discover Cures for Diseases
Many people argue that large language models (LLMs), in their current form, cannot discover cures for diseases like cancer because they are trained to predict tokens rather than understand or interact with the physical world. Or, rather, because they simply aren't intelligent enough. People have all kinds of different arguments for why LLMs can't cure diseases, and regardless of what those specific arguments are, they typically all underestimate what LLMs can accomplish with external tools and experimental feedback. More fundamentally, they misunderstand what scientific discovery involves.
Scientific discovery is not simply a matter of knowing something nobody else knows. It is a process of forming hypotheses, designing experiments, interpreting results, questioning earlier assumptions, and using new evidence to decide what to investigate next. Consider a scientist investigating a possible cancer treatment. They identify a compound that appears to kill malignant cells, then conduct experiments to determine its effectiveness, understand its mechanism, assess its effects on healthy cells, and investigate whether some cancer cells survive. Each result informs the next decision. An unexpected observation might reveal a previously unknown mechanism, while evidence of resistance might lead the scientist to investigate a different treatment strategy.
There is simply no inherent reason why these decisions must always be made by a human. An LLM can generate hypotheses, select experiments, run scientific software, interpret measurements, and use the results to determine what to investigate next. Through tool calls connected to laboratory instruments and robotics, it can also initiate and execute physical experiments. These capabilities do not require an entirely new type of AI model. Tool calling is already widely established, and laboratory automation provides the physical infrastructure through which an LLM can act on the world. The more important question is whether the model can make the decisions necessary to conduct a successful investigation, and I believe they can. An additional related misconception concerns the fact that an LLM's training data is fixed. People unfamiliar with the architecture of LLMs sometimes assume this means the model cannot acquire new knowledge after training. In reality, an LLM can propose an experiment whose outcome is unknown, receive measurements from the physical world, and incorporate those measurements into its subsequent reasoning. This is new empirical information, regardless of whether the model's parameters change. Provided that the information remains accessible through its context or an attached persistent memory, the model can use it to revise hypotheses and make decisions without additional training.
For example, suppose an LLM hypothesizes that an existing drug has an unrecognized effect on a particular cancer. It directs an experiment that reveals this effect for the first time, then investigates the underlying mechanism, tests alternative explanations, and uses the findings to guide further experiments. The discovery was not contained in its training data. It emerged from an investigation that the model directed, using evidence that did not exist when it was trained. Even an initially incorrect hypothesis can lead to a discovery if the model recognizes unexpected results and follows them toward a better explanation. I believe most of us already see this as common sense.
The argument that token prediction prevents meaningful scientific research is, similarly, mistaken. Human scientists communicate experimental instructions through written procedures, software commands, and directions to laboratory personnel. An LLM generates instructions that external systems can interpret and execute. Whether those instructions are represented as tokens has little bearing on their scientific value. The important thing is that the model can select appropriate actions, evaluate the resulting evidence, and make sound decisions about what to do next. Token generation is simply the mechanism through which the model produces its outputs. There is already experimental evidence that LLM-based systems can perform important parts of this process. Google's Co-Scientist generated drug-repurposing hypotheses for acute myeloid leukemia, including the compound KIRA6, which subsequently reduced the viability of several leukemia cell lines in laboratory experiments. Human experts selected the candidates for testing, but the hypotheses were generated by the AI system.
Another system, Robin, investigated potential treatments for dry age-related macular degeneration. It proposed experimental approaches, identified candidate drugs, analyzed results from experiments conducted by scientists, and recommended further investigations. Among its findings was that ripasudil, a glaucoma medication approved in Japan, increased phagocytosis in retinal pigment epithelium cells, a cellular function impaired in the disease. Robin also proposed follow-up experiments to investigate the mechanism behind this effect.
Separately, Coscientist, an unrelated system with a similar name to Google's, demonstrated that LLMs can plan chemical experiments and operate laboratory equipment through connected software. Together, these studies establish that several important parts of the process, from generating therapeutic hypotheses and interpreting new evidence to executing physical experiments, are already within the capabilities of LLM-based systems.
Of course, identifying a promising compound in cultured cells is not equivalent to curing a disease. A treatment must ultimately demonstrate efficacy, acceptable safety, and lasting clinical benefit in patients. Cancer presents additional difficulties because different cancers have different biological mechanisms, and even tumors of the same type can respond differently to treatment. An LLM directing this research must deal with the same biological complexity, experimental uncertainty, and demanding standards of evidence as a human scientist. These difficulties should not, however, be confused with inherent limitations of the LLM architecture. Discovering a treatment requires identifying promising approaches, testing them, recognizing failures, and progressively improving the evidence until a treatment can be established as effective and safe. There is no known computational principle that restricts these activities to human researchers. Whether a particular model can perform them reliably enough, especially across a prolonged investigation, is a separate question that must be answered through experimentation.
I also find it worth mentioning that, despite all of this, I do not believe that LLMs are necessarily the most efficient architecture for scientific research. As I argued in I Don't Think LLMs Are the Final Architecture of Superintelligence, autoregressive language models are increasingly being used for tasks that are not fundamentally linguistic. Although this approach has proven remarkably capable, representing and processing information through sequential token generation can introduce unnecessary computational costs, latency, and constraints on how information is handled. These inefficiencies become particularly apparent when an AI system interacts with the physical world. Precise robotic control often requires continuous feedback and rapid adjustments, making sequential language generation poorly suited to controlling every movement. Likewise, analyzing large biological datasets or complex microscopy images may be more efficiently handled by specialized software and models. But, an LLM directing an investigation does not need to perform every operation itself. It can delegate calculations, data processing, and low-level control to systems better suited to those tasks, much as human scientists rely on instruments, software, and other specialists.
There are also undeniable limitations in the reasoning capabilities of current models. A prolonged investigation might involve thousands of experiments, extensive datasets, conflicting observations, and decisions whose consequences only become apparent much later. Maintaining an accurate understanding of the accumulated evidence, identifying subtle inconsistencies, and planning effectively over long periods are difficult even with external memory and specialized tools. These limitations can affect whether an LLM-based system successfully completes a complex research program, and improving its reliability remains essential.
Future AI architectures are undoubtedly going to handle these problems much more efficiently, particularly as we develop alternatives to relying so heavily on sequential token generation for reasoning, planning, and interaction with the physical world. But, it should go without saying that being inefficient or imperfectly suited to a task does not make it fundamentally incapable of performing that task. The possibility of developing better scientific AI systems in the future does not diminish what LLMs can already accomplish.
Any and all of the challenges discussed above are going to be solved with time, you can't stop technological progress. One of the most exciting possibilities of advanced AI is its potential to help humanity eradicate some of the diseases that cause the greatest suffering, and we can either choose to be ignorant AI pessimists in the present (until proven wrong), or we can recognize its potential, confront challenges that remain, and promote or actively contribute to developing AI systems capable of solving some of humanity’s greatest problems.