For years, some of the biggest names in artificial intelligence have made extraordinary predictions about what AI could do for medicine. Faster drug discovery. Personalized treatments. Diseases solved in years instead of decades. And, inevitably, the promise that AI might someday help cure cancer.
There is just one uncomfortable problem: AI can only learn from the biological data we give it. And much of the data available today may not be good enough to predict what actually happens inside a living human body.
That is the argument being made by biotech startup Vivodyne, which believes the bottleneck holding back AI drug discovery may not be bigger models or more computing power. It may be the quality and type of biological data those models are trained on.
Vivodyne is developing automated robotic laboratories capable of growing and experimenting on human tissue. Its HIVE systems can reportedly grow 20 types of human tissue, expose them to drugs and other interventions, monitor the response, and generate large quantities of experimental data.
The idea sounds deceptively simple: before asking AI to revolutionize human medicine, perhaps we need to give it much better information about actual human biology.
The current AI boom has created the impression that enough data and enough computing power can solve almost anything. That logic has worked remarkably well in areas such as language, images, coding, and mathematics.
Human biology is different.
A model can learn patterns from enormous datasets, but identifying a correlation is not the same as understanding why one biological event causes another. Drug development ultimately depends on those causal relationships.
Much of today's biomedical AI training data comes from animal experiments, isolated cells, proteins, genomic information, scientific literature, and snapshots of biological systems. All of these sources are valuable. But none perfectly recreates the enormously complicated environment inside a human body.
Vivodyne CEO and co-founder Andrei Georgescu summarized the problem provocatively: without better human data, AI systems may become extremely good at curing diseases in mice without becoming equally good at curing them in humans.
That distinction matters more than the AI hype cycle sometimes acknowledges.
Perhaps the most sobering number in pharmaceutical development is the extraordinarily high failure rate of experimental drugs.
According to the figures cited by Vivodyne, roughly 90% of drugs that perform well enough in animal studies to enter human clinical trials ultimately fail to receive regulatory approval.
Think about what that means.
Scientists can identify a promising biological mechanism. Researchers can develop a drug around it. The treatment can perform well enough in preclinical testing to justify enormous additional investment. Companies can spend tens of millions of dollars moving it into clinical trials.
And then the human body says no.
That gap between preclinical success and clinical reality represents one of medicine's biggest unsolved problems. AI can potentially help researchers search through chemical and biological possibilities faster, but speed does not automatically solve the underlying problem.
If the data feeding the model does not adequately represent human biology, a more powerful model could simply generate incorrect predictions faster.
This is where Vivodyne's approach becomes particularly interesting.
Rather than focusing exclusively on developing another AI model, the company is building infrastructure designed to generate the biological data future models might need.
Vivodyne's robotic HIVE laboratories grow human tissue and automatically perform experiments on it. Drugs can be administered, biological responses measured, and the resulting information collected at a scale that would be difficult to reproduce manually.
The company says its tissue models have already shown significant predictive performance. Vivodyne reports that its liver models achieved 94% predictive accuracy against human toxicity results, while its airway tissue matched real human tissue behavior 96% of the time. Its bone marrow system reportedly achieved 100% concordance across tests involving 20 chemotherapy drugs.
Those are company-reported results and will need to be evaluated in the broader context of independent research and real-world pharmaceutical development. But they illustrate the larger strategy.
Vivodyne recently opened what it describes as the world's largest "human data center" near San Francisco. Instead of racks filled only with GPUs, think of this as biological infrastructure: automated experiments continuously producing information about how human tissues respond to different interventions.
That could become extraordinarily valuable in an AI-driven pharmaceutical industry.
One of the most important ideas behind Vivodyne's argument is the difference between observing a biological state and understanding how that state developed.
Imagine showing an AI millions of photographs of patients before and after developing a disease. The system might become excellent at recognizing the differences between those states.
But recognition does not necessarily mean the model understands the chain of biological events that moved the patient from one state to another.
Georgescu argues that existing datasets frequently give AI models static snapshots rather than enough information about biological cause and effect.
Vivodyne wants to generate something different.
Its systems can observe tissue before an intervention, administer a drug or stimulus, and then continuously measure what happens afterward. Instead of simply telling an AI that biological state A and biological state B exist, researchers can potentially show the model how an intervention transformed A into B.
That is much closer to the question medicine actually needs answered:
If we want a particular therapeutic outcome, what biological intervention is most likely to cause it?
For AI drug discovery, that could be far more valuable than simply adding another billion static biological data points.
The excitement surrounding AI in medicine is not baseless.
Google DeepMind's AlphaFold represented an enormous advance in predicting protein structures and demonstrated that machine learning could solve scientific problems that had frustrated researchers for decades.
But predicting biological structures and producing safe, effective medicines are very different challenges.
Even extraordinarily sophisticated AI systems must eventually confront the complexity of living organisms.
A drug can successfully bind to its intended target and still fail because of toxicity, unexpected interactions, inadequate absorption, immune responses, differences between patients, or countless other biological variables.
This is why AI drug discovery should not be judged solely by how quickly models generate potential molecules.
The real test is much harder:
Do more of those molecules eventually become medicines that improve outcomes in actual patients?
Until that happens consistently, claims that AI is about to "cure cancer" deserve scrutiny.
There is a broader lesson here for the AI industry.
We have become obsessed with models.
Bigger models. More parameters. Larger context windows. More GPUs. More agents. More compute.
But in healthcare, model intelligence is only one piece of the puzzle.
The quality of the underlying clinical and biological data may ultimately determine how useful these systems become.
A model trained on weak, incomplete, biased, or poorly representative medical data does not magically escape those limitations because more compute is thrown at it.
This is particularly important when moving from administrative healthcare AI into clinical decision-making and drug development. An AI system summarizing a medical note can tolerate a very different error profile than a system predicting whether a cancer therapy will work.
The closer AI gets to decisions that affect human biology, the more important data quality, validation, causality, and human evidence become.
None of this means AI's potential in medicine has been exaggerated beyond usefulness.
Quite the opposite.
AI may eventually become one of the most important technologies ever introduced into pharmaceutical research.
Models can already help scientists explore enormous chemical spaces, predict molecular interactions, identify potential drug targets, analyze scientific literature, generate candidate molecules, and prioritize experiments.
The opportunity becomes even more interesting when AI is connected to automated laboratories.
Imagine an AI proposing a hypothesis, an autonomous laboratory testing that hypothesis on human tissue, the results feeding back into the model, and the AI using that new information to design the next experiment.
That creates a continuous learning loop between artificial intelligence and experimental human biology.
If systems like Vivodyne's can operate at sufficient scale and accuracy, the future of AI drug discovery may look less like a chatbot magically inventing a cancer cure and more like an enormous automated scientific feedback system.
That future may be less dramatic than the headlines.
It may also be much more realistic.
As a physician and someone deeply interested in AI, this is where I think healthcare needs to become much more disciplined about the conversation.
We should absolutely be excited about what artificial intelligence can do for medicine.
But we should be equally willing to ask what the models actually know, where their knowledge came from, and whether their predictions survive contact with real patients.
Healthcare is full of situations where something looked promising in theory, worked beautifully in a laboratory, succeeded in animals, and then failed when tested in humans.
AI does not eliminate that reality.
If anything, AI's ability to generate predictions at extraordinary speed makes rigorous validation even more important.
The race to cure cancer may therefore depend on something less glamorous than the next giant foundation model.
It may depend on building better experiments, collecting better human biological data, establishing causality, and giving AI a much more accurate representation of how the human body actually behaves.
Because the question isn't whether AI can generate another promising cancer drug.
The question is whether that drug will work when a real human life depends on it.
