AI Isn't Training on Cause and Effect — And That's Why It Can't Cure Cancer Yet

"Absent human testing, what are these [AI] models going to do?" asks Andrei Georgescu, CEO and co-founder of the biotech startup Vivodyne. "They're going to cure cancer in mice." A cure in mice is a real result that rarely survives contact with a human patient; TechCrunch puts the industry-wide gap at "90% of drugs that are effective enough in animal testing to enter clinical trials don't receive regulatory approval for humans". Georgescu's argument: the models aimed at drug discovery inherit that gap rather than closing it, for reasons sitting in their training data, not their architecture.

The models learn states, not transitions

His complaint, in his words: "All the training is done on static snapshots of these cells", so "the model learns 'this is cell state A,' 'this is cell state B,' but never 'cell state B is the effect of inflaming cell state A.'"

Strip the biology out and most engineers have met this. A recommender trained on click logs learns that shoppers who saw a banner bought more often. That is a fact about the log, not the banner: it went to shoppers the ranker already scored as close to buying. Turn it on for everyone and the lift never arrives. Observational data answers what tends to occur together; acting requires knowing what happens if you change it. The two answers come apart when something unrecorded moves both — normal in biology, common in production.

He is not alone. A Nature Methods study that pretrained 400 models on 22.2 million cells across 6,400 experiments found "single-cell foundation models show no clear data scaling laws." More cells did not buy more capability — not even on the paper's perturbation-response task, which asks a model to predict how cells react to an intervention. That measures the ceiling without confirming Georgescu's explanation of it.

Why this reaches past medicine

A model that only describes can be graded against held-out data from the world it trained on. A model that acts cannot, because acting changes that world. Ship a pricing model and you alter what people buy; ship an agent that files tickets and you alter which tickets exist. The corpus was recorded while the model was not intervening; every deployment bets that correlation now behaves as a lever. Drug discovery is the expensive extreme: a clinical trial is a costly way to learn it did not.

Combination therapy is the multi-step planning problem

The next step for medicine has a shape agent developers will recognise — the parallel is ours, not Georgescu's. "If we want combination therapies, the space that has to be searched explodes—it can't be an experimental approach", Georgescu told TechCrunch.

The arithmetic behind that is ours, not his. Screen candidates one at a time and the work grows with the shortlist. Screen every pair and it grows with the square of it; every triple, with the cube. A hundred candidates means a hundred single-drug experiments, roughly five thousand pairs, and over a hundred and sixty thousand triples. Nobody runs that in a wet lab.

An agent choosing a sequence of tool calls meets the same shape: k-step sequences grow exponentially in k, so enumeration dies around the depth where multi-step agents get interesting. What replaces brute force is a model accurate enough to predict an action's effect before the action — Georgescu's own framing of the goal: "You have to say, 'I want this effect to happen, so what cause should I invoke?' Establishing causality in human biology is the basis of all of this."

What Vivodyne says it will take

Its answer is to manufacture the missing data, not mine more of it. Its HIVE machines are modular robotic labs that grow 20 human tissue types and then dose and monitor them autonomously, running "hundreds of thousands of ongoing experiments" exposing diseased tissue to a stimulus and recording the response. It spun out of Penn in 2021, has raised "just under $80 million across two rounds led by Khosla Ventures", and last week opened a facility outside San Francisco its own site calls "the world's highest capacity human data center."

Its fidelity figures are self-reported, and belong quoted separately rather than averaged — they cover different tissues and endpoints. Vivodyne says its liver cells reach "94% predictive accuracy compared to human trials that test for toxicity", that its airway tissue "matches the behavior of real human tissue 96% of the time", and that its bone marrow has achieved "100% concordance in tests of 20 different chemotherapy drugs". None of it is peer-reviewed here and its pharmaceutical partners are unnamed. Georgescu's call for "a sanity check" is aimed at the field's existing models, not at these figures.

Georgescu is not the only sceptic. TechCrunch characterises a weekend post by Anthropic's Dario Amodei as calling AI-cures-cancer talk more cliche than credible, quoting him at "the thing that will work is actually curing cancer" — TechCrunch's characterisation, unverified here. The scoreboard supports it: AlphaFold won a Nobel and has yet to produce a drug, while Isomorphic Labs, set up to carry it forward, now expects trials it had slated for 2025 to begin by year's end.

Whether robotic tissue labs get there is unsettled. The narrower point survives either way: a model trained only on descriptions of states has not been taught which of them it can cause. Any product built on one is quietly assuming otherwise.

Sources