Research
AI Minute Newsroom
2026-08-19
The reason AI has not cured cancer may be that almost none of its training data comes from human tissue
Vivodyne, a company that grows human tissue in robotic laboratories, put an argument to TechCrunch on 19 August that is worth taking seriously even if you discount the sales pitch. Drug discovery models are trained largely on static snapshots of cells and on animal experiments, and roughly nine out of ten drugs that work in animals fail once they reach human trials. The company's position is that the models are not short of parameters, they are short of causal data about human biology — records of what a human tissue does over time when you actually give it something. Its answer is a robotic system it calls HIVE, which grows twenty types of human tissue, doses and monitors them without human hands, and which the company says now runs at a scale it describes as roughly twice that of all US clinical trials combined. It reports predictive accuracy of 94 percent for liver toxicity, 96 percent for airway behaviour and complete agreement with known outcomes for twenty chemotherapy drugs tested on bone marrow. Eight large pharmaceutical companies have paid for early access. Vivodyne has raised close to $80 million. The accuracy figures are the company's own and have not been independently replicated.
Why it mattersNearly every promise made about AI and medicine assumes the bottleneck is intelligence. This is an argument that the bottleneck is measurement — that we have been fitting ever larger models to data drawn from mice and from cells in dishes, and then acting surprised when the predictions do not hold in people. It is a claim you can test rather than admire, which already puts it ahead of most of the genre. Treat the specific percentages as marketing until somebody outside the company reproduces them; treat the underlying point as the more important thing, because if it is right, the next real advance in AI medicine will come from laboratory robots rather than from a bigger model.
✓ Verified · 3 sources
Read in the app — free, in 9 languages
Related stories
OpenAI says its new models block nearly all reasoning-theft attacks.
2026-11-22Lean's own community says the AI-assisted Riemann proof is correct.
2026-11-22Researchers say an open AI model improved by editing its own weights.
2026-11-21A US watchdog found that no AI therapy chatbot has FDA authorization.
2026-11-21Outside mathematicians say the AI Riemann proof passed first checks.
2026-11-21