In 2025 Hoffman co-founded Manas AI with Siddhartha Mukherjee, betting that artificial intelligence would transform cancer drug discovery. The bet was not eccentric. It was the logical extension of an argument Hoffman has been making for years: that biological systems are structured substrates that can be read, predicted, and rewritten with the same tools used for text and code. Biology, in this reading, is just another language — written in nucleic acids and proteins rather than in words, but tractable by the same methods of statistical learning that turned human languages into something machines could manipulate.
The intellectual move is striking. For most of the twentieth century, biology was treated as a fundamentally different kind of science than chemistry or physics — more historical, more contingent, less reducible to general principles. AlphaFold, Demis Hassabis's protein-folding system, was the first dramatic demonstration that this assumption could be inverted. A neural network, trained on existing protein structures, could predict the structures of new proteins at a level of accuracy that decades of biochemistry could not match. Hoffman watched this and drew a generalized conclusion: if proteins are tractable, so are pathways, so are cells, so are organisms.
The conclusion is not yet proven. Proteins are a relatively well-bounded domain. Cancer is not. Cancer involves the breakdown of regulatory systems whose normal operation is still imperfectly understood, the interaction of dozens of cell types in tumor microenvironments, and patient-specific variation that makes generalization hard. Manas AI is betting that modern AI methods are now powerful enough to compress what was previously decades of trial and error into months of in silico exploration followed by targeted experiments. The bet is high-variance. The payoff, if it works, is enormous.
The deeper significance of Manas AI, for Hoffman's broader thought, is that it instantiates the techno-humanist wager in a domain where outcomes are measurable. A drug either works or it does not. A patient either lives longer or does not. The argument about whether AI expands or contracts human agency gets cashed out, in oncology, in survival curves. This is a more honest test than the arguments about social media or productivity software, where the outcome variables are murkier and the counterfactuals harder to construct.
Hoffman's framing of biology-as-language has critics. Biologists point out that the language metaphor breaks down in important places — that organisms are not just running code, that the relationship between genotype and phenotype is mediated by environmental and developmental factors that resist clean computational treatment. Hoffman would accept this critique while pointing out that the language metaphor is not a claim about identity but a claim about tractability. You do not need biology to be language to apply language models to it. You need biology to be statistical enough that models trained on existing data can make useful predictions about new data. That weaker claim is what AlphaFold demonstrated, and it is what Manas AI is trying to extend.
The implication, if biology really is the next language, is that medicine becomes another instance of the broader argument about superagency. A drug discovery pipeline that took twenty years and a billion dollars could be compressed to a fraction of that. The economic implications are large; the human implications are larger. Diseases that have killed people for generations could move from incurable to chronic to absent. This is not a guaranteed future. It is the future Hoffman is putting capital and time into, on the bet that the techno-humanist compass, applied to medicine, points in a direction worth walking. The walk has started. The next decade will report on whether the compass was reading true.