Biomedicine is the pursuit of easing the suffering that biological aberration inflicts; and for many afflictions, drugs are the instrument through which medicine acts. The process by which such drugs are developed is technical, laborious, and capital-intensive—with the consequence that the firms which exist to produce them have organized themselves around three constraints: regulation, the criteria a compound must satisfy before it may be approved; the cost of experimental capital, the capital that a preclinical or clinical experiment requires; and knowledge work, the scientific and operational labor required to coordinate experimental capital to meet regulatory criteria. Every stage of a drug’s development, from target identification through compound synthesis to trial design, is accordingly driven, and gated, by decisions taken against the same three variables: whether the step would move the compound closer to regulatory approval, what it would consume in experimental capital, and what knowledge work must be allocated to it. The object of that knowledge work is singular throughout: to maximize the likelihood of approval and reduce the uncertainty associated with that estimate.
The development of AI promises to shift the cost of knowledge work to the cost of inference. Since the organization of a firm is the coordination of its constraints, the shift is not simply the same firm run more cheaply but a differently organized firm. A company building medicines need no longer organize itself around scarce knowledge work: we can now run a company whose development is driven by autonomous agents built to learn scientific and operational workflows. In doing so we lift not the largest bottleneck but the one whose removal is most consequential. Clinical trials appear to be the principal constraint when in fact they carry fundamental lower bounds that no technology can change: one cannot compress the time it takes for biology to unfold in a person. Intelligence, by contrast, has a lower bound that is essentially the cost of inference. Beyond cost, agents enable new modes of research and entire workflows that were previously cost-prohibitive or impossible. These changes in the organization of knowledge work reach the firm’s outputs by way of the other two constraints: a better allocation of experimental capital to the programs that merit it, and better strategies for meeting the requirements of regulation. We can develop better drugs and design better trials, all for less; in essence, we can execute the process at lower cost and with better results.
The most compelling areas to approach first are those that incumbent firms have historically overlooked but that the capabilities of AI have now made viable. Thousands of rare diseases have been passed over by pharmaceutical companies for years in favor of other indications with the majority of R&D dollars being concentrated on a small minority of diseases. The neglect is more striking because many rare diseases exhibit, at the level of clinical development, far lower development costs and potentially greater capital efficiency than their more competitive counterparts. The impediment has, until now, been structural: the cost of knowledge work for a disease program does not scale with the expected value of its market; that the patient population is smaller does not entail that the judgment required is less. Agentic drug development significantly reduces those costs, opening the long tail of diseases: a class of diseases where the biology is tractable but conventional program economics are unfavorable because development costs overwhelm expected market value.
Moreover, we focus on drug repurposing, an area where scientific knowledge work is most impactful. The headwinds generated by AI drug design, together with the increased supply of drugs from Chinese biotech markets, have driven down the price of preclinical development; and cheaper preclinical development has yielded a growing number of drugs. Consequently, their price has fallen to where a seed-stage startup can reasonably own a portfolio of clinical-stage assets. What remains is scientific and operational judgment: in which diseases to develop the drug, and how to run each trial. These are exactly the decisions agents can make at scale: searching the vast combinatorial space of thousands of drugs against thousands of diseases, they can identify repurposing opportunities, and then design the optimal trial.
We have already started building agents for these tasks. We currently have state-of-the-art performance in identifying drugs for repurposing and for predicting clinical trial success, and we intend to continue building agentic scientific systems that can discover new biomedical knowledge and increasingly automate the process of drug development.
Our ambition is ultimately to extend the economic frontier of medicine until market size ceases to determine which forms of suffering deserve a treatment. Superintelligence may not remove every constraint on drug development but it may remove one of the most arbitrary: the requirement that a disease afflict enough people to justify the cost of thinking about how to treat it. We mean to help move medicine toward a world in which no biologically-tractable disease goes untreated merely for want of a market.