Thirteen designs versus a few hundred thousand

From a stage at a private summit, the CEO of an AI "science factory" made a claim with no paper behind it: his company let software design, build, and test in vivo CAR-T candidates in the lab autonomously, churning through somewhere between 200,000 and 300,000 variants — the figure shifts depending on who is telling it — in under six months for a few million dollars. CAR-T is a way of reprogramming a patient's own immune cells to hunt cancer; doing it "in vivo," with an mRNA-and-lipid payload injected straight into the body, is the field's hot frontier. Lila says it produced the best compositions ever shown in monkeys. There is no preprint, no release, no independent look — just the talk.

Thirteen versus a few hundred thousand, a few million dollars versus two billion — the gap is the whole pitch.

The comparison doing the work is the throughput gap. Lila frames its hundreds of thousands of design rounds against Capstan Therapeutics, a startup whose in vivo CAR-T program AbbVie bought for up to $2.1 billion last summer, and which Lila's telling says reached its lead candidate on roughly thirteen designs. That contrast — thirteen versus six figures, a few million dollars versus two billion — is the entire pitch: that an automated design loop can search a space orders of magnitude wider than a human team for a rounding error of the cost.

It is worth knowing where the comparison frays. Capstan wasn't beaten in any market; it was acquired, and its lead asset treats autoimmune disease, not cancer — so the head-to-head sets an unpublished monkey result beside a clinical-stage drug in a different indication. The claim surfaced at an investor's own summit, relayed through that investor's newsletter, with no peer to check it. What's verifiable is only the price tag on the rival and the size of the number Lila is throwing around. If the underlying result ever clears a journal, the question that decides whether the pitch holds won't be whether AI can design a CAR-T — the field is crowded — but whether brute-force search at that scale finds something a small, careful team never would.