Finn Brooks: Clara, hi — okay, weird Tuesday. I'm reading Amodei's 'Machines of Loving Grace' essay on my couch at like midnight and I just — I had to text you.
Clara Bennett: Mm, I got that text. The all-caps one.
Finn Brooks: Because he published that in October 2024 predicting AI compresses 50 to 100 years of biological progress into 5 to 10 years — and now, a year and a half later, Anthropic actually launched Claude Science. An internal drug discovery program. June 30th. He's not just theorizing anymore.
Clara Bennett: Right, and the framing he keeps reaching for is pattern-matching — drug discovery is finding functional patterns in high-dimensional biochemical data. Which, in practice, is a reasonable description of one part of the process.
Finn Brooks: Wait, one part?
Clara Bennett: That's exactly where I want to start — because $60 billion invested, 175 programs in trials, zero FDA approvals. The pattern-matching works. The clinical gauntlet is a completely different problem.
Clara Bennett: Think of it like a really fast search engine for molecules. Instead of a chemist spending years testing combinations at a bench, the AI ranks billions of possibilities overnight and hands you the ten most promising ones. That's it. That's the acceleration.
Finn Brooks: Okay that actually clicks. Billions overnight.
Clara Bennett: And the two places where that's genuinely powerful — peptide drugs and something like CAR-T. With peptides, short amino acid chains, you can generate enormous libraries of sequence variants and rank them for binding affinity and stability in silico, before a single thing gets synthesized. No lab time burned on the obvious dead ends. With CAR-T — engineering a patient's T cells to target cancer — AI can iterate on receptor architecture designs in ways that would take, I mean, years of manual work.
Finn Brooks: Wait, so the Eli Lilly oral GLP-1 thing — that's the same move? Same search-engine logic?
Clara Bennett: That's actually the important distinction here. Eli Lilly isn't discovering something unknown — they're optimizing something that already worked. GLP-1 injectables existed. The computational problem is reformulating for oral delivery, which is a defined search space with clear success metrics. ADMET prediction fits perfectly — you're filtering absorption, toxicity, metabolism failures before expensive trials. That's signal. That's not hype.
Finn Brooks: No but — okay, China. Someone unveiled an AI system that designs personalized cancer vaccines in 24 hours. Like, per patient. That's not optimizing a known thing, that's — I don't even know what that is.
Clara Bennett: That benchmark is real and it's a striking number. Though — and I want to be honest here — the sourcing on how robust that system actually is in clinical terms is thin. The speed is genuine. Whether the output survives downstream validation, that we don't fully know yet.
Finn Brooks: Okay but — hang on — I want to name the bad take that's actually circulating right now. Because people are reading the Claude Science announcement and they're saying biotech renaissance, we're here, decades compressed. And that's just... that's not what happened. Faster molecular design means you get to Phase 1 trials faster. Which means you spend your billion dollars sooner. That's it.
Clara Bennett: That's the correction, yes. You arrive at the clinical gauntlet faster.
Finn Brooks: And the gauntlet is — what, 10 to 15 years minimum? Even if Isomorphic Labs hands you a perfect candidate tomorrow?
Clara Bennett: Historically, yes. And that timeline isn't bureaucratic friction — it's irreducible biological data collection. You need humans to respond, fail, recover, over time. No in silico model replaces that. Now, ADMET prediction does filter some late-stage failures earlier — catching toxicity problems before trials — but I want to be careful here, the evidence that it meaningfully bends the overall approval curve is... thin. I mean, it's promising, but I wouldn't assert it confidently.
Finn Brooks: Which makes the Deerfield number kind of insane to me.
Clara Bennett: The $1.6 billion autoimmune bet.
Finn Brooks: Deerfield Management, $1.6 billion on a single autoimmune asset — and that's the capital story and the approval story running on completely different clocks. Like, the money is flowing in as if the renaissance is already here. The drugs are not flowing out. That gap is — that's the whole thing, right? Amodei's 'Machines of Loving Grace' framing made discovery sound like the bottleneck. It was never the bottleneck.
Clara Bennett: And that's where Claude Science actually gets interesting to me — or maybe uncomfortable, I'm not sure which. Anthropic is an AI safety company. That's the mission. And now they're running an internal drug program targeting neglected diseases. Which is a genuinely altruistic framing — diseases that traditional pharma won't touch because the commercial incentive isn't there. But I keep turning it over and wondering... is that a coherent extension of what Anthropic is, or is it a market entry where Isomorphic Labs and others are already competing, and neglected diseases are just the lower-competition lane?
Finn Brooks: Wait, both things can be true though, right? Like, genuinely altruistic AND strategically convenient.
Clara Bennett: They can. And I'm not saying it's cynical. I'm saying — the real test isn't the framing at all. It's whether Claude Science actually produces a candidate that clears a Phase 2 trial. That's it. Because if it does, the neglected diseases angle stops being a question about identity and becomes evidence. If it doesn't — I mean, then we're left asking whether this whole renaissance narrative was always about capturing biotech's mindshare rather than actually compressing anything.
Finn Brooks: Yeah. And we genuinely don't know. Nobody does yet. That's the answer, I guess.
