Ben Okonkwo: Jonathan, hey — tell me something, did you see the Anthropic thing drop last week?
Jonathan Ingles: The antivenom story. Yeah, it landed in my inbox three times in one morning.
Ben Okonkwo: Right, and — okay, so my sister's a bench scientist, she texts me the headline, she's excited, and then I actually read the report and I'm... sitting there going, wait, these are not the same thing.
Jonathan Ingles: The headline and the footnotes aren't even in the same building.
Ben Okonkwo: That's exactly where we're going today. So — the core thing. On October 20th, Anthropic published a technical report: Claude generated 318,484 candidate protein designs targeting scorpion neurotoxins. Of those, 211 were actually synthesized and tested in a lab. Seventeen bound. And here's my analogy for it — imagine a master key-cutter who runs three hundred thousand key blanks through a computer and finds seventeen that slide into the lock. That's real. That happened. But nobody has checked yet whether any of the doors actually open.
Jonathan Ingles: And Anthropic's press release called it designing an antivenom.
Ben Okonkwo: The word antivenom is in the framing, yes — while the fine print explicitly says no designs were tested for toxin neutralization in vitro or in vivo. Not in cells. Not in animals. Results are against recombinant toxins only. So the question we're actually trying to work out: is this a milestone in what AI can do, or is it a milestone in what a press release can imply?
Jonathan Ingles: Those are different claims with completely different evidence bars. And the gap between them is the whole episode.
Ben Okonkwo: And what actually breaks that clean version — the pipeline itself. Because Claude didn't just run one model. It coordinated 12 specialized protein-design tools. No wet-lab feedback during the design phase, none. It handed off to Adaptyv Bio and Twist Bioscience cold. They synthesized, they tested, they got 35 design-toxin pairs hitting — 17 cleared Anthropic's binding bar. That's the actual sequence.
Jonathan Ingles: Purely computational hand-off to contract labs. No iteration.
Ben Okonkwo: Right — but the part that doesn't fit is what they didn't get. No multitoxin binder. They targeted five toxins, they cleared four. And Anthropic flags that limited backbone diversity may have constrained the search space. So the thing that would actually matter clinically — something broad-spectrum — wasn't there.
Jonathan Ingles: Okay, wait — recombinant toxins only. Lab-grown proteins. Not actual venom from a Centruroides noxius. That distinction is doing a lot of work that the framing just... buries.
Ben Okonkwo: It's doing enormous work. Cn2 — the specific neurotoxin from Centruroides noxius — two of the 17 binders hit it with particularly high affinity. That's the highlight result. But those binders have never met real venom, which has 200-plus other proteins that could cross-react or interfere.
Jonathan Ingles: And that's exactly where the outside reviewers land. Susana Vázquez Torres and Timothy P. Jenkins called it 'a promising first step.' That phrase is chosen carefully — because the therapeutic distance remains large.
Ben Okonkwo: That language is — I mean, 'promising first step' from domain experts is actually a specific signal. It's not dismissal. But it's also not endorsement of the framing. They're saying: if this approach can confirm binders on hard-to-target toxins in weeks, it could accelerate neglected disease work. That's different from saying Claude designed an antivenom.
Jonathan Ingles: The narrow claim is falsifiable, actually. Anthropic released all 211 designs and the lab data publicly. That's real. But no independent lab has replicated it yet. So you have a specific, checkable claim sitting there — and nobody's checked it.
Ben Okonkwo: Which is... that's the thing I keep returning to. The claim is credible enough to be replicable. It's just — nobody has yet. And until they do, 'promising first step' is the honest ceiling.
Jonathan Ingles: But that ceiling — 'promising first step' — only holds if binding was actually the hard part. And that's the thing Joy Karmakar at UCSF punched directly: it isn't. Publicly, on the record, she flagged that binding affinity against recombinant toxins stopped being the rate-limiting step around 2018.
Ben Okonkwo: So Anthropic solved a problem the field had already moved past.
Jonathan Ingles: That's the sharp version of it, yeah.
Ben Okonkwo: And the bottleneck that actually matters now is — okay, so it's specificity under biological complexity. Does the binder hold when the toxin is folded differently, or when it's swimming in actual venom alongside 200-plus other proteins that might cross-react? That's the experiment. Anthropic explicitly didn't run it. No neutralization data, no cell assay, no animal model. The entire result lives at computational screening to binding confirmation, and that's... I mean, that's a real thing, but it's stage one of a much longer sequence.
Jonathan Ingles: It's like — picture a first-year med student who aces every anatomy diagram but has never seen a patient. The diagrams are accurate. The gap is the whole point.
Ben Okonkwo: Right — but the part that actually stops me is who ran this. Citrini Research. A financial investment firm. Zero drug-discovery infrastructure. One Anthropic scientist directly involved. And Hunterbrook journalists came along for a two-day visit, used an unreleased Claude model without standard bioscience guardrails, and that's what produced the result that got framed as a breakthrough.
Jonathan Ingles: No peer review published alongside. Anthropic self-reported everything.
Ben Okonkwo: Which is — look, I don't want to collapse two separate questions. Because the Citrini thing is actually interesting on its own terms. If a finance firm with no lab coat, no structural biology PhD, can coordinate this in weeks, that tells you something real about how much of early-stage drug discovery has become coordination and computation rather than specialized bench work. Whether that's democratization or a mirage about what's actually happening in wet labs — that's the question we haven't touched yet.
Jonathan Ingles: And the compute cost hiding behind all of this makes that question considerably harder to answer.
Ben Okonkwo: Right, and that compute cost is — okay, so MelindaBChu1, the VC scout on X, she put an actual number on it: roughly 10,000 H100-hours. That's not an abstraction. Most university research labs cannot run that. Like, physically cannot afford it.
Jonathan Ingles: That's the number the press release buries. And it flips the democratization framing entirely.
Ben Okonkwo: It might. But — wait, actually I want to stress-test that, because the 8% hit rate, 17 out of 211, that looks competitive if you're comparing to blind de novo runs. The August 2026 Anthropic study showed 22 to 35% on known targets, field norm is 10 to 15. So — I mean, there's a version where the compute is doing real work. The problem is we don't know the denominator. What would a matched human expert team cost to do the same run?
Jonathan Ingles: Anthropic didn't run that comparison. That's not an oversight — that's a choice.
Ben Okonkwo: No matched cohort, yeah. So 'beats human experts' — that framing from the August study gets attached to this one without the evidence to support it here.
Jonathan Ingles: And frankly — 318,484 candidates to 17 confirmed binders is a 0.005% hit rate at the compute layer. The field's 10 to 15% is at the synthesis-confirmation layer. Those aren't the same denominator. You can't compare them. The numbers sound related and they're not.
Ben Okonkwo: Which means — hm. If Anthropic solved the wrong bottleneck, the one Joy Karmakar said moved past binding in 2018, then the 10,000 H100-hours isn't buying you progress toward a drug. It's buying you a very expensive binding result.
Jonathan Ingles: That's the counterintuitive core. Citrini, a finance shop, getting access to unreleased Claude without standard guardrails — that's not democratization. The compute wall means the moat is widening. Only who's inside the moat has changed.
Ben Okonkwo: And that's — I keep turning over what Anthropic's own report says at the end. They frame this as a step toward antivenom for neglected diseases. Which is a claim that requires leaping over neutralization, in vivo work, toxicity profiling, clinical trials. Multiple stages, none of them demonstrated. The data is public, all 211 designs, the full lab results from Adaptyv Bio and Twist Bioscience — so the narrow binding claim is actually falsifiable. Any lab could check it. Nobody has yet.
Jonathan Ingles: And if they do — if an independent lab runs it and the binders don't neutralize — we won't say AI failed. We'll say Anthropic announced progress at the wrong milestone.
Ben Okonkwo: Which is a quieter and more uncomfortable conclusion than either side wants.
Jonathan Ingles: Your sister texted you a headline about an antivenom. The footnotes say: seventeen molecules that slide into a lock. Nobody's tried the door.
Ben Okonkwo: Yeah. And I still don't know what to tell her. Which is — honestly, that's probably the right place to stop.
