November 21, 2026
Heard in AI

Five investment researchers used Claude to design scorpion binders

Anthropic and Citrini Research say five investment researchers, three with no science background, used an internal Claude model to design anti-toxin proteins: 17 of 211 tested designs met the binding bar, but neutralization is untested.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Moonshots with Peter Diamandis, episode published October 23, 2026 (recorded October 23, 2026)

Five people at an investment research firm decided to try making a drug. Three of them had no background in science, and the other two, in the firm's words, "weren't exactly veteran drug developers." Three months later, Anthropic and the firm, Citrini Research, each published an account of what came out of it: 17 small proteins, designed by an internal version of Claude, that latch onto scorpion neurotoxins in laboratory tests.

Peter Diamandis, the founder of XPRIZE and Singularity University, put that project at the head of the biotech segment on the October 23, 2026 episode of his podcast Moonshots. He framed the block with one claim: four stories, one message, "we're learning to program matter itself." He credited a post by Andrew Coté arguing that people are not prepared for how smoothly biology will become nanotechnology and programmable matter, interoperable with the machine world, and pointed out that every cell is already full of molecular machines that build, repair and move things. His supporting example was Jeff Holden's Atomic Machines, which emerged from six years in stealth with what it calls a matter compiler: a fabrication system that the company says is directed by software, so that the difference between one device and the next is a change in code rather than new masks or a retooled production line.

An old problem with no money in it

Scorpions sting at least a million people a year and kill about 3,000, mostly children in places with limited medical care, Diamandis said. Today's antivenoms are made from animals such as horses, which makes them slow to produce and expensive, and each one works only against certain species. Because there is little money in the area, he said, almost nobody develops new ones.

What Claude actually did

Working with Citrini, Anthropic asked its internal Claude model to design binders: small proteins whose shape lets them grab a toxin and hold on to it, in the hope of a future antivenom. According to Anthropic's write-up, published on October 20, 2026, the model ran for 64 hours with four follow-up prompts and generated 318,484 designs. Diamandis said on air that Claude then picked the best 211; the study's own account is that Claude selected 180 and the researchers added 31. Those 211 were split by ambition: 60 aimed at a single toxin, 30 at two toxins from the beta family, 30 at three from the alpha family, and 91 at all five at once.

Seventeen met the bar, and the bar was narrow. Binding was measured by surface plasmon resonance, an instrument method that detects whether two molecules stick together, and a design counted only if it showed an affinity of 1 micromolar or tighter in every replicate and did not stick to an unrelated control protein. Confirmation generally had to come from two commercial laboratories, Adaptyv Bio and Twist Bioscience, with LqhIII assessed only at Adaptyv. Seven of the qualifying designs bound Cn2, four bound AaH II, four bound AaH I and two bound LqhIII; none bound Ts1. Two of the Cn2 binders reached single-digit nanomolar affinity, more than a hundred times tighter than the cutoff, and competed with a neutralizing antibody for the same toxin.

The limits that came with those results are just as specific. Every one of the 17 qualified against a single toxin only: none of the designs aimed at more than one toxin met the criteria for more than one. Eleven of the 60 single-toxin designs qualified, an 18% rate that Anthropic calls typical for protein design campaigns, so six of the 151 multi-toxin designs did clear the bar, but each against a single toxin — including exactly one of the 91 designs aimed at all five, and only against AaH I. That matters because an antivenom that is useful in a clinic has to cover the species that actually sting people. The experiments used recombinant toxins, made in the laboratory rather than drawn from scorpions, and neither cellular neutralization nor protection in animals was tested. So "17 worked" means 17 designs met a binding standard, not that an antivenom exists. Diamandis's other claim holds up: the designs and the data were released publicly.

A $200,000 budget

Citrini's own account, also published on October 20, 2026, supplies the part that makes the project repeatable. Five investment researchers spent three months on it. After running into biosecurity restrictions, they worked with Anthropic's life-sciences team, using an internal model on Anthropic's premises. They budgeted $200,000 and reported $49,997 in compute against nearly $150,000 of laboratory work. Software coordinated the specialist design tools, the correspondence with vendors and the preparation of candidates; synthesis and the assays then took about two months. Their results matched Anthropic's. Citrini also states plainly that it remains an investment research firm and not a biotechnology company.

The panel's reading

Alexander Wissner-Gross, a computer scientist and the founder of Reified, started with hyperbole and corrected himself inside two sentences: "Neurotoxins are cooked. It's over. It's almost over, I should say." The broader trend, he argued, is not new. Molecular biology has been "cooking, present continuous tense" since at least AlphaFold 3, the protein-structure prediction system, "when arguably structural biology was solved almost overnight."

What he treated as strategically new was Anthropic opening its own wet lab and partnering outside, together with a stated strategy of bulk-solving problems he struggled to label. "It's not quite correct to call these rare, and it's not a disease," he said, before settling on the company's own words, "underfunded or underappreciated." The asymmetry he pointed to is financial: "It's rather difficult, my understanding is, to put together, scrape money for, say, antivenoms against scorpion or other neurotoxins." It is not as lucrative as the GLP-1 drugs used for diabetes and weight loss, "but nonetheless, it's a very real problem." His conclusion was that this is an early case study of what a frontier laboratory can do to "bulk solve entire classes of problems in medicine physiology." Diamandis made it personal: what matters is where you point superintelligence, and if your child is injured or your family has an issue, you point it there.

The twist the panel kept returning to

Wissner-Gross then drew out the collaborator's name, telling the others to wait for it. The partner, he said, was an industry analyst for the broader IT industry, a firm that his understanding was that Dylan Patel's SemiAnalysis had acquired, and that had set up a therapeutics outfit and decided to go and solve neurotoxins. "That this is unheard of," he said. "Like, you don't just, you don't have, like, some random industry analyst decide, oh, we're going to, we're going to solve the problem of scorpion neurotoxins." He hesitated over what to call it and landed on "casual grand challenge problem solving from an industry analyst. And we should ask Dylan about it."

Dave Blundin, founder and general partner of Link Ventures, asked the podcast team on air to research the story and write it up, and drew the general lesson he said he had seen repeatedly: "if you get inside the building at Anthropic or OpenAI or Google, no matter who you are and what you know, you come out successful. You just have to get in there." He called it the ultimate case study of IT consultants going off to solve biology, enabled by AI.

Wissner-Gross read the line from Citrini's site about five people, three with no background whatsoever in science, seeing if a new drug was possible, and finished: "And they go and they solve scorpion neurotoxins. That's the age that we live in now." Citrini's write-up is more guarded about exactly that point, since no binder has been tested for neutralizing a toxin and the firm describes itself as still doing investment research.

Diamandis took the story as an instruction to listeners. "Whatever you thought you could do, blow the ceiling off it. Go ten times bigger," he said, addressing entrepreneurs and kids. "What is a problem? Find a great problem and solve it." The discussion ended on a method for finding one: scan Bill Gross's list of problems, or have a portable AI memory scour the list and return the 20 entries that best fit who you are, then go after those.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03

Connected ideas and articles

From the conversation

Podcast episodes

Moonshots with Peter Diamandis

SpaceX’s AI Business Up 830%, OpenAI Agents Lie, and Startup Turns Mercury Into Gold | EP #305

Episode published (recorded )This article draws on 1:38:01–1:43:50 and 1:43:52–1:44:56 (approximate times)

Article history

Updates to this article

Tags