Colossal Biosciences co-founder and CEO Ben Lamm has a line he returns to often: "the data set is more valuable than the model." He made the argument onstage at Moonshots Live 2026, in an episode of Moonshots with Peter Diamandis recorded on 25 September 2026 and published on 30 September. Peter Diamandis, the XPRIZE founder, answered with "you've made that point so many times." Colossal uses genetic engineering and AI in its work on species preservation and de-extinction. The conversation came days after Anthropic's own move into biology, so Lamm offered the view of someone who runs a lab. In his view, general-purpose AI already helps with biology's paperwork and language. It is starting to help design experiments. On its own, it cannot yet turn a raw genome into answers.
From term papers to middleware
Dave Blundin, founder and general partner of Link Ventures, framed the question as one about business models. Will general AI models do biology "out of the box," or do biology companies need to build their own?
Lamm said Colossal had early access to tools from both OpenAI and Anthropic. For a long time, though, the company used them mostly as what he called "thoughtful middleware." That meant software glue linking the scientists' lab notebooks to Jira, the project-tracking tool that is common in software companies. Getting scientists to work in Jira was, he joked, "harder than stem cell reprogramming at times." Using AI for these systems and reports meant Colossal did not have to hire large consultancies. He named Deloitte and Accenture.
The science was a different matter. Lamm said earlier language models were "really good at writing term papers." They were poor at the work Colossal needs, such as ancestral state reconstruction, which infers what an ancestor's genes probably looked like, and comparative genomics, which compares genomes across species. That is changing, he said. Models can now run simulated experiments. Companies such as Lila are building automation around laboratories. Lila describes a system that links a scientific reasoning model to computational tools and automated instruments, then uses measured results to improve its next hypotheses. Lamm said Colossal has been careful about how and when it invests in that category.
Biology's language problem
Lamm expects the next useful gains outside small-molecule drug discovery to come from what he calls "language problems." His example was research on mice. Different papers may give the same gene a different name, classify it differently or run the same experiment slightly differently. He said that if you tried to repeat published experiments exactly, somewhere between 40 and 60 percent would fail. That is his own estimate.
The gaps also hide useful work. Lamm said Colossal has sometimes found out afterwards that a peer-reviewed paper had already dealt with a problem it was working on. The team had missed it because the way they were looking for it did not surface it in the literature searches they used. In the short term, he said, AI will "bridge the communication layer" between groups that describe the same biology in different words.
After that, he expects AI's role in biology for the next three to five years to be running simulated designs across many possible experiments and "helping creatively come up with the next experiment." Wet-lab experiments with real cells and samples will still be needed to test the results.
Anthropic's announcement follows the same pattern. On 23 September 2026 the company said that about 950 Claude agents had searched DNA data for 21 hours. They identified a family of enzymes in bacteriophages, viruses that infect bacteria, which the company calls array-associated reverse transcriptases. The system pairs an enzyme with a neighbouring partner gene and a repeated stretch of DNA. That repeat layout is the reason for the comparison with CRISPR, the gene-editing system. Human scientists reviewed the candidates that survived and ran the laboratory experiments. Anthropic said most candidates were ruled out during follow-up, and the biological function of the system was still under investigation. Lamm said Ginkgo and others would now test the approach.
Why one genome is not enough
Blundin pushed back. Searching thousands of papers is "LLM city," he said, meaning exactly the kind of text work large language models handle well. But if the input is just a gene sequence, "if I dump that right now into Anthropic, pretty sure nothing good is going to come out."
Lamm agreed, then added: "but at scale." He does not expect anyone to feed a single genome into a frontier model and be told that six changes would make an animal immortal. What he thinks will work is comparison across very large numbers of genomes. Suppose thousands of bird genomes across different evolutionary branches belong to species that resist many diseases. Researchers could start from that known outcome and trace its genetic roots back down the tree of life.
He said one of Colossal's companies is already doing this on a small scale, with examples such as the tumour-suppressor gene P53 and the "immortal" jellyfish. He said "early indications are positive." He gave no results.
This is where Colossal's planned network of biovaults comes in. These are collections of biological samples that the company wants to set up with governments around the world. Lamm described doing T2T sequencing on the samples. "Telomere-to-telomere" sequencing reads a genome from one end of each chromosome to the other. According to the US National Human Genome Research Institute, longer sequencing reads and better computing closed repetitive regions that earlier human genome references had left unresolved. That revealed more than two million additional variants. Lamm also mentioned ATAC-seq. The original methods paper describes it as a way to map which parts of DNA are physically open. Those regions can include switches that control whether genes are active. In Lamm's view, this depth of data across many species is what would make large-scale comparison worthwhile.
Astromech's "inflection models"
Later, Diamandis asked about Astromech, which he described as a company Colossal had just spun out and "a multi-billion dollar company from the start." The valuation was his claim.
Lamm said today's foundation models, the large general-purpose models, do not look at the whole tree of life and will not "magically overnight" turn a genome into an answer. Astromech is trying to build what the team calls "inflection models" internally. These would capture how a DNA sequence evolved, when it evolved and why it did not evolve in related groups of species. They would also bring in outside factors, such as climate, that may have driven those changes. The aim is a prediction model of where evolution went and why.
Lamm was clear that he does not expect Astromech to build a frontier model that "solves all things all ways." Instead, it would plug into one. Once biovaults hold millions of samples, Astromech's data and its methods for classifying that data would act as the "traffic control cop" telling the larger model where to look. Astromech's website describes a "Large Life Model" still in development. It is meant to combine data across species with measurements of gene activity and ancestral reconstruction. The site lists Lamm and the geneticist George Church as founding members.
"The door's barely cracked open"
Near the end of the session, Diamandis passed on a question from Cathie Wood, who had asked it backstage. As he relayed it, Anthropic had announced wet labs able to create something CRISPR-like. Did that light a fire under Lamm's work, or kill its momentum? In Anthropic's own account, Claude identified a system that already occurs in bacteriophage DNA.
Lamm called the news "massively validating." Some people had concluded that "biotech's going to be dead" because the AI companies would do the work in their own labs. "That's just not true," he said. "The door's barely cracked open." To him, the moment showed that AI companies see biology as one of the most widely accepted uses of their technology, because people want healthier families and longer lives. He noted that Anthropic chief executive Dario Amodei has a background in biology. He called it "a great thing for this society."