The Mismatch
Earlier today I fed an AI my entire research library. 6,576 papers. I asked it a question I have been chewing on for years.
Do the official vocabularies meant to encode my field actually match how practitioners think?
I study zoonotic spillover. How diseases jump from animals to humans. Like most fields we have formal ontologies, curated catalogs of concepts meant to keep us organized. I suspected they were thin.
So I stopped arguing and tested it.
The system read 490 of those papers. Just on spillover and emergence. It pulled causal claims from the text, assembled a vocabulary from the bottom-up, and compared it to the official standard.
The gap was huge.
Of the 915 relationships repeatedly used in the literature, 864 had no counterpart in the reference schema. Twelve hundred conceptual categories appeared in my small sample but nowhere in the formal maps. Mostly clustered around environmental drivers and ecology. The actual working language of a fraction of my field is four times richer than the official one.
The experiment cost $26 and change.
A Harness, Not a Model
This was a test drive of Claude Science. Anthropic launched it today. Their goal? To do for the lab bench what Claude Code did for software developers.
It is an ambitious bet. Six months ago Zubair Jandali claimed AI could help with digital tasks in life sciences. Today the pitch shifted. It can run the work.
Here is the catch. Claude Science isn’t a new model.
Anthropic is blunt about it. It runs on Claude Opus 4. Same engine everyone else rents. No secret weights. No special access. The intelligence is identical.
What changes is the harness.
In AI terms, the harness is the scaffolding. It turns a general model into a tool. Data connections, code execution, memory, safety checks. A naked model might reason about a protein. A harnessed one pulls structure data, folds a variant on a GPU cluster, renders the image, and logs every step for reproducibility.
Claude Science is a substantial harness.
- It connects to over sixty scientific databases.
- Ships with pre-built skills for genomics, chemistry, and structural biology.
- Manages jobs across your laptop or rented cloud GPUs.
- Every output carries its full provenance, the code, environment, and conversation history bundled for later regeneration.
Does this make the model smarter?
No. It makes it useful. Which is far more valuable.
A computational biologist could have built half this herself using Claude Code and GitHub over a few weeks. But that assumes every lab has the bandwidth to wire their own tool from scratch. Anthropic is betting that curation beats duplication. Turning raw capability into reliable science.
Pharma First, Everything Else Later
The demo focused on drug discovery.
From a single sentence prompt, Claude planned a campaign to stabilize an enzyme behind phenylketinuria. It screened 2,200 molecules across 80 GPUs. Narrowed them to four candidates. Produced a go/no-go memo.
Then it ran the same process on 100 rare diseases simultaneously.
Why stop there, the presenter asked. Could do 10,000.
Impressive? Yes.
Molecularly limited? Absolutely.
Every database and partner model points toward pharmaceutical science. Genes, proteins, small structures. Where the money is. OpenAI and Google are aiming at the same target.
But the rest of science is wide open.
Earth sciences. Atmospheric data. Ecology. Social sciences. Much of epidemiology.
None of it is configured in the launch. The data these fields rely on, biodiversity records from GBIF, climate reanalysis, census data, remote sensing, are absent from the initial sixty databases. The “dry lab” work, field ecology and population dynamics, remains an untapped frontier.
It is easy to visualize, though.
Imagine a harness like this pointed at zoonosis. It pulls species occurrence records. Overlays climate reanalysis. Fits a distribution model. Flags counties where a tick-borne pathgen is likely to spread next season.
Drafts the surveillance brief. Generates the figure.
Compresses weeks of assembly into an afternoon for a public health department. The intelligence exists to do this. What’s missing is the wiring. The connectors. My $26 experiment is a small proof of concept.
Judgment Over Generation
The technology is fast. It will accelerate everything.
But it clarifies where human scientists matter.
Once generation becomes cheap, judgment becomes the bottleneck. Auditing, validating, correcting. These are the rate-limiting steps for any work moving through a screen.
That is just good science anyway.
Claude Science ships with a “reviewer” agent to flag bad citations and mismatched numbers. For now, it’s the same model checking its own output, not an independent oracle. But the direction feels right.
The real risk is regression toward mediocrity.
Models trained on existing literature excel at reflecting that literature. They talk to themselves.
But the same tool that maps consensus can also map gaps. Relationships nobody has tested. Concepts used everywhere but defined nowhere.
I spent $26 finding the edges of what is already written. Finding what hasn’t been imagined yet? That’s the harder problem.
It also looks like the only one left to solve.
