Who Gets to Do Science When Science Gets Expensive

There's a version of this story where I'm supposed to be impressed. Anthropic launches Claude Science, a workbench that hooks into PubMed and Jupyter and R and cluster terminals and NVIDIA's BioNeMo models, and every figure you generate comes out the other side with the code that made it, a plain-language description of the process, the whole message history sitting right there like a receipt. There's even a reviewer agent that checks your citations and flags errors before you publish. And I'm sitting here thinking: yeah, okay, but who's "you" in that sentence?

Because I remember what it felt like to be three days from a grant deadline, running the same figure six times because the pipeline kept choking on a dependency that hadn't been updated since the Obama administration, eating cereal for dinner because the journal subscription ate the grocery money. I remember what it felt like to be the lab that was already behind. And when I look at a tool that promises reproducibility and auditability and speed, the first thing I want to know isn't whether it works. It's who gets to use it.

The reproducibility piece is genuinely good. I want to say that clearly before I get into the rest of it, because I've watched too many replication crises unfold in slow motion, watched papers fall apart because the figure-making process lived inside one postdoc's head and that postdoc graduated and moved to industry and took the undocumented shell script with them. An auditable chain from raw data to published figure is not a small thing. It's the kind of thing that should have been standard twenty years ago, and the fact that it wasn't is an indictment of how we've organized scientific publishing, not a compliment to any particular tool for finally getting there. But here's where the good news starts to curdle a little: reproducibility that only certain labs can access isn't reproducibility as a scientific value, it's reproducibility as a competitive feature. It's a selling point. And those are different things, even when they look the same from the outside.

The reviewer agent is the part that really gets me, though, not because it's a bad idea (it isn't) but because of what it's being called. Checking citations and flagging calculation errors is proofreading. It's useful proofreading, the kind you want after you've been awake for thirty hours and you're no longer capable of seeing your own mistakes, but it is not peer review and calling it a reviewer agent is doing a lot of rhetorical work in a very small space. Peer review is adversarial in the good sense. It's someone who doesn't share your assumptions, doesn't share your institutional pressures, doesn't share your career incentives, reading your work and asking the questions you didn't want to ask yourself. A model running on the same architecture that generated the paper, checking the paper's internal consistency, is not that. It's closer to asking your own brain to audit your own brain, which, as anyone with ADHD can tell you, goes great.

And then there's NVIDIA. NVIDIA doesn't build scientific infrastructure because they want better science. They build scientific infrastructure because they want more GPUs in more data centers running more inference workloads, and better science is the story they tell while that happens. Which is fine, honestly, I'm not expecting a semiconductor company to be a nonprofit, but it matters that we name the incentive clearly when we're evaluating what gets built and what doesn't, what gets funded and what doesn't, whose questions get asked and whose don't. The underlying architecture here is still a next-token predictor with a very good coat of paint and a very compelling product roadmap. The business model is "sell more hardware." That's not a conspiracy, it's just a business, but it shapes what the tool optimizes for, and what the tool optimizes for shapes what science gets done with it.

Here's the thing I can't stop turning over: what happens when publishing with an auditable chain becomes the norm fast enough that publishing without one starts to look suspicious? Not because anyone mandated it, but just because the field moves, because the well-funded labs adopt it first and their output velocity increases and the journals start to notice and the grant reviewers start to notice and suddenly the labs that can't afford the subscription are not just slower, they're legible as slower in a way they weren't before. That's not a hypothetical, that's how every previous generation of expensive scientific infrastructure worked. Sequencing, cryo-EM, mass spec, the tools that transformed their fields also transformed who got to do the transforming, and the gap between the labs that could afford them and the labs that couldn't didn't close on its own. It closed, when it closed at all, through policy fights and funding mandates and a lot of people making noise about access that the field would have preferred to ignore.

So the question I keep coming back to is the one I started with: who does this serve, really? Not who does the press release say it serves, not who does the product roadmap say it serves, but who actually benefits when this becomes the way science gets done? And the honest answer is that I don't know yet, which is not the same as saying I'm optimistic. The tools we build shape the questions we ask. The questions we ask shape what we know. What we know shapes everything that comes after it, including who gets to be part of the "we" in that sentence at all. That's not a small thing to get wrong. That's not the kind of mistake you patch in a later release. You can build a tool that makes science faster and more auditable and more impressive in a demo and still build a thing that makes the actual pursuit of knowledge narrower and more concentrated and more dependent on who can afford the subscription, and those two things can be true at the same time, and usually are.

I want better science. I want the kind of science where the figure doesn't live in a dead postdoc's shell script, where the calculation error gets caught before it propagates into ten downstream papers, where the researcher in the underfunded lab at the regional university can compete on the quality of their ideas instead of the size of their compute budget. I want that badly enough to say clearly that a tool which moves us toward some of those things while moving us away from others is not automatically progress. It's a choice. And the choice is still being made, right now, by the people building these things and the people funding them and the people deciding whether to mandate access or let the market sort it out (the market will not sort it out). The future of what we know depends on who gets to ask the questions. That's not a metaphor. That's just how knowledge works.