Claude Designs Proteins Now, and I Have Complicated Feelings About That

Saw the Anthropic protein design news this week and sat with it longer than I expected to. Claude apparently designed disease-targeting proteins autonomously, with actual wet-lab validation, and hit something like a 35% success rate against a human average of 10-15%. Not a benchmark. Not a demo. Real proteins, in a real lab, doing real things. That’s a different category of claim.

My first reaction was the one I imagine a lot of people in adjacent technical fields had: cautious excitement, immediately followed by the reflex to check for the catch. I spend my working life around systems that look impressive in a demo and fall over the moment someone points them at anything messy. So the scientist-adjacent part of my brain went looking for the asterisk. There’s always an asterisk. Someone in the discussion thread made the point well: PR from AI labs about breakthroughs has a track record, and not always a good one, of being technically true in a way that only specialists can properly unpack. Fair scepticism. I’d want outside labs replicating this before getting too far ahead of myself.

But there’s a second reaction sitting right next to the sceptical one, and it doesn’t cancel it out. It just sits there, uncomfortable, refusing to resolve. Because buried in that same discussion was a comment from someone with stage 4 inflammatory breast cancer, on her eighth line of treatment, saying she might have just missed the wave of treatments that could have saved her. She wasn’t being dramatic. She was doing the maths on her own life against the pace of medical progress, and the maths wasn’t landing in her favour. That’s not an abstraction. That’s someone’s actual Tuesday.

I don’t know how to hold both of those things at once, the professional scepticism and the visceral hope, but I don’t think I’m supposed to resolve them. They’re both correct. AI hype is genuinely exhausting and has burned plenty of people who got excited about things that turned out to be nothing (LK-99, anyone?). And also: if there’s a real chance this kind of tool shortens the distance between “diagnosed” and “treated” for people who don’t have decades to wait for the normal pace of science, that matters more than my instinct to be the guy who’s seen too many overhyped launches.

I think about this differently now that I’ve got a teenager in the house who’s growing up assuming AI tools are just part of how things get done, the way I assumed calculators were. She doesn’t carry my scar tissue from twenty years of watching tech promises evaporate. There’s something in that. Not naivety, just a different starting point.

The bit of the discussion that stuck with me longest wasn’t the science at all. It was someone in a university protein design workshop saying they’re rewriting tomorrow’s session to basically tell students: you might not need to know this the old way anymore, the model can do it. I’ve felt an echo of that in my own field. Watching a skill you built over years get partially automated is a strange, specific kind of grief, even when the outcome is good. Nobody’s written a great word for that feeling yet.

What I keep coming back to is that the actual value of something like this was never going to be decided in a Reddit thread, or by me either. It gets decided by whether independent labs can reproduce it, whether it holds up outside a controlled demo, and whether the gains actually reach the person with stage 4 cancer rather than getting swallowed by patent lawyers and drug pricing that has nothing to do with the science. That second part worries me more than the AI does, honestly. We’ve had cures and treatments before that existed and still didn’t reach people because of who could afford them. A faster drug discovery pipeline sitting on top of the same broken access model doesn’t fix much.

So I’m not popping any champagne yet. But I’m not rolling my eyes either, which is unusual for me. If even some of this replicates, it’s worth being glad about, cautiously, with one eye still open.