The thing that I noticed while I was doing my degree was that the CS people refused to talk to the bio people, and the bio people refused to talk to the CS people. The CS people would go on and make solutions to problems nobody asked, and the bio people would try to solve data problems in complicated ways that made no sense. Overlapping expertise is and was very rare between those two fields. AI however CAN be that bridge, because it can basically turn your average bioinformatician/biologist into a really competent programmer, without them having to give up time away from the work they are actually skilled at.
Maybe I'm just completely AI-psychotic, but I've always been a researcher at heart. I hated doing literature review because it was so frustrating falling through dead ends. These last few months with agents I have been able to go deeper and find better leads than ever before. AI is like having an insanely smart lab partner, available to you 24/7/365. Never gets tired, never has a pissy day, never gets too hungover to read papers. I do legitimately think it's going to accelerate the field so much faster than people expect. I don't know about the 10 year horizon, but I'd love to see it happen, and I'm cheering for them.
When I hear Amodei et al talk about biology, it's obvious to me that there's an intuition and understanding gap here. There's a famous study in synbio that led to standardization called, Reproducibility of Fluorescent Expression from Engineered Biological Constructs in E. coli
https://journals.plos.org/plosone/article?id=10.1371/journal...
88 labs measured "fluorescence from three engineered constitutive constructs in E. coli." You should eyeball the results yourself first,
https://journals.plos.org/plosone/article/figure/image?size=...
That's the result these machines spit out for the same set of samples being measured across 88 labs. This was termed a "remarkable degree of precision" (for biology) of 1.54x sd(!).
Teams couldn't converge on instrument-to-instrument variation within the SAME lab, https://journals.plos.org/plosone/article/figure/image?size=...
So the reading is different depending on when you measure it. For some reason, people in silicon valley think that biology is like classical logic and computing. You write some genetic code and you get expression out. It's not.
That's the pop-sci view of a real process. In reality, viruses and bacteria are extremely fickle, and hard to grow. A lot of the synthetic biology results aren't easily reproducible even if you know the protocol.
So you're telling me that you can fix that with "just" AI and "robots?"
Well... I would hate to break it to folks. But applying neural networks and ML to biology have been a thing since the dawn of AI. It was one of the very first applicaitons people thought about.
It used to be called computer-aided drug design, then they went all in on expert systems, then there was a time when quantitative structure–activity relationship AIs were a thing, then there was teh deep learning phase... D.E. Shaw actually did a ton of work on it, https://www.deshawresearch.com/technology.html he used his money to buy a supercomputer and used that to train neural networks...
There have been blockbuster drugs that have come out of this. Tirzepatide / mounjaro was discovered by algorithmically iterating through sister-compound families of the peptide they'd found in Gila monster venom until they found just the right one.
The issue that Anthropic and Open AI face here is that there's no low hanging fruit. There's no free lunch. It's very unlikely that Claude or GPT can do something that hasn't already been tried at some point in every AI/ML research era.
So they won't be able to "just" discover a new drug. If that was possible, then it would have already been iterated into by sheer accident based on the number of projects that have tried some variation of this with some sister architecture over the many decades preceding us.
Now where these machines do help is that they help people get started, and they also help working scientists. It's boring stuff, but think stuff like, "hey my machine isn't working, it's made by such and such. Can you write code to help me save my data?" or, "Can you proof read this grant?"
It will even accelerate the already tried approaches as it allows a smaller team to try more AI/ML algorithms than they would necessarily have had the expertise to execute before.
But will Claude magically solve medicine? No. I'm sorry, but it just doesn't work that way. And I'm not sure why they believe this, but I keep repeating this over and over again... A lot of this AI and biology stuff is nonsense. Standard undergrad and post-grad stuff whipped into fervor and fluff.