Answering Seb’s question: ‘Is AI for science underrated or overrated?’
Can I cheat by saying both?
Say we’re thinking about AI and protein structure prediction for drug discovery, for example. I’m quite excited about what that could make possible—it could help narrow down potential drug targets, improve our understanding of protein structure and function, and also give us a better sense of which drugs might fit particular protein structures. Protein design is also really exciting, including improving proteins or enzymes that are used as drugs or in industry.
But I’m often thinking about the bottlenecks too, and I feel like those are underrated right now. That includes the inputs — like the datasets that need to be collected to train models, who will actually gather them and how they’ll do it — and the outputs of AI-driven research. For example, validating predictions in clinical trials is still one of the biggest bottlenecks in medical research.
Some parts of that process could be sped up with AI, but actually running experiments, recruiting participants, securing science funding, navigating policy and regulation, and coordinating how scientists work with other people; all of that broader system that influences science will still be there and still often inefficient, I think. (This isn’t to say that AI can’t influence lab science or other parts of the process, though; it’s already used in improving DNA sequencing and microscopy, for example.)
But I liked Owlposting’s recent blog post on why AI didn’t replace pathologists, which makes a good case in point.
Answering Seb’s question: ‘Is AI for science underrated or overrated?’
Can I cheat by saying both?
Say we’re thinking about AI and protein structure prediction for drug discovery, for example. I’m quite excited about what that could make possible—it could help narrow down potential drug targets, improve our understanding of protein structure and function, and also give us a better sense of which drugs might fit particular protein structures. Protein design is also really exciting, including improving proteins or enzymes that are used as drugs or in industry.
But I’m often thinking about the bottlenecks too, and I feel like those are underrated right now. That includes the inputs — like the datasets that need to be collected to train models, who will actually gather them and how they’ll do it — and the outputs of AI-driven research. For example, validating predictions in clinical trials is still one of the biggest bottlenecks in medical research.
Some parts of that process could be sped up with AI, but actually running experiments, recruiting participants, securing science funding, navigating policy and regulation, and coordinating how scientists work with other people; all of that broader system that influences science will still be there and still often inefficient, I think. (This isn’t to say that AI can’t influence lab science or other parts of the process, though; it’s already used in improving DNA sequencing and microscopy, for example.)
But I liked Owlposting’s recent blog post on why AI didn’t replace pathologists, which makes a good case in point.