Today's generative LLM model fundamentally cannot generate new breakthroughs like that. The tech simply doesn't allow it, for one simple reason: LLMs not only don't consider, but are fundamentally incapable of considering, the science of their output before creating said output. They only consider the likelihood that each word they generate should follow the previous based on a trillion inputs.<p>We might as well use a random chemical element generator, plug that into our models, and save a ton of money on compute instead.<p>At best Gen AI can be used to create inputs for massive N genetic simulation problems. Most of its suggestions will be bad, but maybe another program can filter the good ones out.
This headline does not make much sense. What the article describes is generative AI for molecular design. That’s a perfectly valid and useful thing to do, but it’s not drug design. Drugs are molecules that affect specific targets in human physiology and reduce disease. There’s a lot of work before molecular design that goes into finding the right target. Then a ton of work to make sure that the drug molecule affects the target in a way that actually reduces the disease. These are the hardest and most expensive parts of drug design. Further, they're not particularly amenable to just plugging a transformer into a huge dataset to do next token prediction. We don’t datasets at the scale of all text on the internet for cellular biology or for human clinical trials. There are certainly groups like the Chan Zuckerberg Initiative trying to put together “atlases” that are very comprehensive datasets of RNA in lots of different types of cells, but they’re fundamentally limited by our measurement technology. Clinical trials will just never have that kind of data. It’s too expensive and too dangerous to try out a million different drugs.<p>Maybe I’m just too close to the problem and I’ll suffer the same fate as pre-deep-learning computer vision researchers. That said, I really think this is a domain where details like physics and biochemistry actually matter a great deal, and I’m suspicious of ML hype that claims we can just wave those away and that some AI will just figure out drugs.