Evolutionary psychologist (a genius, everyone should read his short essay on the evolutionary roots of virtue signalling) published a very interesting blog post arguing that AI won't solve longevity and cure diseases as quickly as we expect it to. His argument is that we lack two ingredients that made LLMs widely useful in other domains (such as language or code): first, massive biomedical datasets to train models, and two, rapid feedback loops that would allow us scale RL on the relevant use-cases (feedback loops are slow: we must wait for chemical reactions to happen, for patients to answer to treatments, etc.).
But I think this misses the fact that - holding constant the biomedical data/knowledge humanity currently possesses (and that LLMs therefore do too) - scientific progress is largely bottlenecked by intelligence. Even without running a single new physical experiment, unleashing immense cognitive processing power on our current scientific corpus could allow for important breakthroughs.
LLMs build internal models of the world, modeling the laws of physics, the rules of molecular reactions, causality within the human body, etc. In addition, they are very good at abstract, logical reasoning. Combining these, they could excel at medical discovery from first principle thinking rather than from analyzing more empirical data.
LLMs can apply frontier findings in a domain (such as statistics or material sciences) to a problem in other domains (such as molecular biology), something which no individual scientist can do (human knowledge is now so vast that researchers must specialize to reach the frontier). Breakthroughs (including medical innovations) are often the results of importing paradigms from one field to solve a problem in another field; LLMs can find these connections.
Combining 1) and 2): A lot of medical research is theoretical, and on that front, it seems absurd to imagine that millions of ultra-knowledgeable 800 IQ AI scientists won't be significantly more efficient than humans.
A year ago, Amodei made two interesting points in favour of such an argument:
1% of major scientific discoveries per year drive >50% of progress in biology, and these 1% of major discoveries are often made by the same scientists. If breakthroughs are concentrated among a few outlier minds, what does that tell us ? Well probably that there are very high returns to intelligence in the field. If brute-force data collection were the major bottleneck, discoveries would be more evenly distributed.
A lot of medical breakthroughs could have been made earlier than they were. For example, CRISPR was a naturally occurring component of the immune system in bacteria that’s been known since the 1980s, but it took another 25 years for people to realize it could be repurposed for general gene editing. The data was available, the bottleneck was creativity and intelligence (and probably, having enough scientists work on the issue - something which AI also solves, because the work can be massively parallelized in GPUs running 24/7 in datacenters).
Of course medical research has incompressible constraints. Science can't work without an empirical feedback loop. But nevertheless, the scientific process will probably be massively sped up by AI.
