Writing a paper on streamlining the FDA approval process is nice, and I agree with all of it, but it's one thing to do that (about 4 hours of work) and quite another thing to cure cancer.
My point was more that LLMs are exceptionally weak at developing drugs. They're strong at every math/logic/language task, but curing cancer is something of an entirely different nature. Again, if he wanted to "solve physics," or "develop a universal real-time language translator," or even "solve global warming," I wouldn't be nearly so skeptical.
...The first is something that can be credibly attempted with a next-gen super-LLM and a ton of compute. The second is something that LLMs are exceptionally good at, and a dedicated engineering effort might get you there. The third is a modeling/simulation problem with solutions that are, at least in principle, known but haven't yet been thoroughly weighted. (i.e., how bad really is ocean seeding or launching SO2 into the atmosphere?)
But curing cancer? It's not even clear how to approach the problem. Cancer is also heterogeneous... so when I say that LLMs are weak at developing drugs, the problem is compounded, as curing cancer might require dozens or hundreds of distinct new drugs. It's like Dario picked the least likely, most difficult problem to solve. So it's weird.
"We" as in the readers here. We all know the basics around how cancer research is hard and how getting new drugs through trials takes time and money.
Did you know that Amodei was working on research about cancer biomarkers at Stanford?
Given his background in biology, it is hardly like he is a random tech guy with no knowledge on the topic, as many here seem to imply.
Amodei mentioned "actually curing cancer" as an example of undeniable public benefit, not as an exclusive research focus or anything like that. He is clearly talking about generally intensifying biomedical efforts, not promising a cure for cancer specifically.
Writing a paper on streamlining the FDA approval process is nice, and I agree with all of it, but it's one thing to do that (about 4 hours of work) and quite another thing to cure cancer.
My point was more that LLMs are exceptionally weak at developing drugs. They're strong at every math/logic/language task, but curing cancer is something of an entirely different nature. Again, if he wanted to "solve physics," or "develop a universal real-time language translator," or even "solve global warming," I wouldn't be nearly so skeptical.
...The first is something that can be credibly attempted with a next-gen super-LLM and a ton of compute. The second is something that LLMs are exceptionally good at, and a dedicated engineering effort might get you there. The third is a modeling/simulation problem with solutions that are, at least in principle, known but haven't yet been thoroughly weighted. (i.e., how bad really is ocean seeding or launching SO2 into the atmosphere?)
But curing cancer? It's not even clear how to approach the problem. Cancer is also heterogeneous... so when I say that LLMs are weak at developing drugs, the problem is compounded, as curing cancer might require dozens or hundreds of distinct new drugs. It's like Dario picked the least likely, most difficult problem to solve. So it's weird.