> Say they cure cancer.
A pet peeve of mine.
People keep mouthing this same fantasy. But cancer is so many different cancers. And "the cure(s)" aren't just 1000 different drugs, but maybe 10,000 different candidates for each that pass into lab tests, of which maybe 1% make it to phase 1 trials, and of those maybe 10% make it all the way through phase 3, each after 5-10 years of clinical trials and $100s of millions invested each. For the most part, AI-related biotechs are largely focussing on accelerating the discovery of good candidates. But that's a tiny fraction of the pipeline, which includes lots of lab work (which can be automated) and clinical trials (which OTOH require actual human guinea pigs and years of observation and assessment). How do you meaningfully accelerate that? Presumably by improving the odds of success in phase 3 trials from your initial candidates.
But nobody has proposed yet how LLMs can be a magic bullet here. At best they'll give you those 10,000 candidates per cancer type faster. But then you have all that lab work and clinical work to slog through, and all the likely failures you'll see in clinical trials. And there's just not enough data out there to train ML on to meaningfully improve the odds of success in clinical trials, since you'd really have to understand how your drugs work with all proteins and systems in every type of human body, e.g. "in silico" clinical trials. And we're only merely scratching the surface of that with stuff like OpenADMET.
That is, we can't even model well "how a human body works", much less "how those mechanisms go wrong in each type of cancer". But even if you could do all that, for the cures you'd need to figure out what novel, unusual, out-of-distribution thing (relative to your human model) you can throw into the system to avoid the bad behavior, that doesn't kill the person or interfere with other systems or get degraded too quickly or ... tons of other stuff.
Highly non-trivial. Maybe lower-hanging fruit is with a novel peptide or antibody, since those are better modeled with ML. But small drug-like molecules are galactic in number and not well modeled yet. ("Hey, what happens if I introduce this weird platinum-based drug into a body?" Well yeah, cisplatin is used for chemotherapy, but the body of knowledge needed to assess platinum-based drugs in a human body just doesn't exist, and ML isn't gonna intuit it from known data, nor "learn it" since that would require going down to QM, and then integrating that into its model of all human peptides and pathways.)
This is not a side-project for a chatbot company. Its a (multi-)generational project that requires buy-in from labs that can produce the requisite data.
(And okay so maybe I'm fast-and-loose above. Ideally a single "cure" would be to fix the apoptosis pathway, i.e. repair the built-in cell-death path for fast reproducing cells. But cancer cells mutate quickly and evade known avenues. Perhaps there's a generic fix? Plausible but same reasoning applies: ML doesn't have enough data to intuit a solution, if one exists. And there's no reason, a prior, to assume that one does exist. Only hope.)
Great take and useful for the thought experiment.
I just feel like your taking it all as a given/fact. And maybe you’re right, but I’m not that well versed into it to say things for sure.
What I'm trying to get at is that if we’re considering a black swan event, why not consider both the bull and the bear case or even the neutral one and at least accept that we don’t know which one it will be and it may well be a coin toss…
Perhaps using “cancer” deviated things into a specific subject, but it could be anything. A drug discovery, quantum computing, etc.
Again, I’m no quantum physicist or chemist, but as a world citizen, these feel within the realm of possibilities. Just like it destroying humanity or simply just being yet another technological advancement like the computer or the internet was.
—-
I guess the people working on computers could have argued there was a 10% chance it was going to end humanity and so the people working on the internet because AI was going to come about.
—
I’d say that my final conclusion is that we don’t really know and the signals we’re getting from founders, investors, the market, here, and other people we talk to is that NOBODY KNOWS. All we can do is to continue to speculate and make bets based on hunches.
Yeah. What I think is deceptive is something like solving Navier-Stokes with $10M of compute. The bigger question is how many conjectures and hypotheses they tried, and what fraction of those actually succeeded. If they expended say $1B on various open problems, and only 1% succeeded, is that the kind of investment that one will make? (The $10M alone is more than what would have been devoted to academics investigating these issues. But realistically one has to quantify the cost of the entire enterprise, and the probability of success vs failure.
From OAI we're hearing only about 700+ successes. Relative to what failure rate? The cost of success isn't merely the cost of each success, but rather the overall cost of all experiments, of which -- I'd guess -- a huge fraction failed.
I think that’s what VCs do. Only a fraction of investments succeed and the ones that do pay off big time.
Now, imagine having a mechanism that will constantly improve, so you can go from 1% success rate to 5% success rate in say a year.
I see your point. What I think could be different is the speed at which things can be done compared to conventional ways of discovery.
Think of a COVID/MRNA like solution for other areas.
I'm all in on that, and I own Moderna stock ... so there's that. But we sort of understand the pathway there: we want to introduce a protein that induces an immune system reaction. Traditionally you re-engineer a virus that produces that protein that produces the RNA that gets transcribed into that protein. But instead, with better tech, you introduce that RNA directly. Simpler and faster. And perhaps you can use that mechanism to introduce other proteins for other ends (as recently with its personalized melanoma treatments).
I'm all for the notion that ML can improve 1% to 5%. But I haven't seen anyone suggesting HOW that happens. So far, it looks like taking 1000s of candidates to 10,000 candidates, which just further clogs the pipeline. Using ML to automate lab tests seems like a win there. But still, pushing those through phase 1, phase 2, phase 3 clinical trials is the bottleneck. 90% of what enters that pipeline dies, and each involves many years and many $100s of millions of investment.
But as I noted: the quicker path would seem to be through peptides rather than small drug molecules. ML does pretty well with proteins/peptide chains and mRNA provides a way to introduce them (or injection, like GLP1's). Still requires all the testing, though, but presumably we know more about how they degrade, etc. over time.
It seems you just don't wanna see the bubble. Or maybe there isn't a bubble and that's your call to invest more? Amodei would be happy.
Oh, I see it. I’m just thinking 5–10 years down the line not short-term volatility.
Could there be more ChatGPT moments down the line in other areas like science, genetics, space and something we haven’t considered? I think there’s a chance.
I also think AI is here to stay and there may be some confirmation bias from the community regarding open source models. It’s like someone saying back then that Linux was going to take over… (?) Open source software seems to be served for technical people. Not for the average Joe looking for an answer or help generating a social media post or help writing an email.
Indeed, ML is here to stay. But there's a difference between biology/chemistry and math, e.g. There isn't enough data out there on which to train MLs for the former.
I'm thinking here of LLMs as a lossy compression of all our published knowledge. And then agents doing some smart/heuristic search relative to that. They can surely find things that we have missed -- just because they'll tireless try all avenues of investigation that humans might have given up on -- but if the underlying data is not there they can't manufacture it.
LeCun has made this point elsewhere wrt world models and more ordinary physics. This issue is exaggerated in the case of biophysics and quantum mechanics.