djierardi

Probability of "if"? I'll give you 1000:1 odds, if you're game.

Its plausible. Its just that the effort will be huge. LLMs can do a lot of stuff because they've ingested all of our published text. But, unless we're extremely lucky, there's just not enough data out there to solve this problem yet.

Maybe like with problems in maths: they've solved 700+ open conjectures, but how many did they try? Are the ones that they solved merely those that they put $10M effort into where humans just gave up because there wasn't any concrete benefit in putting in that sort of effort?

yes!

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.

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.

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.

> 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.)

Not "know" in a generic sense. But an LLM trained in the intricacies of contract law can find more loopholes than you or your cheap lawyer can.

Now extrapolate that to LLMs RL-trained on many niche applications. Good and bad arguments available, but ultimately it seems like bot-vs-bot, as long as the bots are well trained for those niches.

So ultimately a huge penalty on companies that must use US AI, ceding a huge advantage to global-exUS competitors. Game over.

Open source already has the "the collective data of humanity as a whole". That's the LLM leg. Now focus on what they train their models on for RL. Do other models have less code for training coding? for math? Those are pretty much the only cases where we can have reliable validation. Then, what's the next step. So many branches: legal contracs? etc. So much more obscure detail, which is why (I'd guess) so many startups exist in these areas.

AGI, I think, is an illusion: you need a whole lot of specialized bots. Heck, when I query a bot wrt something I know nothing about, I'm blown away by its ability to inform me. But when I ask about something I DO know about, its answers are half-assed. Heck, even when it can access the internet, it avoids all the sites that have paywalls, so it sends me to little reddit to conversations that are entirely untrustworthy.

No, I think not. Copious genomic info but very very little wrt the huge number of possible antigens. Many pharmaceutical companies have data here, but that's proprietary. But even in aggregate, if possible, that's a "drop in the bucket". Maybe useful, but still a drop in the bucket. And even then, if aggregated, perhaps useful in some cases. But go through clinical testing, and you're still talking about ~10 years before anything is proven.

If "cure cancer" is do-able via ML, its still DECADES away. IF.

Yeah but qualify the "know". Their "intentions" aren't conscious, but unknown.

Hah!

This assumes that everyone doesn't reach "good enough" at some point. The best developed area here -- software -- seems to suggest a "good enough" plateau, if not an absolute plateau.