danielmarkbruce

If by "slightly" you mean "not slightly", yes.

No, no it isn't.

Talking in weird abstractions and analogies is what is blowing your mind.

Like, looking at the specifics, it's cool and interesting, a little bit contrived, but it's not insane, there is no philosophy required, no need to shake.

I don't know if you ever spent time with some of the guys that are in the stratosphere math wise, but they are different. I'm not talking about the average math phd - I mean the guys that are really at the top of the field. And you have to be pretty strong yourself to see it... they are way out there.

You keep saying "meritocracy" and the like, as though it's just by nature a good thing. It is not. High IQ isn't distributed evenly or fairly or anything of that nature, and yet "meritocracy" seems fine with the idea that "some people are just smarter than others, deal with it".

Rawls had a concept of the "veil of ignorance" which is probably better for thinking about fairness rather than "meritocracy". No one should lose sleep over the fact that the tiny number of corporates with enough capital can now beat out the tiny number of 180 iqs at a task (mostly generating proofs) that is not really valuable.

I never had a chance against Terence Tao. Neither did you. There are a few dozen people on the planet who had a giant advantage in math and could do things no one else could do.

It was not some great awesome fair thing that math was dominated by some outrageous outliers on the iq spectrum. It's not some great awesome fair thing now either.

Yup, and it should. There is a place for it in drug discovery, but it's likely an Amdahl's law type thing.

Maybe read the comment i responded to.

The US is founded on the idea of individual freedom, and that one of the few areas of "collective" spending should be defense.

This is 101. If you don't like it that's fine, but not understanding it can be resolved with a simple google search.

Phase 1 isn't the big problem. It's phase 2 and phase 3 where drugs fail hard and the trials cost the most and take the longest.

Honestly once you go through all the details... drug discovery feels like the one of the overhyped and unlikely to work areas for AI. Other areas of medicine (like diagnosis) are already seeing pretty good results, and anecdotally seems to work well and tends to fit within the realm of text analysis where most of the AI gains have happened.

Talk to chatgpt about how a checking account works, and really understand what happens when you put some money in the bank, and when you transfer that money to another person at another bank, how it's accounted for, how the banks actually do that transfer.

That's 80% of the thing right there.

Alphafold solved a waayyy narrower problem that had lots of experimental structural data. Guessing how a drug interacts with an entire human is a whole different scale of thing. Even knowing protein interactions wouldn't tell you how those interactions propagate through an entire human system or whether a drug will be safe and/or effective.

Drug trials in humans.

Who said anything about easy and fun?

It's quite a leap to go from the study you linked to thinking AI will create models that will be useful for new trials. The big leaps in AI have required enormous amounts of data. There is no such analogous data sets in biology and no foreseeable way to get it.

It really is hard to see AI improving the trial process meaningfully. We are more likely to see progress with better trial design and regulatory regimes.

It sounds lazy tbh. No one said privatize. It can be open source can't it?

This is a largely agreed upon problem. What isn't clear is why it seems to get sidestepped in discussion so frequently. No one knows how a molecule will interact with human biology. It can't be simulated computationally to generate data (too slow), there isn't a lot of data in the grand scheme of things, so you can't train a model to actually predict the interaction of a molecule with a human.

And nothing I've said here is an original thought or piece of insight.

why not just do it?

Under examined? No. This is a very common talking point.

A glut of models doesn't mean there is no moat. There are several dimensions on which to compete, and it might turn out that being able to produce models that optimize on different dimensions is a difficult and valuable thing to do. It also might not, but the "there is obviously no moat" only comes from people who don't use these things systematically.

Lol, it's a bad deal huh. Tell that to the guys who bought the IPO of Facebook, Google, Amazon, MSFT, Nvidia.

Just because something might be not great in aggregate, it doesn't mean any single IPO is a bad idea. Some people have insight into the future of a specific industry/business. If you can pick one, good for you.

People said similar things about facebook, and similar things about Nvidia stock recently (not ipo, but similar "why buy at this price"). Global GDP is over $100 trillion every year, there is a lot of business to do in the world, AI might be able to do a lot, it might be very profitable.

The number might be too high, but doesn't take more than a little imagination and some basic facts about the world to see how it could be worth 20 trillion in 10 years.

The comment below/above is right. Pre-training isn't the path (but you really do need to learn how it's done). Learn how it works, learn the architecture, then take models and do post training. Fine tuning and RL on them. There was a fun/clever/useless project recently posted here about building a terminal emulating model thing, as an example.

You don't need to code to follow it, just read along

You are completely off base. His initial book was released almost two years ago. I post train models, for work, and I learned more from his book on LLMs than any other source on LLMs. There is a lot more to it than that book, but you have to start somewhere, even if you have a decent theoretical foundation. There are a bunch of people who learned how to build them from his book.

How do you think people learn to build/train models? It comes down from the sky?

You can take any open source model and continue training it. It's nothing like building a CPU.

Get Sebastian's book and build an LLM from scratch. The problem with most textbooks (on any subject, it's not unique to ML/AI) is they usually try to build the theoretical foundations first. In many cases things are better taught top down.

There are some cases where that is still the case - medical diagnosis where you have a lot of text information that you can pass in can be pretty mindblowing with LLMs. If it's text inputs, it's going to be pretty darn good with an LLM. If not, probably not.

This seems silly. It's analogous to an electricity market where every megawatt is a different voltage, in a different city with no transmission lines... might stop working halfway through the job, and only works properly if you buy another 500 megawatts from the same supplier... good luck getting this market off the ground.

Correct, latency is not critical.

Yes, it can get expensive. We do something in financial markets where the cost of a single end to end task could be $100. That doesn't mean cost doesn't matter though, it matters a lot - if we screw up and it propagates, it's money down the toilet. And if we could get the task down to $20, it would make a difference.

There are all manner of documented architectures in papers and books, I don't know that there is anything new under the sun at this point, the devil is always in the details. That said, some of the things we use off the top of my head: generate multiple responses using different models and have a third model judge the best answer; generate an answer and have another model critique the answer and then a third take the answer and critique and generate a new answer; multi-model debate; iterate in a loop either x times or until some condition is satisfied; pipelines of specific single calls to check for specific conditions.

I'd say in general, with respect to errors/low quality responses etc you have to be extremely clear about what you are doing because once you get into the weeds, what is and is not an error is very grey in many cases.