Achievements of Artificial Intelligence (AI)

[Utsikt]

I think it's impossible to predict how much inference or elimination/detection of futile directions could be done computationally. Back in the day of Watson, Crick and Franklin, how many researchers could've foreseen GWAS and WGS analyses and their power? I think many scientists back then thought that wet lab experiments would be the only way forward.

Genetics is based on wet lab work that has been automated. The reason genetics is such a powerful tool is that genes are always causal. It’s essentially a cheat code that allows you to hone in on only relevant bits. Any other biometric can be anything in relationship to anything else.

Maybe that was badly worded by myself, "you cannot expect humans to make progress via in vitro studies if in vivo studies are necessary to make progress" was the intention of what I was saying.

That’s a given. In that scenario AI won’t make any progress either.

From what I'm gathering you're arguing that in vivo work is necessary and the legal system will not allow for automation?

Yes, and the legal system is the only thing that stands a chance of stopping it, other than maybe (big maybe) public perception unless we’re getting into dictatorships.

I guess you could be right, but it might also be reasonable to place a bet on the companies who are backed by the largest law firms and have more political influence than major countries.

If anyone can do it it’s them or dictators, but if your position requires what’s essentially a complete breakdown of a fundamental aspect of modern western society that’s probably something to state clearly when going into the discussion.

Too many influential and rich people who want to be richer have vested interests in AI. Humans will find a solution to reduce the costs.

That assumes the benefits will be realised before the resources run out or the financials collapse. Lots of ultra wealthy people have invested in lots of things before without that turning out well. Dynasties collapse because of it.

 

[EndME]

That’s a given.

Which is what I've been saying this whole time. There's no point for this argument.

If the main argument is one of legality that would simplify things. But I don't think it is. I think these things have to be discussed at much greater depth by people in the various fields.

 

[Felis Catus]

Genetics is based on wet lab work that has been automated. The reason genetics is such a powerful tool is that genes are always causal. It’s essentially a cheat code that allows you to hone in on only relevant bits. Any other biometric can be anything in relationship to anything else.

You are forgetting bioinformatics and computational biology, and how much of it is based on mathematical and probabilistic models which use simplified and unrealistic assumptions, and can't accurately model known and unknown unknowns.

~15 years ago, >98% of the gene/protein function annotations in big databases like UniProt were computationally inferred, and admittedly unreliable. I don't know how it's now. The amount of data has been growing exponentially, unlike wet lab human labour. With projects like the Darwin Tree of Life (which is aiming to sequence 70,000 species in Britain and Ireland), only a tiny tiny percentage of the most crucial data can be experimentally examined or validated (I don't know if they will do it).

People don't sit in the lab and dissect species to find out which genes evolved from a common ancestor. Databases of homologs, orthologs, gene families, gene trees are 100% computational inference based on probabilistic models (to put it simply). So if you are interested in "shared" genes between human and mouse, fruit fly, chimp or whichever one is your favourite, you can check one or more online databases where they will be purely computationally inferred, or you can run some algorithms yourself.

ETA: These are only a couple of examples just to illustrate that it's not all about automating wet lab work.

 

[nafione]

AI doesn’t care about «efficient» because it has virtually endless resources. Humans do not.

And as the world of chess has demonstrated, AI performs far better than anything humans were able to code on their own. Players are now learning from AI by trying to figure out why on earth they did pawn to H4 at that point in the game. They can also learn by using various techniques within the field of explainable AI (XAI) to figure out why AI got the result it did.

The problem AI still faces is that real life if very very complex. Mathematics or chess are relatively narrow problems in comparison. But that doesn’t mean that more advanced algorithms and increased computing power won’t eventually be able to figure out harder real life problems as well, or that it can at least outperform humans that by no means make perfect decisions.

I don't really agree with 'virtually endless.' That claim would need to be considered relative to the relative 'expense' of artificial intelligence, which we don't really know. Plus, we see evidence of the constraints in the economics already (which are major current events.)

I'm quite familiar with the progression of machine intelligence, as it's adjacent to my field of engineering. I understand Moore's law, as well as the evidence that algorithms can get more efficient or 'smarter' with time. That said, I also know that diminishing returns cannot be ignored, as it frequently applies to other technologies. I'm just not convinced diminishing returns won't apply here.

I recognize that LLM's have flipped the field of software engineering on its head. Still, I find current LLM's brittle, and advancement isn't obvious to me in the last year. Maybe my cognitive dysfunction is worse than I thought, or maybe the vast majority of the hype is just to steal investor dollars.

I'm not arguing that machine learning isn't going to continue making advancements in narrow fields, I'm just skeptical about extrapolating that to the general sense.

 

[hotblack]

But for programming and mathematics the costs argument, that was omnipresent argument until recent progress, seems to have become irrelevant due to the speed of progress.

But I have given multiple real world examples of where they do matter now. I have seen no evidence given to adequately underpin the argument that they do not matter. Cheap is not free or without costs worth considering. And we have not hit cheap enough to be inconsequential yet and have no idea on what the costs may be in the future. It’s amazing what we can do and computing power has always increased but that doesn’t make the costs vanish or irrelevant.

To perform biological experiments (which I think is the core push if your argument) physical interactions and equipment are also needed. Are those costs vanishing too? That’s the abundance argument. Even assuming a lot of upsides with AI we spend a tiny fraction of GDP on medical/biological research and I don’t see that changing. Choices will be made by people based upon those costs.

I don't think the financial costs of computing matter.

But they do. You quote the reducing costs of per human genome and I think this is a great example. Have they reduced? Of course, at no point have I said hey will not reduce. But do the costs still matter? Yes! Otherwise we wouldn’t be fighting for money SequenceME or have dine DecodeME before doing so. Cost reduction and cost irrelevsnce are very different things.

To you both, people can buy into the idea that costs will drop forever to nothing and that infinite abundance will take them away. I agree they will drop but the extrapolation of the argument to me is unsound. It is not based in evidence and is an argument used by certain people to bypass a lot of scrutiny. That’s perhaps why I push back on it so hard. If costs ‘don’t matter’ then you can do anything you want. The real world has proved again and again that costs (either direct or in the form of externalities) do matter. Enormously.

To be the costs, even if we limit ourselves to just the financial ones, seems one or the most important and deciding factors on what we choose to do and what impacts are made by this technology. As they have with most other technologies we have developed.

It feels like I’m just repeating myself now saying costs matter and others are arguing back with no they don’t. Which is a sign a good discussion is at an end so I’ll leave it here, we obviously disagree on this point. I’ve really enjoyed the discussion, I see a lot of potential upsides to AI (as well as ML) in science and am excited about the possibilities, people here must know that by now. But I cannot get my head around how easily people say or buy into ‘costs don’t matter’ given the world we live in. Maybe I’ve misunderstood and the argument being made is different or more nuanced.

 

[nafione]

But I cannot get my head around how easily people say or buy into ‘costs don’t matter’ given the world we live in.

big tech propaganda forces stronk

they better be to keep the bubble going

but yes its tough to argue compute is cheap, when the biggest "AI" companies are burning cash at an unprecedented rate

 

[EndME]

But I have given multiple real world examples of where they do matter now.

Have a look at the posts by @ChronicallyOverIt. All mathematics PhD students are outperformed in problem solving by models, where costs are marginal (I think it was you who mentioned that graduate students are cheap but they still "cost" thousands per month). People won't have to spend thousands on tokens that outperform people in these professions in the near future. This is becoming the accepted argument as far as I can see. Pricing of current models and future investment doesn't really matter in that argument. Someone can rather "easily" build a model that is just focused on running a "outperform current mathematicians in problem solving and programmers code" with negligle costs and investment. The argument is that in those scenarios there is little need for future investments. One can argue whether that is true, for example mathematicians or programmers could shift focus and discover that whilst classical problem solving cannot be done by them anymore there are other tasks where they outperform LLMs (Tao listed a few of those, I also wonder whether "mathematical model selection" could be something like that, since it appears to be less described as an "optimisation procedure" in the literature).

To perform biological experiments (which I think is the core push if your argument) physical interactions and equipment are also needed. Are those costs vanishing too?

Yes, that is a much harder argument because there is a genuine need for extrapolation and you have to predict the future costs (which is not really the case above) and investments necessary in robotics. Which nobody knows. It is far harder to predict and that is why it should be openly discussed by experts in the different disciplines today, after all entire disciplines are at stake. If one runs under the assumption that the largest bottleneck is programming skills that are equivalent to those of humans, then I believe you eventually land in the above scenario again, at least if the necessary computing grows somewhat linearly rather than exponentially. Whether or not that is true, can be debated. If it is resource scarce materials, then the situation may be different.

To you both, people can buy into the idea that costs will drop forever to nothing and that infinite abundance will take them away.

But this is not my argument at all. I'm not talking about Moore's law.

I agree they will drop but the extrapolation of the argument to me is unsound.

You don't have to extrapolate for mathematics and programming. That's the point. For robotics the situation may be different and the question is whether you end up in the first scenario again or not.

It is not based in evidence and is an argument used by certain people to bypass a lot of scrutiny.

No, it is the reality currently faced by mathematicians and programmers. They are faced with extinction if they are unable to have larger conversations. I'm not talking about the Elon's of the world with an never ending investment chase to chase AGI or whatever "costs don't matter" arguments they come up with.

One can also debate the "true" costs of energy, but I think that is not the argument you're making even if it is something that should be discussed by the greater society. The reality seems to unfortunately ignore it even if icebears are soon to vanish.

 

[Felis Catus]

It feels like I’m just repeating myself now saying costs matter and others are arguing back with no they don’t. Which is a sign a good discussion is at an end so I’ll leave it here, we obviously disagree on this point. I’ve really enjoyed the discussion, I see a lot of potential upsides to AI (as well as ML) in science and am excited about the possibilities, people here must know that by now. But I cannot get my head around how easily people say or buy into ‘costs don’t matter’ given the world we live in. Maybe I’ve misunderstood and the argument being made is different or more nuanced.

Sorry if my brief comments made you frustrated. I'm trying to write less given how ill I am.

Re costs, I'm actually not reading the news from tech people trying to oversell their ideas and attract investments. I see how AI has transformed my partner's job. At one point YouTube was giving me videos of artists complaining about the drop in commissions. Whether I call a hospital or John Lewis customer service, I need to get through a machine first which doesn't always understand what I am saying. Some of these examples are less technically advanced than others. The point is that I can see that AI has transformed certain industries and penetrated others. I think there's no way back. Big businesses have seen what it can do. I don't think they will go back doing things the old way. It's gone too far. It's not just about making money either. Humans like challenges and solving problems. So I think that the costs don't matter - as long as they are too high, many people will be working on ways to reduce them.

I put a graph for the costs of sequencing a human genome to show how much costs have dropped despite it's relatively narrow application. AI as a technology has the potential to be everywhere, so the motivation to make it financially cheap (and hopefully less detrimental to the environment) is much higher and way more people are going to be involved to make it happen than it's been the case with sequencing technologies.


EDIT: replaced a word

 

[hotblack]

Sorry if my brief comments made you frustrated. I'm trying to write less given how ill I am.

And I’m sorry for making you feel like you need to expend energy and explain more.

No problem here, I have enjoyed the discussion.