Achievements of Artificial Intelligence (AI)

[Utsikt]

I can't see it. The testing can be automated. There is no need to restrict automation to simulations.

Are you suggesting that we’re somehow going to automate testing stuff on living human beings?

If we ignore the technological capabilities, how is that ever going to become legal or accepted by the public?

 

[EndME]

Are you suggesting that we’re somehow going to automate testing stuff on living human beings?

I'm arguing that you proceed as you would anyways, but that the workforce in labs is at risk of being replaced in upcoming years. That may seem farfetched to some, but I'm starting to become more convinced of not seeing limits than the opposite. The discussion extends beyond LLMs.

 

[hotblack]

I'm arguing that you proceed as you would anyways, but that the workforce in labs is at risk of being replaced in upcoming years

It’s possible that anything could be replaced. Any knowledge work. And yes that includes a workforce in research labs.

I guess the question will come down to cost. What will it be cost effective to replace. Grad students seem to be very cheap. Cheaper per hour than a robot paired with a frontier LLM for now I expect…

This is something already seen in programming where the cost of tokens can be thousands of dollars per month per programmer.

I don’t buy that AI is suddenly going to bring abundance and remove scarcity or resource constraints, perhaps just shift them a bit.

 

[Yann04]

I don’t buy that AI is suddenly going to bring abundance and remove scarcity or resource constraints, perhaps just shift them a bit.

And even if it did, I don’t buy that we would see much of the benefit. Historically the people in charge have been perfectly happy saying wow tech makes this easier. And instead of seeing that as a sort of “lets reduce working hours to improve peoples QOL” it has been, “we keep people working and now we have double productivity”.

 

[hotblack]

Yeah that’s true too @Yann04

It also just occurred to me that the current limit on research isn’t capacity but cost (few countries put significant GDP into it) and possibly what that research is focussed on. I’m not sure how automation changes that given current costs of everything involved.

 

[EndME]

This is something already seen in programming where the cost of tokens can be thousands of dollars per month per programmer.

I think this is not accurate. As discussed by @ChronicallyOverIt the cost equation has dramatically changed because of the rapid speed of progress. There seems to essentially be little developmental progress or investment necessary anymore to outperform programmers. The costs in the past have been artifically low and some in the future may be artificially high due to other investments, but that doesn't really matter much to this particular problem.

And even if it did, I don’t buy that we would see much of the benefit. Historically the people in charge have been perfectly happy saying wow tech makes this easier. And instead of seeing that as a sort of “lets reduce working hours to improve peoples QOL” it has been, “we keep people working and now we have double productivity”.

Seems like the key issue.

 

[Utsikt]

I'm arguing that you proceed as you would anyways, but that the workforce in labs is at risk of being replaced in upcoming years. That may seem farfetched to some, but I'm starting to become more convinced of not seeing limits than the opposite. The discussion extends beyond LLMs.

You’re kind of dodging the question. What will the robots in the lab work on? Will they have access to live humans to do with as they want, or will they work on samples with the limitations that entails?

I don’t have any issue with the outlook of robots taking over wet lab work eventually, even though robotics at that scale, precision and reliability is incredibly costly in all kinds of resources. But that kind of work would still be limited to in vitro, which as we all know can be quite removed from the realities of in vivo.

 

[EndME]

You’re kind of dodging the question. What will the robots in the lab work on? Will they have access to live humans to do with as they want, or will they work on samples with the limitations that entails?

The will do the tasks that would have otherwise been performed by humans. I see no reason to postulate a different scenario for one but not the other. Just how problem solving in mathematics by LLMs seems to have replaced problem solving by mathematicians. Which a year ago everybody would have argued is hard to do at scale and maybe too costly. And of course there's always the possibility that you can always outsource some mundane tasks to humans that don't need specific training as well, if it is cost-effective, just as is the case with almost all products we consume, but those are not biologists, similar to how Adidas doesn't hire any shoe cobblers.

But that kind of work would still be limited to in vitro, which as we all know can be quite removed from the realities of in vivo.

I don't understand why this argument gets repeated when it applies to both scenarios equally. Nobody is suggesting "biology will be solved" similar to how AI will not "solve mathematics".

 

[Utsikt]

I don't understand why this argument gets repeated when it applies to both scenarios equally.

If some advancements can only come from understanding the in vivo relationships, then you can’t expect AI to further our understanding of those areas from doing in vitro experiments.

We can look at petri dishes for as long as we want, it’s not going to tell us anything about what’s going on with the prolactin response in ME/CFS because we lack so much basic info about what’s going on in vivo.

The will do the tasks that would have otherwise been performed by humans. I see no reason to postulate a different scenario for one but not the other.

So you don’t think there are any relevant practical, legal or ethical differences between letting AI loose on petri dished and letting AI loose on humans? None at all that might affect how realistic they are?

 

[hotblack]

I think this is not accurate

Okay well we have to disagree. I’ve seen very clear numbers on this. I now what I have said is accurate.

What happens in the future we don’t know. There’s great advances in smaller and local models. Some companies spend less. Some nothing. But these are all different points. How much all this can be extrapolated is ofc all unknown but so is much of what we’re discussing.

 

[hotblack]

Just how problem solving in mathematics by LLMs seems to have replaced problem solving by mathematicians.

It hasn’t. It is able to do what they can do. That is different I think.

As covered in the blog I linked the talk is of these tools augmenting and keeping mathematicians honest and costs is an issue again

I was given £1M to run my project over 5 years; Anthropic took only 11 days but I do wonder if they spent more money…

It is worth Anthropic doing that now for attention and because labs are running on debt. What hapoens when they are not? They tried charging customers what it actually cost for a Claude subscription and there was a backlash.

I understand how mich the models are advancing and how small local modes can do a lot of what large ones could 12-18 months ago. But costs and resource allocation are not going to disappear I think. What we choose to spend resources on is going to matter more if anything

 

[EndME]

If some advancements can only come from understanding the in vivo relationships, then you can’t expect AI to further our understanding of those areas from doing in vitro experiments.

Of course not. Just how you cannot expect humans to further our understanding of areas where in vivo studies are necessary. I see no need to repeat that argument.

So you don’t think there are any relevant practical, legal or ethical differences between letting AI loose on petri dished and letting AI loose on humans? None at all that might affect how realistic they are?

I have an abundance of ethical concerns, but I'm much less certain whether they'll impact implementation. I also have ethical considerations for AI doing mathematics (for one I don't think non-open source models should be allowed to be trained on open source data), but it seems that these issues are seldomly discussed at greater depth. As an comparison I also have ethical concerns surrounding OpenAIs contract with the DOD, but things have been still gone ahead.

 

[Yann04]

for one I don't think non-open source models should be allowed to be trained on open source data)

I really like that approach. i feel like thats a very fair constraint without dooming the field to only be viable for the richest companies.

 

[EndME]

It hasn’t. It is able to do what they can do. That is different I think.

I suspect it has. Because it is able to do what they can do, but quicker. Someone would have solved $\theta(p_c)=0$, but they no longer can. A growing number of mathematicians are truely questioning whether there are still problems they can work on, not because the Riemann hypothesis will be solved by Claude tomorrow, but rather because many believe Claude would solve it before them.

As covered in the blog I linked the talk is of these tools augmenting and keeping mathematicians honest and costs is an issue again

As one would imagine there is a different range of views on this in the field of mathematics. The blog is largely about LEAN, but the discussion is evolving rather rapidly due to the nature of the rapid progress. I think the contrast between the blog post by Duminil-Copin which started this discussion (which is a less depressing than the one of Jeff Steif who worked/works in the same area) and the article in the NY Times interviewing Hairer show the progress of the discussion rather well.

 

[EndME]

I really like that approach. i feel like thats a very fair constraint without dooming the field to only be viable for the richest companies.

My understanding is that Arvix was set up to benefit the knowledge of humanity under the condition that humans couldn't consume all of it in any case. I think the conditions are different when you have something consuming all of it with the goal of reaping personal profit.

 

[hotblack]

I suspect it has. Because it is able to do what they can do, but quicker.

But it comes back to cost. Why would Anthropic or anyone else continue to poor money into this? They are using these as marketing tools atm.

As one would imagine there is a different range of views on this in the field of mathematics

Yeah that’s a good point. I suppose that’s an important one too. There are a range of views and we don’t know. I’m ok being uncertain on all this. I don’t know. It’s peoples certainty on the future that sometimes worries me. And thanks for the extra links and reading!

I agree with a lot of what you say but am a but am really thrown by the ‘costs doesn’t matter’ argument. It seems to me costs (and more widely resource constraints and allocation) are the one thing that does. At least in the here and now which is something we can know. Maybe I’m missing something though!

 

[Utsikt]

Of course not. Just how you cannot expect humans to further our understanding of areas where in vivo studies are necessary. I see no need to repeat that argument.

Of course humans can further our understandings where in vivo studies are necessary because humans are currently the only ones that are able to do them.

I have an abundance of ethical concerns, but I'm much less certain whether they'll impact implementation. I also have ethical considerations for AI doing mathematics (for one I don't think non-open source models should be allowed to be trained on open source data), but it seems that these issues are seldomly discussed at greater depth. As an comparison I also have ethical concerns surrounding OpenAIs contract with the DOD, but things have been still gone ahead.

I agree about open source and I agree about defence contracts. There are also examples of companies getting people to hand over genetics or other samples for dubious benefits.

But that’s different from experimenting with interventions because that’s already covered by current laws and regulations. So if you want to do it faster you would need to change the entire legal system. That seems like a pretty tall order. In comparison, most of AI is currently exploiting unregulated spaces so they can move as quickly as they like.

 

[Felis Catus]

Of course humans can further our understandings where in vivo studies are necessary because humans are currently the only ones that are able to do them.

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.

 

[EndME]

Of course humans can further our understandings where in vivo studies are necessary because humans are currently the only ones that are able to do them.

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.

But that’s different from experimenting with interventions because that’s already covered by current laws and regulations. So if you want to do it faster you would need to change the entire legal system. That seems like a pretty tall order. In comparison, most of AI is currently exploiting unregulated spaces so they can move as quickly as they like.

Aren't people advocating for AIs combing through EHR records, subsequently contacting people, then analysing the data and so forth? I could be wrong but I find it hard to imagine that ethical considerations will be a bottleneck when they seem to not hamper contracts with the DOD. From what I'm gathering you're arguing that in vivo work is necessary and the legal system will not allow for automation? I agree that in vivo work is necessary and I guess you could be right about the legal things (and I hope you are), but I might place my bets on the companies who are backed by the largest law firms and have more political influence than major countries.

In comparison, most of AI is currently exploiting unregulated spaces so they can move as quickly as they like.

The people that I know that work for AI companies (or tech in general) all work for companies that all very much knowingly and happily break the GDPR.

But it comes back to cost. Why would Anthropic or anyone else continue to poor money into this? They are using these as marketing tools atm.

But that's the point. They don't have to pour much more in these areas, certainly not as much as everybody was thinking 8 months ago. Of course it is marketing, but does that matter for the argument?

I agree with a lot of what you say but am a but am really thrown by the ‘costs doesn’t matter’ argument. At least in the here and now which is something we can know. Maybe I’m missing something though!

Of course costs matter. 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. This means investments necessary to outperform humans at these tasks are now seemingly negligible, at least for the majorty of people working in these field. An hypothetical analogy: The development of ChatGPT 30.9 might cost further trillions, but there will be a competitor product that is worse but requires negligible costs to develop that can cover mathematics and programming sufficiently to outperform humans.

Of course there are a whole load of others costs that are not discussed here. The cost to society, the cost to future society, the true cost of fossil fuel etc, but that tends to be ignored...

 

[Felis Catus]

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

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.