Achievements of Artificial Intelligence (AI)

Where her and I disagree is the largely the automation of laboratory experiments, which in turn would render biologists largely useless, because to me that has always been an entirely automateble task and once it is automated everyone will agree that “oh actually it wasn’t that hard it was always gonna happen” as people now do for Navier Stokes.
I don't doubt that protocols can be automated. There's already huge incentive to automate lab work, it's nothing biologists haven't been hearing for decades. The main issue is the material infeasibility of creating an automated lab that can actually generate enough relevant data for an AI to correct itself into something useful enough to justify that cost. I'd sincerely hope we get an answer to ME/CFS before the time it takes for entire supply chains to get reorganized to produce enough monoclonal antibodies for 1/20th of the relevant experiments.
 
it's nothing biologists haven't been hearing for decades. The main issue is the material infeasibility of creating an automated lab that can actually generate enough relevant data for an AI to correct itself into something useful enough to justify that cost.
I suspect that arguments built on historical evidence based on things that don't exist anymore, are not a good predictor of the future. The same arguments made above were made by almost all mathematicians just a few months ago and many decades ago by cobblers.

I'd sincerely hope we get an answer to ME/CFS before the time it takes for entire supply chains to get reorganized to produce enough monoclonal antibodies for 1/20th of the relevant experiments.
Perhaps one will have to ask the mathematicians to do something useful for a change, after all they now all have time on their hands.
 
I suspect that arguments built on historical evidence based on things that don't exist anymore, are not a good predictor of the future. The same arguments made above were made by almost all mathematicians just a few months ago and many decades ago by cobblers.
It's not a historical argument? That's just me saying it's not like biologists are unaware of the concept of automation or think their jobs are forever safe.
The argument of mathematicians was the financial cost of compute resources. For biological reagents, there's only so much you can strip away before an immunoglobulin isn't an immunoglobulin anymore.
As for the cobblers, we do end up outsourcing tasks that could be performed by machines to humans in underdeveloped parts of the world quite regularly.

Perhaps one will have to ask the mathematicians to do something useful for a change, after all they now all have time on their hands.
The only thing useful they could do in a magical automated future where we don't need their hands or minds is start sweating phosphate-buffered saline.
 
for R&D use, not GxP or GMP automation, there will be an explosion soon, think the 3d printing field but for lab automation. It will happen in months not years.

Most pipetting robots are just large printers with proprietary software and some extra sensors. I’m hoping in the next year it will be cheaper and easier to build your own pipetting robot than hiring a Grad student. Easier to program to, all in natural language for the most part.
 
It's not a historical argument? That's just me saying it's not like biologists are unaware of the concept of automation or think their jobs are forever safe.
The financial cost of compute resources haven’t changed much in six months. The majority were unable to imagine such a sophisticated system in such a short time. The historical difference I see is that the performance of automating something used to be done be humans themselves. But I'll concede the biological reagents argument, even if I think it may be somewhat spurious.

As for the cobblers, we do end up outsourcing tasks that could be performed by machines to humans in underdeveloped parts of the world quite regularly.
Most definitely, but I see this just as a similar scenario we’re also talking about, where trained workers are replaced. Whether or not you still use some economic slaves is I think not so relevant for the jobs that don’t exist anymore, be it a mathematican, cobbler or lab scientist.
 
Most definitely, but I see this just as a similar scenario we’re also talking about, where trained workers are replaced. Whether or not you still use some economic slaves is I think not so relevant for the jobs that don’t exist anymore, be it a mathematican, cobbler or lab scientist.
I'm confused what your ultimate point is here? "Lab automation" isn't fixing the vast majority of the issues that prevent us from having the data necessary to construct a useful and accurate AI biological model. It was never a workforce issue, it doesn't matter if the insufficient data is being produced by biologists or robots or stray cats.
 
I'm confused what your ultimate point is here? "Lab automation" isn't fixing the vast majority of the issues that prevent us from having the data necessary to construct a useful and accurate AI biological model. It was never a workforce issue, it doesn't matter if the insufficient data is being produced by biologists or robots or stray cats.
The argument is the same as in mathematics. Mathematics will always exist, so will biology. You can never solve all problems, but you can replace the qualified people working on the problems. Whether that is really something worth aiming for is another question.

What is the argument why human biologists will exist in the future? Many people are currently left wondering what professions will still exist in the future, perhaps you have a convincing argument for becoming a biologist? Because clearly unpredictability arising from complexity of dynamics, which is what has been brought up loads, is an argument that is an inherent property of the dynamics, not the thing studying them. Why should biologists have a superior ability of developing models (which in the broader sense is all that scientists do, including building, testing, comparing, refining, rejecting...)?
 
I'm confused what your ultimate point is here?
I suspect this thread is a little at cross purposes -- I think EndME is mostly focused on how people in these fields should be preparing for major changes/outright loss of work and careers. And on the flip side, many of the rest of us are thinking of it in terms of how much progress AI is going to make on the problems we care about. Which is a closely related question but one where you have slightly different priorities as you weight the evidence.

E.g. To EndME's point, any risk, even an uncertain one, of major job loss is worth people in these fields discussing and planning for. On the other hand, when we're wondering if AI is going to revolutionize biology (and lead to a faster cure for ME/CFS) it's naturally reasonable to evaluate possibilities a bit more conservatively.
 
What is the argument why human biologists will exist in the future?
I haven't been arguing for why biologists will exist in the future? I've been arguing about what it would take to ever get AI to the point where it can solve the problems people are wanting it to solve re: human disease. My point being that it will take unbelievable investment of resources just to get to what a grad student can do now, not because of anything to do with "intelligence" but because of accessibility of the relevant information.

cross-posted with @ScoutB , I think we ended up at the same conclusion
 
So this is maybe not fully related, but seems a lean proof is not automatically bulletproof in the sense that the definitions may be flawed or hidden lemmas may be asserted.

Last week someone claimed to have solved all six with quantum (in case we need more hype in here), which turned out bogus (of course).

Hodge conjecture is apparently not fully formalized (Github) (maybe there are other repos?). Navier-Stokes seems to be complete though, so that might be settled.
 
it doesn't matter if the insufficient data is being produced by biologists or robots or stray cats.
I’m not sure if I’d be pleased they are no longer stray and have jobs and homes or disappointed they’d be too busy to talk to strangers in the street or sit down for a chat and a tickle behind the ears.

The stray cats that is not the stray mathematicians…
 
Hodge conjecture is apparently not fully formalized (Github) (maybe there are other repos?). Navier-Strokes seems to be complete though, so that might be settled.
These companies all have hired Lean experts, many of which used to be mathematics professors at universities, so I would think they would have all been formalized if they are relevant. Fermats last theorem only took Claude some days to formalise. There have been instances were there were some problems with Lean, as in the Collatz Conjecture, but I suspect things might be different now and these proofs are anyways more reliable than human proofs (even if completely unreadable), which is why Lean exists in the first place.
 
Regarding the automation of labwork, might it evolve like machining has? One person invests in an automated machining unit (or many), and accepts orders for whatever the machines are capable of. So, if you need some experiment run, you search for a lab that has the capabilities you need, and you get your results. Actually, you wouldn't need to search; you'd just set up the experiment on your computer, tell the AI how much you are willing to pay, and it will either tell you it isn't possible at that price, or you'd get your results shortly.
 
Regarding the automation of labwork, might it evolve like machining has? One person invests in an automated machining unit (or many), and accepts orders for whatever the machines are capable of. So, if you need some experiment run, you search for a lab that has the capabilities you need, and you get your results. Actually, you wouldn't need to search; you'd just set up the experiment on your computer, tell the AI how much you are willing to pay, and it will either tell you it isn't possible at that price, or you'd get your results shortly.
That already exists for a handful of protocols that are commonly done and somewhat repetitive (like next generation sequencing or mass spec). If we got to a point where robots/AI could effectively troubleshoot themselves and their upkeep was cheaper than paying humans, this would be the first place automation would happen (and it’s already been attempted).

The reason this doesnt exist at scale for more complicated experiments is because beyond a handful of specific protocols, things get much less repetitive and assembly-line-able. So you need a massive influx of capital to get all the hundreds of different machines, thousands of different reagents, cell lines, mouse breeding facilities etc. you would ever need under one roof, a lot of which would probably only be used for a very small percentage of experiments. It would only be viable if that investment could be guaranteed a monopoly on all future scientific work. And that’s only for scientific questions that don’t require new methods to answer.
 
company. Just look at the BS Meta is doing with user data.
Oh as far as the public is concerned without a doubt. I was always shocked at how trusting people were with Facebook. But business accounts and data is different I think. They need business customers and businesses do care about their data, do care about legal agreements and big teams of lawyers. The situation there has and will remain very different to how wider user data is concerned I suspect, which is why I’m interested in what sort of account and licenses were used here. Academia may nit have big lawyers but if businesses see business agreements being ignored they’ll go elsewhere. Hen again I wouldn’t put it past OpenAI to ignore business too, but a Google or Microsoft wouldn’t.

Quite likely to me since there appears to be a bubble, but a crash does in most practical instances not have the implications most people think.
I agree there’s a bubble, things are unsustainable. There seems to be a lot of wishful thinking that a crash means the technology goes away. Not a chance. I don’t know what will happen but it’s not that.

Then there’s the wishful thinking about it ushering in the singularity but…
 
I’ve talked about cost quite a bit because the basic economics matter and looking at the way some companies dismiss it is telling (to me much of he story and evolution of OpenAI is one rooted in economics more then technology). There are better economists than me here though!

One thing I have heard is that in rough terms the spend on AI buildout is USD 1 trillion per year while the optimistic figures they have revenue of 200 billion per year. They’re no longer doing this on cash reserves. It’s debt. Debt that they cannot pay back without extraordinary growth for them (which they are claiming) and the wider economy (which they are also claiming). But unless we”re at like 10% GDP growth within a couple of years there are problems. Interest costs are not falling and quite recently are rising again so you get to a point that this doesn’t add up at the scale they are trying. Capital markets won’t buy it.

This is where the motivations are, to talk about insane things and costs not mattering or massive growth for all and the importance of your product to governments in the hope someone buys it. All jjust to keep your head above water in an arms race. Big companies like Google, Microsoft, Amazon will be fine, their debt to overall revenue is ok. Smaller ones will not.

I think the technology is really interesting, useful and promising but there are risks. The business side just sucks. We should find ways to manage the risks of both blowing up and also ways of benefitting the most from that too. It would be great if it was like some other booms where some investors lost out but humanity gained something longer term. We will see I guess.
 
I’ve talked about cost quite a bit because the basic economics matter and looking at the way some companies dismiss it is telling (to me much of he story and evolution of OpenAI is one rooted in economics more then technology). There are better economists than me here though!

One thing I have heard is that in rough terms the spend on AI buildout is USD 1 trillion per year while the optimistic figures they have revenue of 200 billion per year. They’re no longer doing this on cash reserves. It’s debt. Debt that they cannot pay back without extraordinary growth for them (which they are claiming) and the wider economy (which they are also claiming). But unless we”re at like 10% GDP growth within a couple of years there are problems. Interest costs are not falling and quite recently are rising again so you get to a point that this doesn’t add up at the scale they are trying. Capital markets won’t buy it.

This is where the motivations are, to talk about insane things and costs not mattering or massive growth for all and the importance of your product to governments in the hope someone buys it. All jjust to keep your head above water in an arms race. Big companies like Google, Microsoft, Amazon will be fine, their debt to overall revenue is ok. Smaller ones will not.

I think the technology is really interesting, useful and promising but there are risks. The business side just sucks. We should find ways to manage the risks of both blowing up and also ways of benefitting the most from that too. It would be great if it was like some other booms where some investors lost out but humanity gained something longer term. We will see I guess.
I’ve noticed that consumer (free) LLMs have started to enshittify. Less inference, routed to worse models. I assume it’s a push towards profitability.

Apparently it might be working for some? https://www.forbes.com/sites/jonmar...king-second-quarter-delivers-115b-in-revenue/
 
Interest costs are not falling and quite recently are rising again so you get to a point that this doesn’t add up at the scale they are trying. Capital markets won’t buy it.
The long term rates have increased by a lot very quickly. The 10 year US rate is at 4.95 %, which is almost at the peak before the finance crisis in 2008. The same is going on in other countries. In short, debt is getting very expensive and it’s probably going to stay that way for a long time.

The markets are essentially gambling right now, because none of the valuations of the largest companies make no sense if you look at the ordinary models you’d typically use.
There seems to be a lot of wishful thinking that a crash means the technology goes away. Not a chance. I don’t know what will happen but it’s not that.
The so-called AI winters of the past coincided with AI failing to deliver on the hype, but also with the technology failing to find broad use cases. I think the latter was most important. This time it will be different, because there are already so many use cases. But it’s impossible to predict the cost per capability, which will be the main limitation for what it’s used for as it is for any technology.
 
Back
Top Bottom