AI has actually made plenty of things easier and is not evil. The problem is AI is so many different things but most people don’t talk about those as different. LLMs are the problem and a type of AI, but they aren’t the only AI.
AI (in the non llm form) is used in medical research, material science, chemistry, and more and has been for several years (though has exploded recently alongside LLMs).
AI as a whole is very much on the level of usefulness of an engine or the Internet. LLMs can go die in a fire.
However LLM is the one really driving the datacenter and investment. Things like machine vision are generally not controversial, except for surveillance.
Most applications of AI/ML don’t need nearly the size of processing they’re throwing into these data centers.
If I put on my crazy hat, I’m actually suspicious of the purpose of these massive data centers for LLMs. In no way does any of this reach the singularity. All it does is counterfeit the thinking process by having all the information shoved into the model. I say counterfeit because people still believe they do think, and that’s not how any AI model works.
AI as a whole is very much on the level of usefulness of an engine or the Internet.
Right so that’s what I’m kind of challenging here. What can machine learning do today that we couldn’t do without it.
I fully understand how you could have weather prediction models that might outperform traditional algorithms, but it would need waaay more power and the gain is a small increment instead of a difference of magnitude.
Remember those pictures in PC magazines from the 90s that showed how much data can be stored in a single CD vs paper? What’s the ML equivalent of that?
I’ll just focus down on one example, but I’ve seen this amount of success with non llm AI elsewhere. Protein folding. I’m not directly in the space but my understanding from friends who are researchers is this has solved massive problems for them. Cutting years of time down to potentially days (it’s still in early adoption so not fully proven. But I trust their judgement).
Alphafold uses an attention network to determine potential fold patterns and fairly well. This let’s them cut research time and materials as they can very quickly actually predict likely outcomes without having to directly test things in an experimental environment. (Where they go from there I get lost, I’ve helped friends write script to sort data so my knowledge of this area ends at the computer phase)
Yes you can argue we technically could do it manually before AI, but ultimately you could say that about most any tool. The cotton gin didn’t do anything we couldn’t do before, but it sure changed things with it’s speed and utility.
But genome and protein sequencing has been getting easier for the last few decades for a number of reasons. While it sounds like ML is contributing positively, idk if I would call this revolutionary.
Is there anything new the average person has access to today that they didn’t have before?
IMO revolutions are like: I have a computer in my pocket that has thousands of times more power than the best ones a few decades ago.
While I would love to see a revolution in medical science, I have yet to see it (maybe I’m just uninformed). We are getting close to designer babies but I would argue that’s more like social media than the internet in terms of benefits to society.
I fully agree that ML can help in niche ways but yeah I would just say we have another tool in the toolbox not that we are dead set on huge changes in the coming years.
AI has actually made plenty of things easier and is not evil. The problem is AI is so many different things but most people don’t talk about those as different. LLMs are the problem and a type of AI, but they aren’t the only AI.
AI (in the non llm form) is used in medical research, material science, chemistry, and more and has been for several years (though has exploded recently alongside LLMs).
AI as a whole is very much on the level of usefulness of an engine or the Internet. LLMs can go die in a fire.
However LLM is the one really driving the datacenter and investment. Things like machine vision are generally not controversial, except for surveillance.
Most applications of AI/ML don’t need nearly the size of processing they’re throwing into these data centers.
If I put on my crazy hat, I’m actually suspicious of the purpose of these massive data centers for LLMs. In no way does any of this reach the singularity. All it does is counterfeit the thinking process by having all the information shoved into the model. I say counterfeit because people still believe they do think, and that’s not how any AI model works.
Also, industrial processing. Predictive maintenance on equipment.
It is likely to be the key needed to make fusion power possible.
While I believe some uses of LLMs are valid, I agree that the emphasis on them is all wrong.
Predictive maintenance on equipment is usually not AI. They’re basic heuristic algorithms.
I only know PID and such so I can’t speak on fusion system control, lol.
https://www.pppl.gov/news/2024/ai-approach-elevates-plasma-performance-and-stability-across-fusion-devices
I know these guys advertise it…
https://www.monolithai.com/
Right so that’s what I’m kind of challenging here. What can machine learning do today that we couldn’t do without it.
I fully understand how you could have weather prediction models that might outperform traditional algorithms, but it would need waaay more power and the gain is a small increment instead of a difference of magnitude.
Remember those pictures in PC magazines from the 90s that showed how much data can be stored in a single CD vs paper? What’s the ML equivalent of that?
I’ll just focus down on one example, but I’ve seen this amount of success with non llm AI elsewhere. Protein folding. I’m not directly in the space but my understanding from friends who are researchers is this has solved massive problems for them. Cutting years of time down to potentially days (it’s still in early adoption so not fully proven. But I trust their judgement).
Alphafold uses an attention network to determine potential fold patterns and fairly well. This let’s them cut research time and materials as they can very quickly actually predict likely outcomes without having to directly test things in an experimental environment. (Where they go from there I get lost, I’ve helped friends write script to sort data so my knowledge of this area ends at the computer phase)
Yes you can argue we technically could do it manually before AI, but ultimately you could say that about most any tool. The cotton gin didn’t do anything we couldn’t do before, but it sure changed things with it’s speed and utility.
That’s great to hear.
But genome and protein sequencing has been getting easier for the last few decades for a number of reasons. While it sounds like ML is contributing positively, idk if I would call this revolutionary.
Is there anything new the average person has access to today that they didn’t have before?
IMO revolutions are like: I have a computer in my pocket that has thousands of times more power than the best ones a few decades ago.
While I would love to see a revolution in medical science, I have yet to see it (maybe I’m just uninformed). We are getting close to designer babies but I would argue that’s more like social media than the internet in terms of benefits to society.
I think your definition of revolutionary tech and mine are just fundamentally different here so I don’t think I can point you at what you want.
That’s fair.
I fully agree that ML can help in niche ways but yeah I would just say we have another tool in the toolbox not that we are dead set on huge changes in the coming years.