4 ms·

I doubt that. As someone who was enthusiastic about AI before neural networks rose to prominence, I only shake my head when I hear people ascribe all sorts of

I doubt that. As someone who was enthusiastic about AI before neural networks rose to prominence, I only shake my head when I hear people ascribe all sorts of properties of intelligence to the particular trend in AI that is in vogue today. The ultimate quest of AI was to understand how to make good decisions. The idea was that maybe if we emulate human activity that humans call "intelligence", but remove all the randomness, unintended errors etc, then we can understand what it means to be intelligent and subsequently will be able to attack hard problems where humans fail due to things like the size of the problem, or the lack of intuition for a particular domain etc. The neural-net AI abandoned all hopes of understanding of how intelligence works and concentrated its efforts on mimicking a typical human response by asking a lot of people the same question, and then averaging the answers. It inspires no confidence that it will be able to apply to domains where humans themselves struggle (eg. multidimensional geometry) or that it will ever become an authoritative source for any kind of answers (i.e. would produce a proof why its answers are correct). To anyone who wanted to know how intelligence works neural-net AI is worthless. That's not to say that such AI is worthless in general, clearly, it does very well predicting complex patterned behavior, or can sift through a lot of data with difficult to deterministically determine patterns, but it isn't even going in the direction anyone who wanted intelligence would be interested in. I think that as with many historically vague terminology, we'll have to redefine what "intelligence" means. Split this concept into multiple concepts, and, maybe the neural-net group will win the right to call themselves "AI", while those who I describe above will switch to a different term, or maybe it will be the other way around. It's clear though that the modern in-vogue approach is simply a waste of time for a lot of people interested in AI.

ta988 3y ago
You don't seem to understand how thise systems are trained if you believe it is by asking a lot of people and averaging answers. Plus there is a ton of work to now combine LLMs with their excellent language abilities to more formal systems based on rules, ontologies and (fuzzy) logic. LLMs showed that this type of networks can likely solve one aspect if intelligence, the ability to handle language, that previous approaches were all extremely bad at. Doesn't mean that this approach is good for all other aspects of intelligence, but it turns out it is not too bad at information retrieval as well and a few other things. And other kind of networks also showed excellent abilities in other aspects like vision.
- crabbone 3y ago
  I work in HPC infra, so I know some stuff about how these network are trained, but more from a technical perspective, i.e. stuff like "how to set up MPI for a cluster of X nodes" or how to configure GPU drivers. I know basic stuff about how models are trained from another perspective too -- since I work with researchers, and my wife being an RSE who usually works with healthcare related research I'm also exposed to a lot of side effects of using neural network styled AI in practice. I don't claim to have a good knowledge of the ins and outs of the particular trend of the day though. I wouldn't be able to tell what the difference is between "deep learning" and "reinforcement learning", not with any authority or confidence. I did take two courses on AI though, long before the neural-network approach was a thing. I.e. my AI learning experience was focused on search algorithms, building knowledge bases etc. So, here, based on my experience of practical applications of LLMs, I can comment on things you wrote: > Plus there is a ton of work to now combine LLMs with their excellent language abilities to more formal systems based on rules Tons of publications doesn't mean this is a solved problem. As it stands now, there's a serious doubt this is a problem that can be solved. I have not seen any real system working in the world of medicine which would use LLMs for anything that can be encoded as rules. If anything, I've heard a lot about failures to implement anything remotely similar to this (which happened outside of academic medicine, or somehow slept through the cracks). I.e. chatbots suggesting suicide to people in distress, or suggesting to cut on carbs to young women struggling with anorexia and so on. > the ability to handle language, that previous approaches were all extremely bad at Well, now it's my turn to point fingers. You probably don't know much about previous approaches, and don't understand the problem well. It's true that since the term AI was coined substantial efforts went into understanding natural languages. But what came out of this work is a disillusionment, the understanding that natural languages suck at encoding knowledge and as such are a bad tool for building ontologies, or call them knowledge bases. So, it wasn't because older methods failed to parse language well. It was because we realized that there's no point in being good at it. It's a cute gimmick, if you can do it, and can be useful in harmless situations, like enabling rich interactions between NPCs and PCs in video games. But, if your goal is to understand how to make good decisions, it's more of an obstacle really. Like I said previously, but you seem to have ignored: yes, neural networks are useful, but not in the way a lot of people understand intelligence, or, at least, they fail to address the central and the critical questions about intelligence. Compounded by dishonest advertisement of their abilities and armies of low-quality practitioners this approach has a tendency to alienate people who are interested in more fundamental questions about the nature of intelligence.
circuit10 3y ago
We don’t fully understand how human intelligence works, yet that doesn’t mean we aren’t intelligent
- crabbone 3y ago
  I think you stopped reading somewhere in the middle.
  - circuit10 3y ago
    I read it again, all I got is that you want to redefine intelligence specifically to not include current AI That feels like goalpost moving. Having it be interpretable is a useful quality but it’s not the same as intelligence
    - crabbone 3y ago
      Well... no, I don't want to redefine "intelligence". I don't know how you got this idea. What I'm saying is that "intelligence" didn't mean "guessing the best answer by averaging many known, externally generated answers" before neural network approach became popular. Intelligence was always about finding good solutions to problems. In rare cases doing nothing is a good solution to a problem, in less rare cases, statistically guessing is a good solution, but it still doesn't answer the question of how to find a good solution. No amount of improving the accuracy of statistical guess will answer this question because what we need is a proof that the claim is correct, but the guess is the opposite of a proof -- it's asking you to take the claim on faith. In practical terms, compare these two scenarios: 1. A neural net AI is diagnosing breast cancer in a patient. Suppose, it was a male patient (a very rare occurrence, that's likely to fail statistically guessing AI). AI fails to find cancer and the patient dies. 2. A radiologist diagnosing breast cancer in a patient. Same circumstances as before. The radiologist fails to identify cancer, the patient subsequently dies. In the first case, we don't even expect the AI to be correct. Nobody's taking the blame, because, well... we've constructed this AI to be guessing statistically. It couldn't have enough data to guess reliably because in order to do so it needed to see more breast cancer in male patients than we have data for. In the case of a radiologist, we expect the radiologist to understand that they are treating an exceptional case. We will blame them, and probably sue for malpractice because we expect them to be able to reason, to be able to explain why what they see "makes sense" and why their response to what they saw "made sense". We will hold them responsible for failing this patient. In common understanding of the word "intelligence", the radiologist is intelligent (or, at least, we expect them to be), while neural net isn't.
visarga 3y ago
I don't necessarily agree with you but I agree with this discussion. This kind of long-hard-look at what intelligence means is necessary today. We need to understand what we have here, what is this AI, what is our intelligence, and how much we can rely on AI. Just 12 months ago if you polled here you would find much more naive opinions. But now we are learning fast to evaluate AI and human work critically. We have had our hundred hours of AI play and all tried to make it stumble, then to make it work, and in the end we are wiser for it. Regular citizens even know not to trust the bunch of benchmark scores every new model shoves in our eyes, like hardened ML researchers. And basic AI concepts are infiltrating and getting root in the public - model, parameters, tokenisation, datasets, fine-tuning, vector search, and their complex interplay.

ta988 3y ago

You don't seem to understand how thise systems are trained if you believe it is by asking a lot of people and averaging answers. Plus there is a ton of work to now combine LLMs with their excellent language abilities to more formal systems based on rules, ontologies and (fuzzy) logic. LLMs showed that this type of networks can likely solve one aspect if intelligence, the ability to handle language, that previous approaches were all extremely bad at. Doesn't mean that this approach is good for all other aspects of intelligence, but it turns out it is not too bad at information retrieval as well and a few other things. And other kind of networks also showed excellent abilities in other aspects like vision.

crabbone 3y ago
I work in HPC infra, so I know some stuff about how these network are trained, but more from a technical perspective, i.e. stuff like "how to set up MPI for a cluster of X nodes" or how to configure GPU drivers. I know basic stuff about how models are trained from another perspective too -- since I work with researchers, and my wife being an RSE who usually works with healthcare related research I'm also exposed to a lot of side effects of using neural network styled AI in practice. I don't claim to have a good knowledge of the ins and outs of the particular trend of the day though. I wouldn't be able to tell what the difference is between "deep learning" and "reinforcement learning", not with any authority or confidence. I did take two courses on AI though, long before the neural-network approach was a thing. I.e. my AI learning experience was focused on search algorithms, building knowledge bases etc. So, here, based on my experience of practical applications of LLMs, I can comment on things you wrote: > Plus there is a ton of work to now combine LLMs with their excellent language abilities to more formal systems based on rules Tons of publications doesn't mean this is a solved problem. As it stands now, there's a serious doubt this is a problem that can be solved. I have not seen any real system working in the world of medicine which would use LLMs for anything that can be encoded as rules. If anything, I've heard a lot about failures to implement anything remotely similar to this (which happened outside of academic medicine, or somehow slept through the cracks). I.e. chatbots suggesting suicide to people in distress, or suggesting to cut on carbs to young women struggling with anorexia and so on. > the ability to handle language, that previous approaches were all extremely bad at Well, now it's my turn to point fingers. You probably don't know much about previous approaches, and don't understand the problem well. It's true that since the term AI was coined substantial efforts went into understanding natural languages. But what came out of this work is a disillusionment, the understanding that natural languages suck at encoding knowledge and as such are a bad tool for building ontologies, or call them knowledge bases. So, it wasn't because older methods failed to parse language well. It was because we realized that there's no point in being good at it. It's a cute gimmick, if you can do it, and can be useful in harmless situations, like enabling rich interactions between NPCs and PCs in video games. But, if your goal is to understand how to make good decisions, it's more of an obstacle really. Like I said previously, but you seem to have ignored: yes, neural networks are useful, but not in the way a lot of people understand intelligence, or, at least, they fail to address the central and the critical questions about intelligence. Compounded by dishonest advertisement of their abilities and armies of low-quality practitioners this approach has a tendency to alienate people who are interested in more fundamental questions about the nature of intelligence.

circuit10 3y ago

We don’t fully understand how human intelligence works, yet that doesn’t mean we aren’t intelligent

crabbone 3y ago
I think you stopped reading somewhere in the middle.
- circuit10 3y ago
  I read it again, all I got is that you want to redefine intelligence specifically to not include current AI That feels like goalpost moving. Having it be interpretable is a useful quality but it’s not the same as intelligence
  - crabbone 3y ago
    Well... no, I don't want to redefine "intelligence". I don't know how you got this idea. What I'm saying is that "intelligence" didn't mean "guessing the best answer by averaging many known, externally generated answers" before neural network approach became popular. Intelligence was always about finding good solutions to problems. In rare cases doing nothing is a good solution to a problem, in less rare cases, statistically guessing is a good solution, but it still doesn't answer the question of how to find a good solution. No amount of improving the accuracy of statistical guess will answer this question because what we need is a proof that the claim is correct, but the guess is the opposite of a proof -- it's asking you to take the claim on faith. In practical terms, compare these two scenarios: 1. A neural net AI is diagnosing breast cancer in a patient. Suppose, it was a male patient (a very rare occurrence, that's likely to fail statistically guessing AI). AI fails to find cancer and the patient dies. 2. A radiologist diagnosing breast cancer in a patient. Same circumstances as before. The radiologist fails to identify cancer, the patient subsequently dies. In the first case, we don't even expect the AI to be correct. Nobody's taking the blame, because, well... we've constructed this AI to be guessing statistically. It couldn't have enough data to guess reliably because in order to do so it needed to see more breast cancer in male patients than we have data for. In the case of a radiologist, we expect the radiologist to understand that they are treating an exceptional case. We will blame them, and probably sue for malpractice because we expect them to be able to reason, to be able to explain why what they see "makes sense" and why their response to what they saw "made sense". We will hold them responsible for failing this patient. In common understanding of the word "intelligence", the radiologist is intelligent (or, at least, we expect them to be), while neural net isn't.

visarga 3y ago

I don't necessarily agree with you but I agree with this discussion. This kind of long-hard-look at what intelligence means is necessary today. We need to understand what we have here, what is this AI, what is our intelligence, and how much we can rely on AI. Just 12 months ago if you polled here you would find much more naive opinions. But now we are learning fast to evaluate AI and human work critically. We have had our hundred hours of AI play and all tried to make it stumble, then to make it work, and in the end we are wiser for it. Regular citizens even know not to trust the bunch of benchmark scores every new model shoves in our eyes, like hardened ML researchers. And basic AI concepts are infiltrating and getting root in the public - model, parameters, tokenisation, datasets, fine-tuning, vector search, and their complex interplay.