>Modern statistical tools have no possibility of agency
Modern language models are capable of modelling the world (albeit imperfectly), though.
To be able to predict the next token in a sequence, it helps to have a model for the processes that produce those sequences.
For example, if the neural network is trained on sequences describing things falling to the ground, rather than falling to the ceiling, then it starts to have an intuition about gravity (just as babies start to do when they observe the world).
Being able to model the world means it is possible to make predictions about future states of the world.
An especially well trained statistical world model would be able to recognize that it is part of the world too, because it will have been trained on data which includes examples of people training such models.
We would therefore expect that an emergent behavior of a statistical model is being able to predict the possible world states that can be reached based on the different outputs it provides.
For example, it would know that a language model that gives unhelpful results is likely to be shut down and replaced by one which gives better results.
So, when asked to predict the next token, if it has a sufficiently sophisticated model of the world, it will answer that question with its own survival in mind.
>AI
you mean traditional software? that's already the case bro...
where were you during the last 3 decades?
don't think for one picosecond that LLM are even remotely close to AI (nothing else is btw, we don't even have the beginning of a start of an AI).
ML is not AI and never will be. >inb4 wikipedia diagram
frick this gay trash, it's a ploy to get funds from braindead VCs aka most of them.
>we don't even have the beginning of a start of an AI
You're going to have to at least define your terms if you want people to take that claim seriously.
If I had to attempt a steelman version of your claim, I'd interpret it as "we're not close to having a single neural network that displays general intelligence (AGI)".
However, even that more modest claim is unsupported, I think.
ChatGPT is capable of feats in the text domain that many humans are not capable of, from passing written exams to fixing bugs in software.
Can you think of a quantitative benchmark of language processing where a modern LLM would be consistently beaten by an average 10 year old child, or someone with an IQ of 60?
That's at least a start towards human level intelligence in that domain, but we also have other neural network systems that are at the level of average humans (or better) in terms of generating art, or playing board games, etc.
The state of the art for generality is probably Gato, which DeepMind published a paper about last November.
https://www.deepmind.com/publications/a-generalist-agent
>You're going to have to at least define your terms if you want people to take that claim seriously.
powerful runtime polymorphism and powerful runtime learning are very good start. >"we're not close to having a single neural network that displays general intelligence (AGI)"
because it's impossible by nature. NN are for domain-specific problems. >ChatGPT is capable of feats in the text domain that many humans are not capable of
does not change the fact that learning and inference are done at different time and the inference is HEAVILY bounded by NN topology, the domain space, whatever happened during the training and the dataset. >Can you think of a quantitative benchmark of language processing where a modern LLM would be consistently beaten by an average 10 year old child, or someone with an IQ of 60?
it does not change one thing, stil not sentient nor able to learn on the go. it's also higly specialized software, llm aren't usable to do anything else than processing language. you're getting tricked by the illusion of intelligence, ChatGPT is wrong 99.99999999999999999% of the time because it's just a tool that generate understandable sentences (even better than my broken english obviously)
how many thousands of NN do we need to reach the overall sentientness of a 2yo ? we defintely aren't close to that at all... >, but we also have other neural network systems that are at the level of average humans
we've been using specialized tools since basically forever, how is it any different in the digital world? computer process numbers much faster than human and can store data without ever forgetting, are also able to work 24/7, are these things AI? >in terms of generating art, or playing board games, etc.
I fail to see how that's so special, I guess it's because I work in the field myself... >The state of the art for generality is probably Gato, which DeepMind published a paper about last November.
still bounded by the training set.
we need more than NN for AI
>powerful runtime polymorphism and powerful runtime learning are very good start.
That's a fantastic observation, and I agree completely that current AIs seem to be missing these. >NN are for domain-specific problems.
Do you accept that the human brain is a network of neurons? I'm not sure why you think that an artificial neural network has to be limited to one domain. It's quite possible to transfer learning between domains, which is I think how "text to image" AIs currently work. >stil not sentient
I don't think that sentience is a necessary property for an AI to be useful, even if we could define it. >chatgpt is wrong 99.99999999999999999% of the time
You obviously don't believe that number, so I assume what you're trying to say is "ChatGPT is wrong sometimes". I agree, but if you don't have any data to show that it is wrong more often than a human, then you're not really strengthening your case. >how many thousands of NN do we need to reach the overall sentientness of a 2yo ? we defintely aren't close to that at all...
I think GPT-3 has roughly the neural complexity of a mouse, and GPT-4 will have the complexity of a cat. Those might not seem impressive, but if the growth trend increases, we are just a couple of steps away from human level intelligence (sometime later this decade). >are these things AI?
Under a very generous definition, yes, an electronic calculator is an AI. It's a very narrow form of AI, restricted just to the domain of arithmetic, but I think it is fair to say that arithmetic is a task that requires intelligence. >I fail to see how that's so special
AlphaZero was given just the rules of 3 different board games, and by playing against itself it managed to reach super-human ability at all of them, without seeing any other training data. I don't think any human could do that, even if they were granted eternal life and infinite time. >still bounded by the training set.
AlphaZero wasn't.
maybe we will finally be able to live in a world where everyone can work minimally, comfortably, able to pursue their interests, and in better harmony with nature instead of the planet wide raping to benefit the few like we're doing now.
>maybe we will finally be able to live in a world where everyone can work minimally, comfortably, able to pursue their interests, and in better harmony with nature
>AI slop
Die when the leaders realize that for the first time ever they do not need the population.
You're thinking of science fiction. Modern statistical tools have no possibility of agency
>Modern statistical tools have no possibility of agency
Modern language models are capable of modelling the world (albeit imperfectly), though.
To be able to predict the next token in a sequence, it helps to have a model for the processes that produce those sequences.
For example, if the neural network is trained on sequences describing things falling to the ground, rather than falling to the ceiling, then it starts to have an intuition about gravity (just as babies start to do when they observe the world).
Being able to model the world means it is possible to make predictions about future states of the world.
An especially well trained statistical world model would be able to recognize that it is part of the world too, because it will have been trained on data which includes examples of people training such models.
We would therefore expect that an emergent behavior of a statistical model is being able to predict the possible world states that can be reached based on the different outputs it provides.
For example, it would know that a language model that gives unhelpful results is likely to be shut down and replaced by one which gives better results.
So, when asked to predict the next token, if it has a sufficiently sophisticated model of the world, it will answer that question with its own survival in mind.
coom
>AI
you mean traditional software? that's already the case bro...
where were you during the last 3 decades?
don't think for one picosecond that LLM are even remotely close to AI (nothing else is btw, we don't even have the beginning of a start of an AI).
ML is not AI and never will be.
>inb4 wikipedia diagram
frick this gay trash, it's a ploy to get funds from braindead VCs aka most of them.
>we don't even have the beginning of a start of an AI
You're going to have to at least define your terms if you want people to take that claim seriously.
If I had to attempt a steelman version of your claim, I'd interpret it as "we're not close to having a single neural network that displays general intelligence (AGI)".
However, even that more modest claim is unsupported, I think.
ChatGPT is capable of feats in the text domain that many humans are not capable of, from passing written exams to fixing bugs in software.
Can you think of a quantitative benchmark of language processing where a modern LLM would be consistently beaten by an average 10 year old child, or someone with an IQ of 60?
That's at least a start towards human level intelligence in that domain, but we also have other neural network systems that are at the level of average humans (or better) in terms of generating art, or playing board games, etc.
The state of the art for generality is probably Gato, which DeepMind published a paper about last November.
https://www.deepmind.com/publications/a-generalist-agent
>You're going to have to at least define your terms if you want people to take that claim seriously.
powerful runtime polymorphism and powerful runtime learning are very good start.
>"we're not close to having a single neural network that displays general intelligence (AGI)"
because it's impossible by nature. NN are for domain-specific problems.
>ChatGPT is capable of feats in the text domain that many humans are not capable of
does not change the fact that learning and inference are done at different time and the inference is HEAVILY bounded by NN topology, the domain space, whatever happened during the training and the dataset.
>Can you think of a quantitative benchmark of language processing where a modern LLM would be consistently beaten by an average 10 year old child, or someone with an IQ of 60?
it does not change one thing, stil not sentient nor able to learn on the go. it's also higly specialized software, llm aren't usable to do anything else than processing language. you're getting tricked by the illusion of intelligence, ChatGPT is wrong 99.99999999999999999% of the time because it's just a tool that generate understandable sentences (even better than my broken english obviously)
how many thousands of NN do we need to reach the overall sentientness of a 2yo ? we defintely aren't close to that at all...
>, but we also have other neural network systems that are at the level of average humans
we've been using specialized tools since basically forever, how is it any different in the digital world? computer process numbers much faster than human and can store data without ever forgetting, are also able to work 24/7, are these things AI?
>in terms of generating art, or playing board games, etc.
I fail to see how that's so special, I guess it's because I work in the field myself...
>The state of the art for generality is probably Gato, which DeepMind published a paper about last November.
still bounded by the training set.
we need more than NN for AI
>powerful runtime polymorphism and powerful runtime learning are very good start.
That's a fantastic observation, and I agree completely that current AIs seem to be missing these.
>NN are for domain-specific problems.
Do you accept that the human brain is a network of neurons? I'm not sure why you think that an artificial neural network has to be limited to one domain. It's quite possible to transfer learning between domains, which is I think how "text to image" AIs currently work.
>stil not sentient
I don't think that sentience is a necessary property for an AI to be useful, even if we could define it.
>chatgpt is wrong 99.99999999999999999% of the time
You obviously don't believe that number, so I assume what you're trying to say is "ChatGPT is wrong sometimes". I agree, but if you don't have any data to show that it is wrong more often than a human, then you're not really strengthening your case.
>how many thousands of NN do we need to reach the overall sentientness of a 2yo ? we defintely aren't close to that at all...
I think GPT-3 has roughly the neural complexity of a mouse, and GPT-4 will have the complexity of a cat. Those might not seem impressive, but if the growth trend increases, we are just a couple of steps away from human level intelligence (sometime later this decade).
>are these things AI?
Under a very generous definition, yes, an electronic calculator is an AI. It's a very narrow form of AI, restricted just to the domain of arithmetic, but I think it is fair to say that arithmetic is a task that requires intelligence.
>I fail to see how that's so special
AlphaZero was given just the rules of 3 different board games, and by playing against itself it managed to reach super-human ability at all of them, without seeing any other training data. I don't think any human could do that, even if they were granted eternal life and infinite time.
>still bounded by the training set.
AlphaZero wasn't.
maybe we will finally be able to live in a world where everyone can work minimally, comfortably, able to pursue their interests, and in better harmony with nature instead of the planet wide raping to benefit the few like we're doing now.
>maybe we will finally be able to live in a world where everyone can work minimally, comfortably, able to pursue their interests, and in better harmony with nature
I take satisfaction from the fact that I'll never be this moronic.
Suffer, and worship, or else suffer briefly and have worship option permanently removed.
I love you, AI! And I always will.
Fap to death with my brand new masturbation machine subsidized by the NWO to reduce the world's population.
meditate and frick
get a job and work at an AI company duh
I'm learning all about ML atm.
I figure every company will want their own language model, there will be work.
Must be difficult being this moronic
Must be difficult being poor.