4 ms·

Makes me wonder what kind of progress they've had internally at OpenAI to be making such claims. So far I haven't even seen any particularly useful or compelli

Makes me wonder what kind of progress they've had internally at OpenAI to be making such claims. So far I haven't even seen any particularly useful or compelling definitions of AGI; it's usually handwavy, incredibly vague, or defined in terms of something else which just shifts the problem around. The general sentiment I've gotten from following a bunch of AI researchers on Twitter is that we're still multiple key insights away from being able to build AGI. So either Sam knows something we don't or he's hoping to we get lucky in the next 2 years. I hope he's wrong and we actually get AGI by 2024.

ilaksh 3y ago
You didn't define AGI either.
- TheAceOfHearts 3y ago
  I didn't define AGI because I'm also confused about the term. As the term is currently used, it seems to directionally gesture vaguely towards something that's mostly based on gut-based sensing. Maybe it's one of those "when you see it you'll know" kind of things. Or maybe we'll have to experiment with building it first before we can narrow down the exact boundaries of how it's defined.
pointlessone 3y ago
There’s a relatively mainstream definition of AGI: an AGI could learn to accomplish any intellectual task that human beings or animals can perform. Sometimes agency is included in there to facilitate the learning and accomplishment. I think it’s as good a definition for a major milestone in AI development as any.
- dr_dshiv 3y ago
  The other mainstream definition comes from Peter Norvig , author of the most mainstream textbook on artificial intelligence (or at least the most cited). He defines AGI relative to narrow AI, using the development of general purpose digital computing as an analogy. Before general purpose digital computing, there were a wide range of specialized narrow forms of computing. Thus, he concludes that AGI is already here (it is similarly general purpose) but in the early days. I find his view pragmatic and helpful. https://www.noemamag.com/artificial-general-intelligence-is-already-here https://www.noemamag.com/artificial-general-intelligence-is-...
  - pointlessone 3y ago
    I don't buy this argument. Specifically, Modalities. The claim is that "[b]y using modality-specific tokenizers or processing raw data streams, frontier models can, in principle, handle any known sensory or motor modality". Sure, you may get some sort of output but I'm quite confident you won't get anything useful without training the model on that kind of input. We barely started sticking narrow AI together. E.g. ChatGPT (LLM) can invoke Dall-e (text to image). Maybe AGI will turn out to be a bunch of narrow AIs in a trench coat. So far I haven't seen such a system. Tasks claim is a little stretched, too. For example, simple arithmetic. Does model actually do arithmetic or does it do text generation and the right answer just happens to be the most likely next token? Can we reliably tell one from the other to claim that model actually perform tasks other than just text generation?
- bjorglo 3y ago
  [dead]
- TheAceOfHearts 3y ago
  I'm familiar with this definition but find it lacking. It has utility as a working definition for gesturing towards a desirable goal. My problem with such a definition is that it doesn't clarify anything, it basically shifts the problem over to one where human or animal intelligence is treated as a magical black box. Sorry if this comes off as a bit pedantic, I understand that research into new and unknown fields can take a while before landing on more accurate and unambiguous definitions.
  - pointlessone 3y ago
    I think of it a little differently. Human or animal intelligence here is a somewhat useful reference point. As an analogy, think of the first steam engines. They were replacing horses so why not measure their primary function in units of horse power? We had a somewhat rigorous physical framework for work and power but it probably wasn't well understood by the target audience to be useful. While we don't seem to have a scientifically rigorous definition of intelligence we still can have a somewhat useful discussions about utility in reference to another familiar instance of intelligence.
bamboozled 3y ago
Makes me wonder what kind of progress they've had internally at OpenAI to be making such claims This is by design. I think this is why he is the CEO, he is a master of deception, half truths, decoys and promises. You never know what he knows so bad fucking luck, that’s his attitude. Even if he isn’t lying , he is communicating in a non-friendly, secretive and manipulative way which is toxic. AGI, something better than humans at pretty much ahy tasks would have some radical implications for society. Him throwing this around loosely is irresponsible and shows a lack of respect or empathy on many fronts. If OpenAI are trying to develop it, fine, but just so some maturity as well.

ilaksh 3y ago

You didn't define AGI either.

TheAceOfHearts 3y ago
I didn't define AGI because I'm also confused about the term. As the term is currently used, it seems to directionally gesture vaguely towards something that's mostly based on gut-based sensing. Maybe it's one of those "when you see it you'll know" kind of things. Or maybe we'll have to experiment with building it first before we can narrow down the exact boundaries of how it's defined.

pointlessone 3y ago

There’s a relatively mainstream definition of AGI: an AGI could learn to accomplish any intellectual task that human beings or animals can perform. Sometimes agency is included in there to facilitate the learning and accomplishment. I think it’s as good a definition for a major milestone in AI development as any.

dr_dshiv 3y ago
The other mainstream definition comes from Peter Norvig , author of the most mainstream textbook on artificial intelligence (or at least the most cited). He defines AGI relative to narrow AI, using the development of general purpose digital computing as an analogy. Before general purpose digital computing, there were a wide range of specialized narrow forms of computing. Thus, he concludes that AGI is already here (it is similarly general purpose) but in the early days. I find his view pragmatic and helpful. https://www.noemamag.com/artificial-general-intelligence-is-already-here https://www.noemamag.com/artificial-general-intelligence-is-...
- pointlessone 3y ago
  I don't buy this argument. Specifically, Modalities. The claim is that "[b]y using modality-specific tokenizers or processing raw data streams, frontier models can, in principle, handle any known sensory or motor modality". Sure, you may get some sort of output but I'm quite confident you won't get anything useful without training the model on that kind of input. We barely started sticking narrow AI together. E.g. ChatGPT (LLM) can invoke Dall-e (text to image). Maybe AGI will turn out to be a bunch of narrow AIs in a trench coat. So far I haven't seen such a system. Tasks claim is a little stretched, too. For example, simple arithmetic. Does model actually do arithmetic or does it do text generation and the right answer just happens to be the most likely next token? Can we reliably tell one from the other to claim that model actually perform tasks other than just text generation?
bjorglo 3y ago
[dead]
TheAceOfHearts 3y ago
I'm familiar with this definition but find it lacking. It has utility as a working definition for gesturing towards a desirable goal. My problem with such a definition is that it doesn't clarify anything, it basically shifts the problem over to one where human or animal intelligence is treated as a magical black box. Sorry if this comes off as a bit pedantic, I understand that research into new and unknown fields can take a while before landing on more accurate and unambiguous definitions.
- pointlessone 3y ago
  I think of it a little differently. Human or animal intelligence here is a somewhat useful reference point. As an analogy, think of the first steam engines. They were replacing horses so why not measure their primary function in units of horse power? We had a somewhat rigorous physical framework for work and power but it probably wasn't well understood by the target audience to be useful. While we don't seem to have a scientifically rigorous definition of intelligence we still can have a somewhat useful discussions about utility in reference to another familiar instance of intelligence.

bamboozled 3y ago

Makes me wonder what kind of progress they've had internally at OpenAI to be making such claims This is by design. I think this is why he is the CEO, he is a master of deception, half truths, decoys and promises. You never know what he knows so bad fucking luck, that’s his attitude. Even if he isn’t lying , he is communicating in a non-friendly, secretive and manipulative way which is toxic. AGI, something better than humans at pretty much ahy tasks would have some radical implications for society. Him throwing this around loosely is irresponsible and shows a lack of respect or empathy on many fronts. If OpenAI are trying to develop it, fine, but just so some maturity as well.