Well, the big, obvious, enormous difference between current deep models and the human brain is that the brain uses WAY more compute. Even the most generous estimates put GPT-3 at using something like 1000x less compute than the brain. OpenAI demonstrated, quite decisively, that increasing model size leads to increasing performance.
Also, generating dialog in video games is NOT trivial (and is well beyond GPT-3′s capabilities). Any AI capable of that would be enormously valuable, since it would need a generalized grasp of language close to human-level proficiency and could be adapted to many text generation tasks (novel/script writing, customer service, chatbot based substitutes for companionship, etc).
I think the big, obvious, enormous difference between GPT-3 and the human brain is that GPT-3 isn’t an agent. It’s not trained for behavior; it’s adjusted for accuracy. It doesn’t even have any agency in choosing its input; it’s given a big wodge of training data, and has to ingest it. It has less agency than a slug, and therefore can’t really learn to do anything “agenty”.
I mean, I could be wrong. Maybe it could do something interesting given 1000 times more compute. But it seems unlikely enough that it doesn’t worry me. Things like DeepMind’s generalized game-playing agents are a lot scarier to me.
Also, generating dialog in video games is NOT trivial (and is well beyond GPT-3′s capabilities).
As I understand it, it’s actually generating dialog in commercial games today. There was some kind of big flap about people getting it to generate text involving children and sex.
But I didn’t mean “trivial” in the sense of “easy”. I meant “trivial” in the sense of “doesn’t really matter in the grand scheme of things”. Even if you add all the applications you listed, it’s still not a big deal by the standards of people who talk about X-risk.
I think the big, obvious, enormous difference between GPT-3 and the human brain is that GPT-3 isn’t an agent. It’s not trained for behavior; it’s adjusted for accuracy.
It’s true that GPT-3 doesn’t do everything that a human brain does, but one of my thoughts when reading Duncan’s post on shoulder advisors was that it really sounds like the brain runs something like GPT-? instances that can be trained on various prediction tasks.
It doesn’t even have any agency in choosing its input; it’s given a big wodge of training data, and has to ingest it. It has less agency than a slug, and therefore can’t really learn to do anything “agenty”.
It’s quite trival to change it in a way where it’s output feeds back into it’s input given that it’s input is text and it’s output is text.
You can make the output console comments and then feed the resulting console answer back into the model. It likely needs a larger attention fields to be practically useful but more compute and clever ways to handle it could lead there.
Our own thinking process is also a lot about having a short term memory into which we put new thoughts and based on which our next action/thought gets generated.
Well, the big, obvious, enormous difference between current deep models and the human brain is that the brain uses WAY more compute. Even the most generous estimates put GPT-3 at using something like 1000x less compute than the brain. OpenAI demonstrated, quite decisively, that increasing model size leads to increasing performance.
Also, generating dialog in video games is NOT trivial (and is well beyond GPT-3′s capabilities). Any AI capable of that would be enormously valuable, since it would need a generalized grasp of language close to human-level proficiency and could be adapted to many text generation tasks (novel/script writing, customer service, chatbot based substitutes for companionship, etc).
I think the big, obvious, enormous difference between GPT-3 and the human brain is that GPT-3 isn’t an agent. It’s not trained for behavior; it’s adjusted for accuracy. It doesn’t even have any agency in choosing its input; it’s given a big wodge of training data, and has to ingest it. It has less agency than a slug, and therefore can’t really learn to do anything “agenty”.
I mean, I could be wrong. Maybe it could do something interesting given 1000 times more compute. But it seems unlikely enough that it doesn’t worry me. Things like DeepMind’s generalized game-playing agents are a lot scarier to me.
As I understand it, it’s actually generating dialog in commercial games today. There was some kind of big flap about people getting it to generate text involving children and sex.
But I didn’t mean “trivial” in the sense of “easy”. I meant “trivial” in the sense of “doesn’t really matter in the grand scheme of things”. Even if you add all the applications you listed, it’s still not a big deal by the standards of people who talk about X-risk.
It’s true that GPT-3 doesn’t do everything that a human brain does, but one of my thoughts when reading Duncan’s post on shoulder advisors was that it really sounds like the brain runs something like GPT-? instances that can be trained on various prediction tasks.
Something of an side, but what exactly is your definition of ‘agent’?
Depends on the context. :)
If I had to give a general definition, something like “a system whose behavior can usefully be predicted through the intentional stance”.
It’s quite trival to change it in a way where it’s output feeds back into it’s input given that it’s input is text and it’s output is text.
You can make the output console comments and then feed the resulting console answer back into the model. It likely needs a larger attention fields to be practically useful but more compute and clever ways to handle it could lead there.
Our own thinking process is also a lot about having a short term memory into which we put new thoughts and based on which our next action/thought gets generated.