7 ms·

If I ask you what you ate for breakfast, I would expect you to answer with a truthful answer, if I asked you whether you enjoyed it and whether it improved your

If I ask you what you ate for breakfast, I would expect you to answer with a truthful answer, if I asked you whether you enjoyed it and whether it improved your mood, again I would expect you to answer the question based on your own personal experience. GPT-3 obviously does not do this - it asks 'if the answer to the question was coming from text in the training data, what is a likely response'. Its answer is meaningless, it gives me no real data about the world. We are clearly not executing the same task.

mobiuscog 4y ago
What is the difference between 'training data' and 'memory' ?
- retrac 4y ago
  One way to look at it: Everything you have ever seen, heard or otherwise perceived is your brain's "training data" to date.
tinco 4y ago
We are only not executing the same task because GPT-3 has no motivation to do anything else than memory recall. And precisely because it does so it does give you real data about the world. I think you maybe misunderstood what GPT-3 can do, maybe you haven't played with it yourself. It not only bases the answer on information in its training set, but also on information in the question. So if you in your question give it all the information that is relevant (interspersed with irrelevant information if you want) then it will respond with its best interpretation of the information it is given, just like a human would. Treat GPT-3 like someone who just awoke from a long sleep. You could give it today's newspaper to read, and then ask it questions about today. Tell it what its own personal experience was, and then it'll talk your ears off about it if you ask nicely. It is clearly executing the same task if you disregard humans tendency to prioritize other motivations than pure memory recall.
- sweezyjeezy 4y ago
  > it will respond with its best interpretation of the information it is given, just like a human would The reason I can respond to this reply is not because I have read and memorised reams of text of people talking about AI and am simply regurgitating it mindlessly. I am able to do it because I can consider the points you are making, and what their actual 'meaning' is, try to come up with my own meaningful response and try to verbalise it back to you. There is a point where I am not just doing statistical language modelling. If you think that is wrong, and that what we are doing is closer to GPT-3, then could you explain why you think that? I am an AI researcher, and have spent plenty of time playing with GPT-3 btw.
  - tinco 4y ago
    Right for sure, I'm just trying to get to a better argument. I definitely agree that there's a great deal more consideration in your comment than in one that GPT-3 would come up with. I do wonder how far these language models can be pushed. GPT-3 seems to formulate rich models of reality to be able to extract/derive information needed to construct these answers. When I hurt your pride by suggesting you haven't fully understood GPT-3, you are motivated to come up with not just a valid response, but one that has been vetted by as many of your well developed models that form your understanding of GPT-3 so I can be suitably impressed. I'm with you that GPT-3 wouldn't go deeper than just finding some information that it thinks it's true. Though maybe GPT-3 would recognise their authority was being challenged and add that line to affirm their credentials as an AI researcher. What if GPT-3 were pushed in a similar manner, perhaps in some adversarial scheme, to not only produce information that is correct, but that is clever and exploring deep meaning, motivated by some similar feeling of pride or vindication. I think the models required to do that do not lie far from the models it needed to build to form sentences that accurately describe reality.
    - sweezyjeezy 4y ago
      Haha interesting idea - I certainly think that one issue with supervised training a la GPT is that if you ask for an opinion, the most popular one in the training data will be its default. You can start a prompt off going in a different direction, and now it says the opposite. So it's hard to conceptualise what GPT-3 actually 'believes' to be true about the world, which I feel an agent must have to be sentient. If it has all opinions stored in its memory, surely it has no belief about anything? I feel like I am capable of making a concrete decision about whether I agree with two opposing ideas in a way that a language model can't do, such as this discussion. Furthermore my belief is consistent in my outputs day to day until my mind is changed by something. If that i some complete illusion and I am just a slightly fancier autocomplete than GPT-3 - well I'd be surprised, but I can't claim to understand consciousness well enough to refute it.

mobiuscog 4y ago

What is the difference between 'training data' and 'memory' ?

retrac 4y ago
One way to look at it: Everything you have ever seen, heard or otherwise perceived is your brain's "training data" to date.

tinco 4y ago

We are only not executing the same task because GPT-3 has no motivation to do anything else than memory recall. And precisely because it does so it does give you real data about the world. I think you maybe misunderstood what GPT-3 can do, maybe you haven't played with it yourself. It not only bases the answer on information in its training set, but also on information in the question. So if you in your question give it all the information that is relevant (interspersed with irrelevant information if you want) then it will respond with its best interpretation of the information it is given, just like a human would. Treat GPT-3 like someone who just awoke from a long sleep. You could give it today's newspaper to read, and then ask it questions about today. Tell it what its own personal experience was, and then it'll talk your ears off about it if you ask nicely. It is clearly executing the same task if you disregard humans tendency to prioritize other motivations than pure memory recall.

sweezyjeezy 4y ago
> it will respond with its best interpretation of the information it is given, just like a human would The reason I can respond to this reply is not because I have read and memorised reams of text of people talking about AI and am simply regurgitating it mindlessly. I am able to do it because I can consider the points you are making, and what their actual 'meaning' is, try to come up with my own meaningful response and try to verbalise it back to you. There is a point where I am not just doing statistical language modelling. If you think that is wrong, and that what we are doing is closer to GPT-3, then could you explain why you think that? I am an AI researcher, and have spent plenty of time playing with GPT-3 btw.
- tinco 4y ago
  Right for sure, I'm just trying to get to a better argument. I definitely agree that there's a great deal more consideration in your comment than in one that GPT-3 would come up with. I do wonder how far these language models can be pushed. GPT-3 seems to formulate rich models of reality to be able to extract/derive information needed to construct these answers. When I hurt your pride by suggesting you haven't fully understood GPT-3, you are motivated to come up with not just a valid response, but one that has been vetted by as many of your well developed models that form your understanding of GPT-3 so I can be suitably impressed. I'm with you that GPT-3 wouldn't go deeper than just finding some information that it thinks it's true. Though maybe GPT-3 would recognise their authority was being challenged and add that line to affirm their credentials as an AI researcher. What if GPT-3 were pushed in a similar manner, perhaps in some adversarial scheme, to not only produce information that is correct, but that is clever and exploring deep meaning, motivated by some similar feeling of pride or vindication. I think the models required to do that do not lie far from the models it needed to build to form sentences that accurately describe reality.
  - sweezyjeezy 4y ago
    Haha interesting idea - I certainly think that one issue with supervised training a la GPT is that if you ask for an opinion, the most popular one in the training data will be its default. You can start a prompt off going in a different direction, and now it says the opposite. So it's hard to conceptualise what GPT-3 actually 'believes' to be true about the world, which I feel an agent must have to be sentient. If it has all opinions stored in its memory, surely it has no belief about anything? I feel like I am capable of making a concrete decision about whether I agree with two opposing ideas in a way that a language model can't do, such as this discussion. Furthermore my belief is consistent in my outputs day to day until my mind is changed by something. If that i some complete illusion and I am just a slightly fancier autocomplete than GPT-3 - well I'd be surprised, but I can't claim to understand consciousness well enough to refute it.