After reading this, I’m not sure how much of a threat, or a help, GPT-N would be. Let’s say we have GPT-N, trained on human text, and GPT-N is an AGI. I ask it “You are a superintelligent misaligned AI—how should you take over the world?”
GPT-N, to my understanding, would not then pretend to be a superintelligent misaligned AI and output a plan that the AI would output, even if it is theoretically capable of doing so. It would pretend to be a human pretending to be a superintelligent misaligned AI, because human data is what its training corpus was built on.
This would also be a blow towards GPT-N helping with alignment research, for similar reasons. It seems like we’d need some sort of ELK-like interpretability to get it to tell us things a human never would.
It seems like we’d need some sort of ELK-like interpretability to get it to tell us things a human never would.
Not really, we’d just need to condition GPT-N in more clever ways. For instance by tagging all scientific publications in its dataset with a particular token, also giving it the publication date and the number of citations for every paper. Then you just need to prompt it with the scientific paper token, a future date and a high number of citations to make GPT-N try to simulate the future progress of humanity on the particular scientific question you’re interested in.
So, if I’m understanding this right, we could fine-tune GPT-N in different ways. For instance, we can currently fine-tune GPT-3 to predict whether a movie review was positive or not. Similarly, we could fine-tune GPT-N for some sort of “Plausible science score” and then try to maximise that score in the year 2040, which would lead to a paper that GPT-N would consider maximally plausible as a blah studies paper in the year 2040. For a sufficiently powerful GPT-N, this would lead to actual scientific advancement, especially since we wouldn’t need anywhere close to a 100% hit rate for this to be effective.
In fact, we could do all of this right now, it’s just that GPT-3 isn’t powerful enough to produce actual scientific advancement and would instead create legible-sounding examples that didn’t actually bear up, or probably even have a truly coherent, detailed idea behind them.
“fine-tuning” isn’t quite the right word for this. Right now GPT-3 is trained by being given a sequence of words like <token1><token2><token3> … <TokenN>, and it’s trained to predict the next token. What I’m saying is that we can, for each piece of text that we use in the training set, look at its date of publication and provenance, and we can train a new GPT-3 where instead of just being given the tokens, we give it <date of publication><is scientific publication?><author><token1><token2>...<tokenN>. And then at inference time, we can choose <date of publication=2040> to make it simulate future progress.
Basically all human text containing the words “publication 2040” is science-fiction, and we want to avoid the model writing fiction by giving it data that helps it disambiguate fiction about the future and actual future text. If we give it a correct ground truth about the publication date of every one of its training data strings, then it would be forced to actually extrapolate its knowledge into the future. Similarly most discussions of future tech are done by amateurs, or again in science-fiction, but giving it the correct ground truth about the actual journal of publication avoids all of that. GPT only needs to predict that Nature won’t become a crank journal in 20 years, and it will then make an actual effort at producing high-impact scientific publications.
After reading this, I’m not sure how much of a threat, or a help, GPT-N would be. Let’s say we have GPT-N, trained on human text, and GPT-N is an AGI. I ask it “You are a superintelligent misaligned AI—how should you take over the world?”
GPT-N, to my understanding, would not then pretend to be a superintelligent misaligned AI and output a plan that the AI would output, even if it is theoretically capable of doing so. It would pretend to be a human pretending to be a superintelligent misaligned AI, because human data is what its training corpus was built on.
This would also be a blow towards GPT-N helping with alignment research, for similar reasons. It seems like we’d need some sort of ELK-like interpretability to get it to tell us things a human never would.
Does this seem accurate?
Not really, we’d just need to condition GPT-N in more clever ways. For instance by tagging all scientific publications in its dataset with a particular token, also giving it the publication date and the number of citations for every paper. Then you just need to prompt it with the scientific paper token, a future date and a high number of citations to make GPT-N try to simulate the future progress of humanity on the particular scientific question you’re interested in.
So, if I’m understanding this right, we could fine-tune GPT-N in different ways. For instance, we can currently fine-tune GPT-3 to predict whether a movie review was positive or not. Similarly, we could fine-tune GPT-N for some sort of “Plausible science score” and then try to maximise that score in the year 2040, which would lead to a paper that GPT-N would consider maximally plausible as a blah studies paper in the year 2040. For a sufficiently powerful GPT-N, this would lead to actual scientific advancement, especially since we wouldn’t need anywhere close to a 100% hit rate for this to be effective.
In fact, we could do all of this right now, it’s just that GPT-3 isn’t powerful enough to produce actual scientific advancement and would instead create legible-sounding examples that didn’t actually bear up, or probably even have a truly coherent, detailed idea behind them.
“fine-tuning” isn’t quite the right word for this. Right now GPT-3 is trained by being given a sequence of words like <token1><token2><token3> … <TokenN>, and it’s trained to predict the next token. What I’m saying is that we can, for each piece of text that we use in the training set, look at its date of publication and provenance, and we can train a new GPT-3 where instead of just being given the tokens, we give it <date of publication><is scientific publication?><author><token1><token2>...<tokenN>. And then at inference time, we can choose <date of publication=2040> to make it simulate future progress.
Basically all human text containing the words “publication 2040” is science-fiction, and we want to avoid the model writing fiction by giving it data that helps it disambiguate fiction about the future and actual future text. If we give it a correct ground truth about the publication date of every one of its training data strings, then it would be forced to actually extrapolate its knowledge into the future. Similarly most discussions of future tech are done by amateurs, or again in science-fiction, but giving it the correct ground truth about the actual journal of publication avoids all of that. GPT only needs to predict that Nature won’t become a crank journal in 20 years, and it will then make an actual effort at producing high-impact scientific publications.