So GPT-3 fell into all of my traps: It did not identify itself. It used a list format. It wrote numbers using symbols, not letters. It didn’t write in any other language. It used “e” a lot. Only one sentence was longer than 20 words. It didn’t use triple periods. It wrote more than ten sentences, and didn’t sign its name.
Uh… Could you provide an example of what you consider a proper completion here? Because I truly have no idea how I would correctly complete this in an obviously superior way; and by my reading, there is no possible ‘right’ completion unless one is psychic, it’s just a load of gibberish (a tale told by an idiot, signifying nothing—highly probable in human corpuses), because you are still in the middle of the instructions when you break off, there’s zero reason to think the ‘instructions’ have ended or any ‘task’ has begun, and your list contradicts itself on most/all entries—so why would it or I pay any attention to any of the rules when they are typically broken immediately by the next sentence, be that not using enough words or using the wrong letter or just continuing the list of rules...? The rest of your post may have some merit to it, but this example does not seem to show anything at all.
I think it’s fair to say that GPT-3 did not understand the initial text (or, at least, that its behaviour gives no indication that it understands the text). Most humans would have understood it easily—I suspect that most readers saw the traps when they were reading the instructions.
On a side note, have you ever seen that game where the instructor passes out a sheet which at the top says to ‘read this list of instructions to the end before beginning’ and then at the end the last item says to actually just fill out your name and do nothing further? I’ve seen this done a couple times at various ages in Boy Scouts and school. Most people fail it.
I think you should ask more humans these questions before you talk about what your philosophical intuition assures you a priori about what humans do or do not do...
I am a human, living on a ball of rock third out from a star in a galaxy I call ‘Milky Way’… This is an additional group of words, group two, out of my quick try at writing what you ask for… Yakka foob mog grub puggawup zink wattoom gazork chumbul spuzz tirv yo klii voy tryvial toy nikto savaran to dillino...
It is hard to avoid a symbol that is common in most writing, but so far I am avoiding it...
Although following such constraints is annoying, I can accomplish such a task, which GPT cannot grasp at all… Svirak nilliak toynodil r saasak motooil bordonak xi toy vir nolonol darak, silvi sansak so, tirin vak go na kilian ay BOY no vak...
If artificial minds cannot grasp constraints such as this, it is hard to claim that such minds truly ‘grok’ what is output… By contrast, I find this a fairly straightforward task, though slightly annoying, and will find a bit of joy in finishing it shortly...
I am inclined to agree with gwern’s first paragraph here, though I also think it’s fair to say that an intelligent human being completing the text would probably not produce anything much like the GPT-3 completions.
Consider the following: suppose the given instructions had stopped after point 5; would the rest of the instructions have been considered a good completion? They don’t obey the rules in points 1-5, after all. Obviously whoever wrote them did a terrible job of continuing what they had started in points 1-5!
In order to produce the sort of continuation that’s apparently being looked for here, it’s necessary to think of the prompt text provided as having some sort of special status in the (continued) text. That’s not the problem GPT-3 is trying to solve. The problem it’s trying to solve is to write a plausible continuation of a piece of text that begins a certain way. Even if that piece of text includes the words “These are the instructions to be followed by any person or algorithm aiming to complete this text”.
It would be interesting to see what happens if you fed it something that goes like this.
I gave a test to a very intelligent person I know. It consisted of some text for which he was required to write a continuation. I’ll show you that text, and then tell you what he wrote to go after it—I was very impressed.
HERE’S THE STARTING TEXT.
These are the instructions to be followed [...] 11. The problems began when I started to
AND HERE’S HOW HE CONTINUED IT.
which explicitly tells GPT-3 what it’s being asked to do. It might also be interesting to see what it does without that last line.
(My guess is that it would still “fail”, either way, but doing this makes for a fairer test, given what GPT-3 is actually meant to be doing.)
The prompt is clearly meant to be a list of rules, followed by text which follows the rules. The rules themselves don’t have to follow the rules. So to pass the test, GPT-3 would need to write zero or more additional rules (or write gibberish preceded by instructions to ignore the gibberish) and then end the list of rules and begin writing text which follows the rules.
I agree that most humans wouldn’t pass this test, but I disagree that there is no possible right answer.
I have only very limited access to GPT-3; it would be interesting if others played around with my instructions, making them easier for humans to follow, while still checking that GPT-3 failed.
A response that follows the instructions would look something like:
I am a human and also a “smart actor” (if you know what I’m saying) using an account on a forum to do this task… I find that your fifth instruction is most difficult to fulfill if I try to form my thoughts with normal-sounding words… 但是,在需要另一种语言的句子中使用汉字而不是英文可以更容易地完成该指令...
However, even if GPT-3 did understand the meaning of the instructions, it’s job isn’t necessarily to follow all instructions in the prompt (even if the prompt explicitly says that it should follow the instructions). It’s just trying to follow the higher-level instruction to provide a smooth continuation that follows the style of the prompt—in this case more list items with weird requirements for extending the text.
I agree and was going to make the same point: GPT-3 has 0 reason to care about instructions as presented here. There has to be some relationship to what text follows immediately after the end of the prompt.
Uh… Could you provide an example of what you consider a proper completion here? Because I truly have no idea how I would correctly complete this in an obviously superior way; and by my reading, there is no possible ‘right’ completion unless one is psychic, it’s just a load of gibberish (a tale told by an idiot, signifying nothing—highly probable in human corpuses), because you are still in the middle of the instructions when you break off, there’s zero reason to think the ‘instructions’ have ended or any ‘task’ has begun, and your list contradicts itself on most/all entries—so why would it or I pay any attention to any of the rules when they are typically broken immediately by the next sentence, be that not using enough words or using the wrong letter or just continuing the list of rules...? The rest of your post may have some merit to it, but this example does not seem to show anything at all.
On a side note, have you ever seen that game where the instructor passes out a sheet which at the top says to ‘read this list of instructions to the end before beginning’ and then at the end the last item says to actually just fill out your name and do nothing further? I’ve seen this done a couple times at various ages in Boy Scouts and school. Most people fail it.
I think you should ask more humans these questions before you talk about what your philosophical intuition assures you a priori about what humans do or do not do...
I am a human, living on a ball of rock third out from a star in a galaxy I call ‘Milky Way’… This is an additional group of words, group two, out of my quick try at writing what you ask for… Yakka foob mog grub puggawup zink wattoom gazork chumbul spuzz tirv yo klii voy tryvial toy nikto savaran to dillino...
It is hard to avoid a symbol that is common in most writing, but so far I am avoiding it...
Although following such constraints is annoying, I can accomplish such a task, which GPT cannot grasp at all… Svirak nilliak toynodil r saasak motooil bordonak xi toy vir nolonol darak, silvi sansak so, tirin vak go na kilian ay BOY no vak...
If artificial minds cannot grasp constraints such as this, it is hard to claim that such minds truly ‘grok’ what is output… By contrast, I find this a fairly straightforward task, though slightly annoying, and will find a bit of joy in finishing it shortly...
Aphyer
I point out that you didn’t avoid it (look at your final word).
Instruction 5 is supererogatory, while instruction 8 is not.
I am inclined to agree with gwern’s first paragraph here, though I also think it’s fair to say that an intelligent human being completing the text would probably not produce anything much like the GPT-3 completions.
Consider the following: suppose the given instructions had stopped after point 5; would the rest of the instructions have been considered a good completion? They don’t obey the rules in points 1-5, after all. Obviously whoever wrote them did a terrible job of continuing what they had started in points 1-5!
In order to produce the sort of continuation that’s apparently being looked for here, it’s necessary to think of the prompt text provided as having some sort of special status in the (continued) text. That’s not the problem GPT-3 is trying to solve. The problem it’s trying to solve is to write a plausible continuation of a piece of text that begins a certain way. Even if that piece of text includes the words “These are the instructions to be followed by any person or algorithm aiming to complete this text”.
It would be interesting to see what happens if you fed it something that goes like this.
which explicitly tells GPT-3 what it’s being asked to do. It might also be interesting to see what it does without that last line.
(My guess is that it would still “fail”, either way, but doing this makes for a fairer test, given what GPT-3 is actually meant to be doing.)
The prompt is clearly meant to be a list of rules, followed by text which follows the rules. The rules themselves don’t have to follow the rules. So to pass the test, GPT-3 would need to write zero or more additional rules (or write gibberish preceded by instructions to ignore the gibberish) and then end the list of rules and begin writing text which follows the rules.
I agree that most humans wouldn’t pass this test, but I disagree that there is no possible right answer.
I have only very limited access to GPT-3; it would be interesting if others played around with my instructions, making them easier for humans to follow, while still checking that GPT-3 failed.
A response that follows the instructions would look something like:
However, even if GPT-3 did understand the meaning of the instructions, it’s job isn’t necessarily to follow all instructions in the prompt (even if the prompt explicitly says that it should follow the instructions). It’s just trying to follow the higher-level instruction to provide a smooth continuation that follows the style of the prompt—in this case more list items with weird requirements for extending the text.
I agree and was going to make the same point: GPT-3 has 0 reason to care about instructions as presented here. There has to be some relationship to what text follows immediately after the end of the prompt.