4 ms·

What does it mean to predict the next token correctly though? Arguably (non instruction tuned) models already regurgitate their training data such that it'd com

What does it mean to predict the next token correctly though? Arguably (non instruction tuned) models already regurgitate their training data such that it'd complete "Mary had a" with "little lamb" 100% of the time. On the other hand if you mean, give you the correct answer to your question 100% of the time, then I agree, though then what about things that are only in your mind (guess the number I'm thinking type problems)?

card_zero 2y ago
This highlights something that's wrong about arguments for AI. I say: it's not human-like intelligence, it's just predicting the next token probabilistically. Some AI advocate says: humans are just predicting the next token probabilistically, fight me. The problem here is that "predicting the next token probabilistically" is a way of framing any kind of cleverness, up to and including magical, impossible omniscience. That doesn't mean it's the way every kind of cleverness is actually done, or could realistically be done. And it has to be the correct next token, where all the details of what's actually required are buried in that term "correct", and sometimes it literally means the same as "likely", and other times that just produces a reasonable, excusable, intelligence-esque effort.
- dylan604 2y ago
  > Some AI advocate says: humans are just predicting the next token probabilistically, fight me. We've all had conversations with humans that are always jumping to complete your sentence assuming they know what your about to say and don't quite guess correctly. So AI evangelists are saying it's no worse than humans as their proof. I kind of like their logic. They never claimed to have built HAL /s
  - card_zero 2y ago
    No worse than a human on autopilot.
- usaar333 2y ago
  https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-general-intelligence/ https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-g... This essay has aged extremely well.
cruffle_duffle 2y ago
But now you are entering into philosophy. What does a “correct answer” even mean for a question like “is it safe to lick your fingers after using a soldering iron with leaded solder?”. I would assert that there is no “correct answer” to a question like that. Is it safe? Probably. But it depends, right? How did you handle the solder? How often are you using the solder? Were you wearing gloves? Did you wash your hands before licking your fingers? What is your age? Why are you asking the question? Did you already lick your fingers and need to know if you should see a doctor? Is it hypothetical? There is no “correct answer” to that question. Some answers are better than others, yes, but you cannot have a “correct answer”. And I did assert we are entering into philosophy and what it means to know something as well as what truth even means.
- _blk 2y ago
  Great break-down. Yes, the older you are, the safer it is. Speaking of Microsoft cooperation: I can totally see a whole series of windows 95 style popup dialogs asking you all those questions one by one in the next product iteration.
usaar333 2y ago
> What does it mean to predict the next token correctly though? Arguably (non instruction tuned) models already regurgitate their training data such that it'd complete "Mary had a" with "little lamb" 100% of the time. The unseen test data. Obviously omniscience is physically impossible. The point though is that the better and better next token prediction is, the more intelligent the system must be.

card_zero 2y ago

This highlights something that's wrong about arguments for AI. I say: it's not human-like intelligence, it's just predicting the next token probabilistically. Some AI advocate says: humans are just predicting the next token probabilistically, fight me. The problem here is that "predicting the next token probabilistically" is a way of framing any kind of cleverness, up to and including magical, impossible omniscience. That doesn't mean it's the way every kind of cleverness is actually done, or could realistically be done. And it has to be the correct next token, where all the details of what's actually required are buried in that term "correct", and sometimes it literally means the same as "likely", and other times that just produces a reasonable, excusable, intelligence-esque effort.

dylan604 2y ago
> Some AI advocate says: humans are just predicting the next token probabilistically, fight me. We've all had conversations with humans that are always jumping to complete your sentence assuming they know what your about to say and don't quite guess correctly. So AI evangelists are saying it's no worse than humans as their proof. I kind of like their logic. They never claimed to have built HAL /s
- card_zero 2y ago
  No worse than a human on autopilot.
usaar333 2y ago
https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-general-intelligence/ https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-g... This essay has aged extremely well.

cruffle_duffle 2y ago

But now you are entering into philosophy. What does a “correct answer” even mean for a question like “is it safe to lick your fingers after using a soldering iron with leaded solder?”. I would assert that there is no “correct answer” to a question like that. Is it safe? Probably. But it depends, right? How did you handle the solder? How often are you using the solder? Were you wearing gloves? Did you wash your hands before licking your fingers? What is your age? Why are you asking the question? Did you already lick your fingers and need to know if you should see a doctor? Is it hypothetical? There is no “correct answer” to that question. Some answers are better than others, yes, but you cannot have a “correct answer”. And I did assert we are entering into philosophy and what it means to know something as well as what truth even means.

_blk 2y ago
Great break-down. Yes, the older you are, the safer it is. Speaking of Microsoft cooperation: I can totally see a whole series of windows 95 style popup dialogs asking you all those questions one by one in the next product iteration.

usaar333 2y ago

> What does it mean to predict the next token correctly though? Arguably (non instruction tuned) models already regurgitate their training data such that it'd complete "Mary had a" with "little lamb" 100% of the time. The unseen test data. Obviously omniscience is physically impossible. The point though is that the better and better next token prediction is, the more intelligent the system must be.