13 ms·

The reason we chucked loads of data at it was because we had no other options. If you wanted to write a function that classified a picture as a cat or a dog, go

The reason we chucked loads of data at it was because we had no other options. If you wanted to write a function that classified a picture as a cat or a dog, good luck. With ML, you can learn such a function. That logic doesn’t extend to things we already know how to program computers to do. Arithmetic already works. We don’t need a neural net to also run the calculations or play a game of chess. We have specialized programs that are probably as good as we’re going to get in those specialized domains.

michaelt 3y ago
> We don’t need a neural net to also run the calculations or play a game of chess. That's actually one of the specific examples from the link I mentioned:- > In computer chess, the methods that defeated the world champion, Kasparov, in 1997, were based on massive, deep search. At the time, this was looked upon with dismay by the majority of computer-chess researchers who had pursued methods that leveraged human understanding of the special structure of chess. When a simpler, search-based approach with special hardware and software proved vastly more effective, these human-knowledge-based chess researchers were not good losers. They said that ``brute force" search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. These researchers wanted methods based on human input to win and were disappointed when they did not. While it's true that they didn't use an LLM specifically, it's still an example of chucking loads of compute at the problem instead of something more elegant and human-like. Of course, I agree that if you're looking for a good game of chess, Stockfish is a better choice than ChatGPT.
- janalsncm 3y ago
  What was considered “loads of compute” in 1998 is the kind of thing that can run on anyone’s phone today. Stockfish is extremely cheap compared with an LLM. Even a human-like model like Maia is tiny compared with even the smallest LLMs used these services. Point is, LLM maximalists are wrong. Specialized software is better in many places. LLMs can fill in the gaps, but should hand off when necessary.
observationist 3y ago
Not so fast - you might have precise and efficient functions that do things like basic arithmetic. What you might not have is a model that can reason mathematically. You need a model to do things like basic arithmetic functions so that semantic and arbitrary relations get encoded in the weights of a network. You see this type of glitch crop up in tokenizing schemes in large language models. If you attempt working with character level reasoning or output construction, it will often fail. Trying to get ChatGPT 4 to output a sentence, and then that sentence backwards, or every other word spelled backwards, is almost impossible. If you instead prompt the model to produce an answer with a delimiter between every character, like #, also to replace spaces, it can resolve the problems much more often than with standard punctuation and spaces. The idea applies to abstractions that aren't only individual tokens, but specific concepts and ideas that in turn serve as atomic components of higher abstractions. In order to use those concepts successfully, the model has to be able to encode the thing and its relationships effectively in the context of whatever else it learns. For a given architecture, you could do the work and manually create the encoding scheme for something like arithmetic, and it could probably be very efficient and effective. What you miss is the potential for fuzzy overlaps in the long tail that only come about through the imperfect, bespoke encodings learned in the context of your chosen optimizer.
- imtringued 3y ago
  Damn, how many problems with LLMs relate to the encoding of the token? Surely every symbolic manipulation task is getting thrown off by this. Memorizing the multiplication table of two three digit numbers is no easy task at all. That explains why the interpreter hack works so well. The python interpreter sees things digit by digit, but the LLM does it token by token. I've asked it so many times to count the number of words or letters and it was incredibly bad at it. Since it is capable of splitting large tokens into smaller tokens, the solution to this problem is to create additional training samples that perform "big token" to "small token" conversion and back, so that the model will learn to dynamically provide the most suitable encoding to itself.

michaelt 3y ago

> We don’t need a neural net to also run the calculations or play a game of chess. That's actually one of the specific examples from the link I mentioned:- > In computer chess, the methods that defeated the world champion, Kasparov, in 1997, were based on massive, deep search. At the time, this was looked upon with dismay by the majority of computer-chess researchers who had pursued methods that leveraged human understanding of the special structure of chess. When a simpler, search-based approach with special hardware and software proved vastly more effective, these human-knowledge-based chess researchers were not good losers. They said that ``brute force" search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. These researchers wanted methods based on human input to win and were disappointed when they did not. While it's true that they didn't use an LLM specifically, it's still an example of chucking loads of compute at the problem instead of something more elegant and human-like. Of course, I agree that if you're looking for a good game of chess, Stockfish is a better choice than ChatGPT.

janalsncm 3y ago
What was considered “loads of compute” in 1998 is the kind of thing that can run on anyone’s phone today. Stockfish is extremely cheap compared with an LLM. Even a human-like model like Maia is tiny compared with even the smallest LLMs used these services. Point is, LLM maximalists are wrong. Specialized software is better in many places. LLMs can fill in the gaps, but should hand off when necessary.

observationist 3y ago

Not so fast - you might have precise and efficient functions that do things like basic arithmetic. What you might not have is a model that can reason mathematically. You need a model to do things like basic arithmetic functions so that semantic and arbitrary relations get encoded in the weights of a network. You see this type of glitch crop up in tokenizing schemes in large language models. If you attempt working with character level reasoning or output construction, it will often fail. Trying to get ChatGPT 4 to output a sentence, and then that sentence backwards, or every other word spelled backwards, is almost impossible. If you instead prompt the model to produce an answer with a delimiter between every character, like #, also to replace spaces, it can resolve the problems much more often than with standard punctuation and spaces. The idea applies to abstractions that aren't only individual tokens, but specific concepts and ideas that in turn serve as atomic components of higher abstractions. In order to use those concepts successfully, the model has to be able to encode the thing and its relationships effectively in the context of whatever else it learns. For a given architecture, you could do the work and manually create the encoding scheme for something like arithmetic, and it could probably be very efficient and effective. What you miss is the potential for fuzzy overlaps in the long tail that only come about through the imperfect, bespoke encodings learned in the context of your chosen optimizer.

imtringued 3y ago
Damn, how many problems with LLMs relate to the encoding of the token? Surely every symbolic manipulation task is getting thrown off by this. Memorizing the multiplication table of two three digit numbers is no easy task at all. That explains why the interpreter hack works so well. The python interpreter sees things digit by digit, but the LLM does it token by token. I've asked it so many times to count the number of words or letters and it was incredibly bad at it. Since it is capable of splitting large tokens into smaller tokens, the solution to this problem is to create additional training samples that perform "big token" to "small token" conversion and back, so that the model will learn to dynamically provide the most suitable encoding to itself.