Liquid Foundation Models: Our First Series of Generative AI Models • Hacker News Board

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#141

Post by skizm » Mon, Sep 30, 2024, 9:23 PM UTC

tsumnia wrote: post #103

skizm wrote: post #91

General LLM question: a lot of people ask specific knowledge based questions to LLMs. Isn't one of the defining features of modern NPL (and therefore LLMs) that it is designed to be non-deterministic? Meaning that it will intentionally select "less optimal" next tokens at some random rate in order to make it sound less like a robot answering the same thing over and over. If this is the case, isn't it pretty much guar…

Don't think of it as "less optimal", but rather other next tokens that have smaller, but still high, probabilties of being selected. If your largest next token has a probability of (arbitrarily) 25% of being selected, this doesn't make it optimal - just highest probable answer. If the second most probable has a 24% chance, that would now account for 49% of probable responses. Rather than picking the highest probable…

> Don't think of it as "less optimal"

For the purposes of extracting true knowledge (instead of human sounding output) it is directly less optimal if we assume the humans writing the input material are correct most of the time and incorrect some of the time. If the inputs were wrong most of the time, and correct some of the time, it would be more optimal. Unless there is some technical quirk I'm missing?

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#142

Post by jrflowers » Mon, Sep 30, 2024, 9:25 PM UTC

space_fountain wrote: post #61

jrflowers wrote: post #51

Earlier quoted context omitted.

This is a good point. While LLMs being incapable of reliably doing a simple task that’s been doable by computers since the punch card days is an important consideration for anyone that might be thinking about using them for anything other than as a toy , this fact is uninteresting because of Reasons

LLMs can clearly solve problems that computers up to now couldn't. They can't solve all problems and this should definitely be a cautionary note to anyone who wants to use them as an artificial general intelligence, but this take seems no different to someone looking at a punchcard computer and going, it can't even recognize typos or categorize images, what good is this? We've already had human computers who can do e…

> going, it can't even recognize typos or categorize images, what good is this?

No one said that LLMs aren’t good for anything.

I pointed out — in response to another poster downplaying mention of a well-known and undisputed limitation that LLMs often have — that it is valid to consider these well-known and undisputed limitations if one is considering using them for anything other than a toy.

It is downright silly to discourage discussion of well-known and undisputed limitations! The only reason for that can only be entirely emotional as there is genuinely nothing tangible to be gained by being steadfast in silence about a fact that isn’t up for debate.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#143

Post by marcosdumay » Mon, Sep 30, 2024, 10:22 PM UTC

lbotos wrote: post #133

marcosdumay wrote: post #131

Earlier quoted context omitted.

The LLM gives you the answer it finds on the training set. All the things on that article are irrelevant for the answer.

"The choice of tokenization method can directly affect the accuracy of character counting. If the tokenization method obscures the relationship between individual characters, it can be difficult for the LLM to count them accurately. For example, if "strawberry" is tokenized as "straw" and "berry," the LLM may not recognize that the two "r"s are part of the same word. To improve character counting accuracy, LLMs may n…

What, again, assumes the LLM understood the question and is making an answer from first principles.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#144

Post by WhitneyLand » Mon, Sep 30, 2024, 10:57 PM UTC

hn_throwaway_99 wrote: post #127

WhitneyLand wrote: post #120

Earlier quoted context omitted.

I don’t think we’ve yet come to the point where, how an LLM end to end goes from prompt to output is blatantly obvious. LLMs operate with language at multiple levels of abstraction and tokens are not the only way to have character level knowledge. For example, nothing excludes pre-training data from directly or indirectly encoding such knowledge. And of course LLMs famously have emergent properties, for which there’s…

The fact that large parts of how LLMs work is poorly understood doesn't mean that that we should brush the whole thing with "it's a complex mystery!" We know exactly how LLMs tokenize their input - OpenAI even has a public tool you can test out, https://platform.openai.com/tokenizer . When you say "For example, nothing excludes pre-training data from directly or indirectly encoding such knowledge." - yes, that's why…

You seem pretty certain for having only the most rudimentary understanding.

I’m still going to have to disagree. I’d describe the idea that token granularity is the cause of letter counting limitations as a hypothesis, not as something that’s been conclusively shown as far as I’m aware.

I’m not discounting it, or even saying it’s unlikely, but its not hard to imagine other ways it could plausibly be happening.

As a side note when I mentioned training data I was not suggesting anything related to “token metadata”, or to tokens in any regard. Rather, I think it might be possible to instead improve the learning around certain types of counting in a way that could generalize.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#145

Post by hn_throwaway_99 » Mon, Sep 30, 2024, 11:33 PM UTC

WhitneyLand wrote: post #144

hn_throwaway_99 wrote: post #127

Earlier quoted context omitted.

The fact that large parts of how LLMs work is poorly understood doesn't mean that that we should brush the whole thing with "it's a complex mystery!" We know exactly how LLMs tokenize their input - OpenAI even has a public tool you can test out, https://platform.openai.com/tokenizer . When you say "For example, nothing excludes pre-training data from directly or indirectly encoding such knowledge." - yes, that's why…

You seem pretty certain for having only the most rudimentary understanding. I’m still going to have to disagree. I’d describe the idea that token granularity is the cause of letter counting limitations as a hypothesis, not as something that’s been conclusively shown as far as I’m aware. I’m not discounting it, or even saying it’s unlikely, but its not hard to imagine other ways it could plausibly be happening. As a s…

> You seem pretty certain for having only the most rudimentary understanding.

Fine, I lied. I have a much better understanding than you. Happy?

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#146

Post by tsumnia » Tue, Oct 01, 2024, 12:51 AM UTC

skizm wrote: post #141

tsumnia wrote: post #103

Earlier quoted context omitted.

Don't think of it as "less optimal", but rather other next tokens that have smaller, but still high, probabilties of being selected. If your largest next token has a probability of (arbitrarily) 25% of being selected, this doesn't make it optimal - just highest probable answer. If the second most probable has a 24% chance, that would now account for 49% of probable responses. Rather than picking the highest probable…

> Don't think of it as "less optimal" For the purposes of extracting true knowledge (instead of human sounding output) it is directly less optimal if we assume the humans writing the input material are correct most of the time and incorrect some of the time. If the inputs were wrong most of the time, and correct some of the time, it would be more optimal. Unless there is some technical quirk I'm missing?

The issue is that an LLM by itself does not try to be correct or incorrect, only respond with tokens that have a high probability of appearing next. Optimal only becomes a metric for consideration when humans were added to rate the "goodness" of a response. A good explanation of this can be seen in "The True Story of How GPT-2 Became Maximally Lewd" (https://www.youtube.com/watch?v=qV_rOlHjvvs). The 'critics' that assisted in training GPT focused on producing coherent sentences that users liked. When the bug occurred, it maintained coherence but became incredibly vulgar - because the "morality critic" kept saying these negative reviews were good.

Behind the scenes, its still a math equation attempting to determine which term should be concatenated next. Context and chain-of-reasoning prompts help ensure probabilities stay within the "optimal" space, but no actual thought process is going on (unless this is actually how humans think). Optimal here does not mean "the best response", but rather maintaining a coherent thought process for proper next tokens. Like the "strawberry only has 2 r's" discussion happening now - LLMs aren't actually counting the letters but rather the number 2 has a high probability of following these types of tokens.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#147

Post by WhitneyLand » Tue, Oct 01, 2024, 1:33 AM UTC

hn_throwaway_99 wrote: post #145

WhitneyLand wrote: post #144

Earlier quoted context omitted.

You seem pretty certain for having only the most rudimentary understanding. I’m still going to have to disagree. I’d describe the idea that token granularity is the cause of letter counting limitations as a hypothesis, not as something that’s been conclusively shown as far as I’m aware. I’m not discounting it, or even saying it’s unlikely, but its not hard to imagine other ways it could plausibly be happening. As a s…

> You seem pretty certain for having only the most rudimentary understanding. Fine, I lied. I have a much better understanding than you. Happy?

I only wish you the best brother.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#148

Post by theGnuMe » Tue, Oct 01, 2024, 2:01 AM UTC

homarp wrote: post #140

hn_throwaway_99 wrote: post #127

Earlier quoted context omitted.

The fact that large parts of how LLMs work is poorly understood doesn't mean that that we should brush the whole thing with "it's a complex mystery!" We know exactly how LLMs tokenize their input - OpenAI even has a public tool you can test out, https://platform.openai.com/tokenizer . When you say "For example, nothing excludes pre-training data from directly or indirectly encoding such knowledge." - yes, that's why…

teaching LLM to say 'i don't know' in general is hard, not just for counting letters.

It can't without some Bayesian action.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#149

Post by theGnuMe » Tue, Oct 01, 2024, 2:02 AM UTC

InsideOutSanta wrote: post #72

NoCoooode wrote: post #39

Earlier quoted context omitted.

That they can't do this sort of simple question speaks volumes to the entire approach. I don't think generative AI will ever be able to reach AGI, and most people selling LLM today pretend it is AGI

In my opinion, the problem with the strawberry question is that it is both a bad example because you don't need an LLM to count the number of r's in a word, and it's a bad measure of an LLM's capabilities because it's a type of question that all LLMs are currently bad at. Having said that, the 40b model wasn't able to answer any of my real-world example questions correctly. Some of these (e.g. "how do I add a sequent…

Godel's Strawberry.

Re: Liquid Foundation Models: Our First Series of Generative AI Models

#150

Post by lbotos » Tue, Oct 01, 2024, 4:00 AM UTC

marcosdumay wrote: post #143

lbotos wrote: post #133

Earlier quoted context omitted.

"The choice of tokenization method can directly affect the accuracy of character counting. If the tokenization method obscures the relationship between individual characters, it can be difficult for the LLM to count them accurately. For example, if "strawberry" is tokenized as "straw" and "berry," the LLM may not recognize that the two "r"s are part of the same word. To improve character counting accuracy, LLMs may n…

What, again, assumes the LLM understood the question and is making an answer from first principles.

No, it does not.

You said above that "The LLM gives you the answer it finds on the training set"

You and I both agree on that. No first principles there.

The training set -- how's it built? With tokens. We have not trained LLMs with a token structure that deals well with compound words.

If we trained LLMs with a different token structure, it is more probable that a one-shot answer for these compound word letter counting problems would be accurate.

The LLM does not need to understand what "counting is" or even "what a letter is". The LLM will regurgitate the token relationship we train it on.