People who think that LLMs having trouble with these questions is evidence one way or another about how good or bad LLMs are just don’t understand tokenization. This is not a symptom of some big-picture deep problem with LLMs; it’s a curious artifact like in a jpeg image, but doesn’t really matter for the vast majority of applications.
You may hate AI but that doesn’t excuse being ignorant about how it works.
These sorts of artifacts wouldn’t be a huge issue except that AI is being pushed to the general public as an alternative means of learning basic information. The meme example is obvious to someone with a strong understanding of English but learners and children might get an artifact and stamp it in their memory, working for years off bad information. Not a problem for a few false things every now and then, that’s unavoidable in learning. Thousands accumulated over long term use, however, and your understanding of the world will be coarser, like the Swiss cheese with voids so large it can’t hold itself up.
You’re talking about hallucinations. That’s different from tokenization reflection errors. I’m specifically talking about its inability to know how many of a certain type of letter are in a word that it can spell correctly. This is not a hallucination per se – at least, it’s a completely different mechanism that causes it than whatever causes other factual errors. This specific problem is due to tokenization, and that’s why I say it has little bearing on other shortcomings of LLMs.
No, I’m talking about human learning and the danger imposed by treating an imperfect tool as a reliable source of information as these companies want people to do.
Whether the erratic information is from tokenization or hallucinations is irrelevant when this is already the main source for so many people in their learning, for example, a new language.
Hallucinations aren’t relevant to my point here. I’m not defending that AIs are a good source of information, and I agree that hallucinations are dangerous (either that or misusing LLMs is dangerous). I also admit that for language learning, artifacts caused from tokenization could be very detrimental to the user.
The point I am making is that LLMs struggling with these kind of tokenization artifacts is poor evidence for drawing any conclusions about their behaviour on other tasks.
That’s a fair point when these LLMs are restricted to areas where they function well. They have use cases that make sense when isolated from the ethics around training and compute. But the people who made them are applying them wildly outside these use cases.
These are pushed as a solution to every problem for the sake of profit with intentional ignorance of these issues. If a few errors impact someone it’s just a casualty in the goal of making it profitable. That can’t be disentwined from them unless you limit your argument to open source local compute.
Well – and I don’t meant this to be antagonistic – I agree with everything you’ve said except for the last sentence where you say “and therefore you’re wrong.” Look, I’m not saying LLMs function well, or that they’re good for society, or anything like that. I’m saying that tokenization errors are really their own thing that are unrelated to other errors LLMs make. If you want to dunk on LLMs then yeah be my guest. I’m just saying that this one type of poor behaviour is unrelated to the other kinds of poor behaviour.
Also just checked and every open ai model bigger than 4.1-mini can answer this. I think the joke should emphasize how we developed a super power inefficient way to solve some problems that can be accurately and efficiently answered with a single algorithm. Another example is using ChatGPT to do simple calculator math. LLMs are good at specific tasks and really bad at others, but people kinda throw everything at them.
what do you mean by spell fine? They’re just emitting the tokens for the words. Like, it’s not writing “strawberry,” it’s writing tokens <302, 1618, 19772>, which correspond to st, raw, and berry respectively. If you ask it to put a space between each letter, that will disrupt the tokenization mechanism, and it’s going to be quite liable to making mistakes.
I don’t think it’s really fair to say that the lookup 19772 -> berry counts as the LLM being able to spell, since the LLM isn’t operating at that layer. It doesn’t really emit letters directly. I would argue its inability to reliably spell words when you force it to go letter-by-letter or answer queries about how words are spelled is indicative of its poor ability to spell.
The problem is that it’s not actually counting anything. It’s simply looking for some text somewhere in its database that relates to that word and the number of R’s in that word. There’s no mechanism within the LLM to actually count things. It is not designed with that function. This is not general AI, this is a Generative Adversarial Network that’s using its vast vast store of text to put words together that sound like they answer the question that was asked.
People who think that LLMs having trouble with these questions is evidence one way or another about how good or bad LLMs are just don’t understand tokenization. This is not a symptom of some big-picture deep problem with LLMs; it’s a curious artifact like in a jpeg image, but doesn’t really matter for the vast majority of applications.
You may hate AI but that doesn’t excuse being ignorant about how it works.
These sorts of artifacts wouldn’t be a huge issue except that AI is being pushed to the general public as an alternative means of learning basic information. The meme example is obvious to someone with a strong understanding of English but learners and children might get an artifact and stamp it in their memory, working for years off bad information. Not a problem for a few false things every now and then, that’s unavoidable in learning. Thousands accumulated over long term use, however, and your understanding of the world will be coarser, like the Swiss cheese with voids so large it can’t hold itself up.
You’re talking about hallucinations. That’s different from tokenization reflection errors. I’m specifically talking about its inability to know how many of a certain type of letter are in a word that it can spell correctly. This is not a hallucination per se – at least, it’s a completely different mechanism that causes it than whatever causes other factual errors. This specific problem is due to tokenization, and that’s why I say it has little bearing on other shortcomings of LLMs.
No, I’m talking about human learning and the danger imposed by treating an imperfect tool as a reliable source of information as these companies want people to do.
Whether the erratic information is from tokenization or hallucinations is irrelevant when this is already the main source for so many people in their learning, for example, a new language.
Hallucinations aren’t relevant to my point here. I’m not defending that AIs are a good source of information, and I agree that hallucinations are dangerous (either that or misusing LLMs is dangerous). I also admit that for language learning, artifacts caused from tokenization could be very detrimental to the user.
The point I am making is that LLMs struggling with these kind of tokenization artifacts is poor evidence for drawing any conclusions about their behaviour on other tasks.
That’s a fair point when these LLMs are restricted to areas where they function well. They have use cases that make sense when isolated from the ethics around training and compute. But the people who made them are applying them wildly outside these use cases.
These are pushed as a solution to every problem for the sake of profit with intentional ignorance of these issues. If a few errors impact someone it’s just a casualty in the goal of making it profitable. That can’t be disentwined from them unless you limit your argument to open source local compute.
Well – and I don’t meant this to be antagonistic – I agree with everything you’ve said except for the last sentence where you say “and therefore you’re wrong.” Look, I’m not saying LLMs function well, or that they’re good for society, or anything like that. I’m saying that tokenization errors are really their own thing that are unrelated to other errors LLMs make. If you want to dunk on LLMs then yeah be my guest. I’m just saying that this one type of poor behaviour is unrelated to the other kinds of poor behaviour.
Also just checked and every open ai model bigger than 4.1-mini can answer this. I think the joke should emphasize how we developed a super power inefficient way to solve some problems that can be accurately and efficiently answered with a single algorithm. Another example is using ChatGPT to do simple calculator math. LLMs are good at specific tasks and really bad at others, but people kinda throw everything at them.
And yet they can seemingly spell and count (small numbers) just fine.
what do you mean by spell fine? They’re just emitting the tokens for the words. Like, it’s not writing “strawberry,” it’s writing tokens <302, 1618, 19772>, which correspond to st, raw, and berry respectively. If you ask it to put a space between each letter, that will disrupt the tokenization mechanism, and it’s going to be quite liable to making mistakes.
I don’t think it’s really fair to say that the lookup 19772 -> berry counts as the LLM being able to spell, since the LLM isn’t operating at that layer. It doesn’t really emit letters directly. I would argue its inability to reliably spell words when you force it to go letter-by-letter or answer queries about how words are spelled is indicative of its poor ability to spell.
I mean that when you ask them to spell a word they can list every character one at a time.
Well that’s a recent improvement. GPT3 was very bad at that, and GPT4 still makes mistakes.
The problem is that it’s not actually counting anything. It’s simply looking for some text somewhere in its database that relates to that word and the number of R’s in that word. There’s no mechanism within the LLM to actually count things. It is not designed with that function. This is not general AI, this is a Generative Adversarial Network that’s using its vast vast store of text to put words together that sound like they answer the question that was asked.