It’s Luna, their small, low-intelligence model. But still impressive to see.
I also found this screenshot on Reddit:
The model spent 26 seconds for some reason 🤔
You must log in or register to comment.
Success at letter counting is not a particularly good way of predicting model capabilities.
There are ads in chatgpt??
We just had to wait.
Isn’t it something to do with tokenisation? Like double e gets into one token or something?
According to this page ChatGPT would tokenize “seventeen” as [“se”, “venteen”] but AFAIK tokens become multi-dimensional vectors which represents its meaning (all vectors are the same length and the numbers in the vector have nothing to do with the letters that spell it out, just so you don’t imagine something like a C++ string). So an LLM can’t answer how many of each letters there is in a token either, unless explicitly trained to do so. To be able to immediately answer this, I think they would need to include huge amounts of training data to cover all tokens and letter combinations.
But I would have expected that since the strawberry incident, all models would have been trained to do what Qwen tried to do here, and break the word down into letters. Because “s e v e n t e e n” tokenizes to [“s”, “e”, “v”, “e”, “n”, “t”, “e”, “e”, “n”] so then they only have to train the model to be able to convert any token into its individual letter, and to count occurrences of something in an array.
However, if you disallow thinking and force the LLM to answer immediately, then you’re back to the original problem and will likely get an incorrect answer.
Kind of but not really. AIs learn to predict text but they can’t really do math the traditional way. They learn the tokens, which are more than syllables and less than a word long. So because they don’t do math, LLMs in particular speculate “something like 3 comes after this question” because in training it has seen similar questions answered with “3”. That’s the jist of it.
Your generalization, you see, doesn’t apply to what “Qwen27b” did here above, right ?
It doesn’t. I wouldn’t say I made a generalization but a very high level explanation so it’s easier to understand for non-tech people.
I suspect modern models are trained specifically on these kinds of questions to spare the embarrassment of this question for future models.
that does mean they stochastically understand the occurrence of other numbers for this question and if you wrote tokens similar to this it would answer with 4 or whatever the specific case. I’m just trying to explain that no matter whether it gets it right or wrong, it goes off of statistics and semantics, not math. Because these things are not made to do math, that’s what algorithms are for, and we know very well how to write them.
Yes he’s talking about why chatgpt failed the task
deleted by creator
This can happen because of tokenization. Tell it to execute code to do the letter counting and both models will succeed every time
