First of all, this link is just to C# bindings of llama.cpp and so doesn’t contain the actual implementation. But it also doesn’t refute my criticism of your claim. More specifically, I take issue with this statement that you said: “today’s LLM’s ingest a ton of text [snip] and builds up statistics of which tokens it sees in that context”.
I claim that this is not how today’s LLMs work because we have no idea what LLMs do with the input data during training. We have very little insight into what kind of data structure it builds and how the data structure it built is organized.
No, it’s done because one letter at a time is too slow. Tokens are a “happy” medium tradeoff.
I think I worded my sentence ambiguously, let me re-word it for you: “Going one token at a time is only considered a limitation because LLMs are not accurate enough”
It makes a “break” of the block, which lets it start a new answer instead of continuing on the previous. How it reacts to that depends on the fine tune and filters before the data hits the LLM.
Once again, my sentence was not written well, my bad. I was commenting on the observed behavior, not on how it works from a technical perspective.
I have just said that LLM’s we have today can’t fix the problems with false data and hallucinations, because it’s a core principle of how it operates. It will require a new approach.
You could add a rocket engine and wings to a pogo stick, but then it’s no longer a pogo stick but an airplane with a weird landing gear. Today’s LLM’s could give us hints to how to make a better AI, but that would be a different thing than today’s LLM’s. From what has been leaked from OpenAI GPT4 has scaling issues so they use mixture of experts. Just throwing hardware at it is already showing diminishing returns. And we’re learning fascinating new ways of training them, but the inherent problem is the same.
Alright, we agree here for the most part so I’m just going to skip this.
For example, if you ask an LLM if it can give an answer to a question, it will have two paths to go down, positive and negative. Note, at the point where it chooses that it doesn’t know how to finish it, it doesn’t look ahead.
This is weird though. How do you know LLMs can’t look ahead? When we prompt LLMs, we are basically asking them this question: “What is the next word of your response?” How do you know it hasn’t written out the entire response in memory already after which it only shows you the first word? LLMs are neural networks. Neural networks have working memory. That’s how neural networks work after all, it’s just a vector of data that is repeatedly transformed as it passes through each layer. Of course, if it does write the entire response in memory, it is all thrown away after every word.
As far as the backspace tokens go, you are right to be skeptical but also do not be surprised if it works out. We’ve had LLMs trained to complete and edit text for some time already. They’ve fallen out of use today but they did perform acceptably well.
First of all, this link is just to C# bindings of llama.cpp and so doesn’t contain the actual implementation.
I know, it’s my code. I refactored it from some much less readable and usable c# code. I picked it because it more clearly shows the steps involved in generating text.
How do you know LLMs can’t look ahead? […] How do you know it hasn’t written out the entire response in memory already after which it only shows you the first word?
Firstly, it goes against everything we know so far of how they operate, and secondly… because they can’t.
If you look at the C# code, the first step is in _process_tokens function, where it feeds the context into llama_eval. That goes through each token and updates the internal memory / model state. Since it saves state, if you already have processed some of the tokens you can tell it to skip them and start on the new ones.
After this function you have a state in memory, the current state of the LLM, as a result of the tokens it’s seen so far.
When we are done with that, we go to the more interesting part, the _predict_next_token function. Note that that takes a samplingparams parameter. It then set some options, like if top k is not set it’s set to length of the model’s vocabulary (number of tokens it knows about), and repeat_last_n, if not set, is set to the length of the existing context.
The code then gets the model’s vocabulary, aka all the tokens it knows about, and then it generates the logits. The logits is an array the length of the vocabulary, with a number for each token showing how likely that one is the next token. The code then adds any specified token bias to that token’s number. Already here, even if it already had a specific answer in mind, you can see problems starting.
Then the code adds token repetition penalty, based on the samplingparams. This means that if a token repeats inside the given history, it’s value will be lowered according to the repeat_penalty. Again, even if it had a specific answer, this has a high chance of messing that up. The same is done for frequency and presence. For more details of what those native functions do, you can see the llama.cpp source - they have the same name there.
After all the penalties are applied, it’s time to pick the token. If the temp is 0 or lower, it just picks the highest rated token (aka greedy sampling). This tends to give very boring and flat responses, but it’s predictable and reproduceable, so it’s often used in benchmarks of various kinds.
But if that’s not used (which it almost never is in “real” use), there are several methods. You have MiroStat, which tries to create more consistent quality between different answer lengths, and the “traditional” using top-k, top-p and temperature.
Common for them is however that internally it produces a top list of candidates, and then pick one at random. And that’s why a LLM can’t plan ahead.
When a token ID is eventually produced it returns the new ID, that gets added to context, the text equivalent of the token is looked up and sent back to the UI, and the new context is fed into llama_eval and the process starts again.
For the LLM to even be able to plan an answer ahead it must know of all penalties and parameters (or have none applied), and greedy token prediction must be used.
And that is why, even if it had some sort of near magical ability to plan ahead that we just don’t know is there, at the end of the day it could still not plan a specific response.
Wow, very nice! First of all, I will preface by admitting that I have not worked with LLMs to the degree of making a toy implementation. Your explanation of the sampling techniques is insightful but doesn’t clear up my confusion. Why does sampling imply the absence of higher level structure in the model?
For example, even though poker is highly influenced by chance, I can still have a plan that will increase my likelihood of winning. I don’t know what card will be drawn next but I can prepare strategies for each possible card. I can have preferences for which cards I want to be drawn next.
First of all, this link is just to C# bindings of llama.cpp and so doesn’t contain the actual implementation. But it also doesn’t refute my criticism of your claim. More specifically, I take issue with this statement that you said: “today’s LLM’s ingest a ton of text [snip] and builds up statistics of which tokens it sees in that context”.
I claim that this is not how today’s LLMs work because we have no idea what LLMs do with the input data during training. We have very little insight into what kind of data structure it builds and how the data structure it built is organized.
I think I worded my sentence ambiguously, let me re-word it for you: “Going one token at a time is only considered a limitation because LLMs are not accurate enough”
Once again, my sentence was not written well, my bad. I was commenting on the observed behavior, not on how it works from a technical perspective.
Alright, we agree here for the most part so I’m just going to skip this.
This is weird though. How do you know LLMs can’t look ahead? When we prompt LLMs, we are basically asking them this question: “What is the next word of your response?” How do you know it hasn’t written out the entire response in memory already after which it only shows you the first word? LLMs are neural networks. Neural networks have working memory. That’s how neural networks work after all, it’s just a vector of data that is repeatedly transformed as it passes through each layer. Of course, if it does write the entire response in memory, it is all thrown away after every word.
As far as the backspace tokens go, you are right to be skeptical but also do not be surprised if it works out. We’ve had LLMs trained to
completeandedittext for some time already. They’ve fallen out of use today but they did perform acceptably well.I know, it’s my code. I refactored it from some much less readable and usable c# code. I picked it because it more clearly shows the steps involved in generating text.
Firstly, it goes against everything we know so far of how they operate, and secondly… because they can’t.
If you look at the C# code, the first step is in
_process_tokensfunction, where it feeds the context intollama_eval. That goes through each token and updates the internal memory / model state. Since it saves state, if you already have processed some of the tokens you can tell it to skip them and start on the new ones.After this function you have a state in memory, the current state of the LLM, as a result of the tokens it’s seen so far.
When we are done with that, we go to the more interesting part, the
_predict_next_tokenfunction. Note that that takes asamplingparamsparameter. It then set some options, like if top k is not set it’s set to length of the model’s vocabulary (number of tokens it knows about), and repeat_last_n, if not set, is set to the length of the existing context.The code then gets the model’s vocabulary, aka all the tokens it knows about, and then it generates the logits. The logits is an array the length of the vocabulary, with a number for each token showing how likely that one is the next token. The code then adds any specified token bias to that token’s number. Already here, even if it already had a specific answer in mind, you can see problems starting.
Then the code adds token repetition penalty, based on the
samplingparams. This means that if a token repeats inside the given history, it’s value will be lowered according to the repeat_penalty. Again, even if it had a specific answer, this has a high chance of messing that up. The same is done for frequency and presence. For more details of what those native functions do, you can see the llama.cpp source - they have the same name there.After all the penalties are applied, it’s time to pick the token. If the temp is 0 or lower, it just picks the highest rated token (aka greedy sampling). This tends to give very boring and flat responses, but it’s predictable and reproduceable, so it’s often used in benchmarks of various kinds.
But if that’s not used (which it almost never is in “real” use), there are several methods. You have MiroStat, which tries to create more consistent quality between different answer lengths, and the “traditional” using top-k, top-p and temperature.
Common for them is however that internally it produces a top list of candidates, and then pick one at random. And that’s why a LLM can’t plan ahead.
When a token ID is eventually produced it returns the new ID, that gets added to context, the text equivalent of the token is looked up and sent back to the UI, and the new context is fed into llama_eval and the process starts again.
For the LLM to even be able to plan an answer ahead it must know of all penalties and parameters (or have none applied), and greedy token prediction must be used.
And that is why, even if it had some sort of near magical ability to plan ahead that we just don’t know is there, at the end of the day it could still not plan a specific response.
Wow, very nice! First of all, I will preface by admitting that I have not worked with LLMs to the degree of making a toy implementation. Your explanation of the sampling techniques is insightful but doesn’t clear up my confusion. Why does sampling imply the absence of higher level structure in the model?
For example, even though poker is highly influenced by chance, I can still have a plan that will increase my likelihood of winning. I don’t know what card will be drawn next but I can prepare strategies for each possible card. I can have preferences for which cards I want to be drawn next.