(I realize that choosing the most likely word wouldn't necessarily solve the issue, but choosing the most likely phrase possibly might.)
Edit, post seeing the video and comments: it's beam search, along with temperature to control these things.
Temperature and top_k (two very similar parameters) were both introduced to account for the fact that human text is unpredictable stochastically for each sentence someone might say as such - as shown in this 2021 similar graph/reproduction of an older graph from the 2018/2019 HF documentation: https://lilianweng.github.io/posts/2021-01-02-controllable-t...
It could be that beam search with much longer length does turn out to be better or some merging of the techniques works well, but I don't think so. The query-key-value part of transformers is focused on a single total in many ways - in relation to the overall context. The architecture is not meant for longer forms as such - there is no default "two token" system. And with 50k-100k tokens in most GPT models, you would be looking at 50k*50k = A great deal more parameters and then issues with sparsity of data.
Just everything about GPT models (e.g. learned positional encodings/embeddings depending on the model iteration) is so focused on bringing the richness of a single token or single token index that the architecture is not designed for beam search like this one could say. Without considering the training complications.
At the output of an LLM the raw next-token prediction values (logits) are passed through a softmax to convert them into probabilities, then these probabilities drive token selection according to the chosen selection scheme such as greedy selection (always choose highest probability token), or a sampling scheme such as top-k or top-p. Under top-k sampling a random token selection is made from one of the top k most probable tokens.
The softmax temperature setting preserves the relative order of output probabilities, but at higher temperatures gives a boost to outputs that would otherwise have been low probability such that the output probabilities are more balanced. The effect of this on token selection depends on the selection scheme being used.
If greedy selection was chosen, then temperature has no effect since it preserves the relative order of probabilities, and the highest probability token will always be chosen.
If a sampling selection scheme (top-k or top-p) was chosen, then increased temperature will have boosted the likelihood of sampling choosing an otherwise lower probability token. Note however, that even with the lowest temperature setting, sampling is always probabilistic, so there is no guarantee (or desire!) for the highest probability token to be selected.
(80% of the time) The answer to the expression 2 + 2 is 4
(15% of the time) The answer to the expression 2 + 2 is Four
(5% of the time) The answer to the expression 2 + 2 is certainly
(95% of the time) The answer to the expression 2 + 2 is certainly Four
This is how you can asp ChatGPT the same question few times and it can give you different words each time, and still be correct.
For example if you ask a model what is 0^0, the highest probability output may be "1", which is incorrect. The next most probable outputs may be words like "although", "because", "Due to", "unfortunately", etc. as the model prepares to explain to the user that the value of the expression is undefined; because there are many more ways to express and explain the undefined answer than there are to express a naively incorrect answer, the correct answer is split across more tokens so that even if eg the softmax value of "1" is 0.1 and across "although"+"because"+"due to"+"unfortunately">0.3, at temperature of 0, "1" gets chosen. At slightly higher temperatures, sampling across all outputs would increase the probability of a correct answer.
So it's true that increasing the temperature increases the probability that the model outputs tokens other than the single-most-likely token, but that might be what you want. Temperature purely controls the distribution of tokens, not "answers".
This is where the semi-ambiguity of the human languages helps a lot with.
There are multiple ways to answer with "4" that are acceptable, meaning that it just needs to be close enough to the desired outcome to work. This means that there isn't a single point that needs to be precisely aimed at, but a broader plot of space that's relatively easier to hit.
The hefty tolerances, redundancies, & general lossiness of the human language act as a metaphorical gravity well to drag LLMs to the most probable answer.
> 2 + 2
You really couldn't come up with an actual example of something that would be dangerous? I'd appreciate that, because I'm not seeing reason to believe that an "output beyond the most likely one" output would end up ever being dangerous, as in, harming someone or putting someone's life at risk.
Thanks.
Can you explain how it chooses one of the lower-probability tokens? Is it just random?
That said, LLM reach their current performance despite this limitation.
An example is Beam Search:https://www.width.ai/post/what-is-beam-search
Essentially we keep a window of probabilities of predicted tokens to improve the final quality of output.
Here is a blog post that describes it: https://huggingface.co/blog/how-to-generate.
I will warn you though that beam search is typically what you do NOT want. Beam search approximately optimizes for the "highest likely sequence at the token level." This is rarely what you need in practice with open-ended generations (e.g. a question-answering chat bot). In practice, you need "highest likely semantic sequence," which is much harder problem.
Of course, various approximations for semantic alignment are currently in the literature, but still a wide open problem.
(No affiliation)
I have no idea why you say this. Most of our pipelines will run greedy, for reproducibility.
Maybe we turn the temp up if we are returning conversational text back to a user.
Thats basically chunking or at least how it starts. I was impressed by the ability to add and subtract the individual word vector embeddings and get meaningful results. Chunking a larger block blends this whole process so you can do the same thing but in conseptual space the so take a baseline method like sentence embedding and that becomes your working block for comparison.
If you haven't seen the first few chapters, I cannot recommend enough.
Andrew Ng's course doesn't use video effectively at all: half of each class is Andrew talking to the camera, while the other half is him slowly writing things down with a mouse. There's a reason why a lot of people recommend watching at 1.5x speed.
Online classes are online classes. If they try to make copy in-person classes, like most Coursera courses do, they will keep all of the weaknesses of online classes without any of its strengths.
3BlueOneBrown videos are a great complement to Karpathy's lectures to aid in visualising what is going on.
I was ignorant enough to try and jump straight in to his videos and despite him recommending I watch his preceeding videos I incorrectly assumed I could figure it out as I went. There is verbiage in there that you simply must know to get the most out of it. After giving up, going away and filling in the gaps though some other learnings, I went back and his videos become (understandably) massively more valueable for me.
I would strongly recommend anyone else wanting to learn neural networks that they learn from my mistake.
Going through Andrej's makemore tutorials required quite a lot of time but it's definitely worth it. I used free tier of Google Colab until the last one.
Pausing the video a lot after he explains what he plans to do and trying to do it by myself was a very rewarding way to learn, with a lot of "aha" moments.
Prior discussion: https://news.ycombinator.com/item?id=38505211
Thank you for sharing.
