Comment by howunfortunate
9 hours ago
As an MLE I feel these takes are too reductionist.
You could say the (nearly) same thing about search. And content recommendation. And clustering. And topic modeling. And outlier detection. And spam filtering. And image diffusion. And dimension reduction. And...
There's a lot in common between these things, but there's also a lot cool and different!
For transformers in particular, it's pretty cool that you get some WILD emergent properties simply from scaling up.
So yes, it's just a next token predictor, but I'm just a bundle of nerves and meat. I don't get a lot out of those descriptions.
Some concrete facts about LLMs are explained by their next token predictor nature. Every time it says "wait, that's wrong." instead of generating the correct thing the first time.
I think that's relatively emergent too though! BERT never really did that (at least to my recollection), presumably because its training was never sufficient for it to develop corrective reasoning in a chain of thought.
What exactly do you mean with "emergent" here?
2 replies →
BERT isn't a next token predictor. It predicts a single token based on the whole surrounding context in both directions.
1 reply →
Obligatory link to the classic copy-pasta:
> They're Made out of Meat
https://web.mit.edu/people/dpolicar/writing/prose/text/think...
It's a little different than that. Your bundle of nerves and meat is not static. It changes over time.
To me the heart of the "next token predictor" is that the distributions are static. You can manipulate what you feed into it through context (and a lot of interesting engineering has been applied there through CoT and other techniques to manipulate the prompt). But these models as implemented will never be able try things and learn from mistakes or adapt. They are a set of weights frozen in time. A set of distributions derived from the original data that created them.
That's not quite true though. The fact that most models are in practice non-deterministic has been a huge point of contention.
And there's nothing inherently stopping labs from continuously fine-tuning the weights after every new invocation. It's just a difficult (not to mention expensive) software and ML problem.
How are they non-deterministic? Ok there is random involved but for watermarking results the random is substituted by pseudo-random, otherwise it won’t work. If you control the randomness, you should be able to reproduce answers to prompts in equal context 100%.
I was not saying that they are deterministic, rather that the distributions (aka weights) are fixed. A model as deployed today at anthropic/open ai/etc is not learning beyond the context as far as I know.
What prevents continuous fine-tuning from what I understand is catastrophic forgetting. You can do things like RLHF which are built to minimize the damage but that is more about bringing out capabilities of the base model than incorporating new knowledge (at least from my understanding, I am obviously not a researcher at a lab).
3 replies →
It's not "not quite true", it's literally true because alternative architectures like RNNs and Mamba fully update their own internal states, whereas transformers only append to the context.
RNNs and Mamaba do not update their weights, but you could hypothetically scale the internal state to be as big as Fable's and GPT 6's parameters.