i hate "stochastic parrot" because it's not even really meaningful
I think it's true that models are statistical, inasmuch as P(A|B) where B is the prior sequence is what the loss is computing, and that's statistical. It's just computing that function in an absurdly complex way, which involves creating topological representations of relationships, etc.
I agree that "just" autocomplete implies the wrong thing. It turns out autocomplete is amazing if you scale it up.
I think it's true that they reason and area creative but these are really hard points because people mean subtly different things when saying "reason" and "creative".
fortunately i wrote an entire post about what the difference is between the parts of this that it is easy to make sense of and the parts of it that it is prohibitively difficult to make sense of and it was posted on hackernews
I agree, what we do is much closer to growing them than to engineering them. We basically engineer the conditions for growth, and then check the results and try again.
My best argument that insights from neuroscience will transfer to neural networks, and vice versa:
For sufficiently complex phenomena (e.g., language), there should only be one reasonably efficient solution to the problem, and small variations on that solution. So there should be some reversible mapping between any two tractable solutions to the problem that is pretty close to lossless, provided both solutions actually solve the problem.
And, yeah, the main advantage of neural networks is that they're white-box. You can also control your experiments in a way you can't in the real world.
My short explanation would be that even for RL, you are training on a next token objective; but the next token is something that has been selected very very carefully for solving the problem, and was generated by the model itself.
So you're amplifying existing trajectories in the model by feeding the model's outputs back to itself, but only when those outputs solve a problem.
This elides the kl penalty and the odd group scoring, which are the same in the limit but vastly more efficient in practice.
Every LLM provider caches their KV-cache, it's a publicly documented technique (go stuff that KV in redis after each request, basically) and a good engineering team could set it up in a month.
I am not deep into the Linus weeds but my impression is that he doesn't especially care if he's on the receiving end of this. It only started to feel different from "well, the Linux list is the PvP zone" when Linus was sufficiently weighty/famous that you almost had to take an insult from him to heart, and he did eventually correct his behavior there.
you might have better luck giving the LM the original document and having it generate its own OCR independently, then asking the llm to tiebreak between its own generation and the OCR output while the image is still in the context window until it is satisfied that it got things correct
tokens are on average four characters and the number of residual streams (and therefore RAM) the LLM allocates to a given sequence is proportionate to the number of units of input. the flops is proportionate to their square in the attention calculation.
you can hypothetically try to ameliorate this by other means, but if you just naively drop from tokenization to character or byte level models this is what goes wrong
2) implement guaranteed citation of the author any time the model gave an answer that was directly derivative, with an option to not do so if the summary was sufficiently vague,
or
3) ignore the author's book completely as training data
These are fancy Markov chains in the sense that humans are just chemicals and computers just do math. Technically true, but not even "overly reductive"; it is just wrong if it is used to imply that, e.g., humans just swirl around in beakers or the most complex thing you can do with computers is trigonometry.
You can make anything sound unimpressive if you describe it sufficiently poorly.
And: So many different variations are published every month. There are a good number of people in serious research trying approaches that don't use cross entropy loss (ie, strictly next-token prediction).
I don't know what the trajectory of the technology is over the next ten years, but I am positive no one else does either and anyone who thinks they do is wrong.
You are uploading 5 billion examples of <something>. You cannot filter it manually, of course, because there are five billion of it. Given that it is the year 2024, how hard is it to be positive that a well-resourced team at Stanford in 2029 will not have better methods of identifying and filtering your data, or a better reference dataset to filter it against, than you do presently?
I am pretty sure if the authors were trying to license their works for this purpose we would just not use them at all; it is difficult to see under what circumstances they would stand to profit from this other than by suing people after the fact over it.
I think it's true that models are statistical, inasmuch as P(A|B) where B is the prior sequence is what the loss is computing, and that's statistical. It's just computing that function in an absurdly complex way, which involves creating topological representations of relationships, etc.
I agree that "just" autocomplete implies the wrong thing. It turns out autocomplete is amazing if you scale it up.
I think it's true that they reason and area creative but these are really hard points because people mean subtly different things when saying "reason" and "creative".