HN user
I'm sorry, this article reads like AI slop.
It has all the hallmarks: grandiose writing ("everything changed"), the classic "It wasn't X, it was Y" (about five times in the first minute of reading the article), undue emphasis on symbolism...
All those indicators are coupled with an unpleasant level of obsession with how statistical models are "a new kind of mind"—even after claiming to "strip away the hype cycle".
Regularization as a concept is taught in introductory ML classes. A simple example is called L2 regularization: you include in your loss function the sum of squares of the parameters (times some constant k). This causes the parameter values to compete between being good at modeling the training data and satisfying this constraint--which (hopefully!) reduces overfitting.
The specific regularization techniques that any one model is trained with may not be publicly revealed, but OAI hardly deserves credit for the concept.
We absolutely should treat it just like any other software tool.
Let's distinguish between papers and preprints, please. arXiv has contributed to a blurring of the distinction. The arXiv preprints are useful but should always be taken with a grain of salt. There is nearly no filtering done on things uploaded to arXiv.
Everyone accessing someone's uncritically reviewed work is a bittersweet gift.
This repo is a joke, right? I'd be embarrassed peddling this as AGI.
We're a long way from AGI existing at all. Even if you disagree, it's agreed upon that we're not there yet. For this repo to call its offering AGI is even more dramatic of a mischaracterization than saying I'm an Olympic sprinter because I went for a morning jog.
What can I say? Hucksters gonna huckster.
Google absolutely has their own internal models that do exactly this. It wouldn't surprise me if Microsoft indeed does have an internal Copilot that is trained on their data, but even on the smallest risk that they leak their code, they wouldn't share that particular model.
Well…one that peeks at the test set labels.
Further analysis shows that it doesn’t perform well at all—successes are tied to things like test set leakage.
https://kenschutte.com/gzip-knn-paper2/
This paper isn’t any surprisingly effective result. It’s thoroughly shoddy scholarship by which the authors should feel embarrassed.
Yes - it's mentioned, but doesn't the framing below make it sound like they're still advocating for this paper?
In essence, it's advisable to take the paper’s reported figures with a grain of salt, particularly as they cannot be precisely reproduced as described. Nonetheless, this approach continues to deliver unexpectedly well.
A "grain of salt" is different from "critical evaluation flaw," and if the reproduction's results are true, then the method doesn't after all "deliver unexpectedly well".
The paper has recently been called into question for overestimating their performance relative to BERT: https://news.ycombinator.com/item?id=36758433. Might be good for the blog's author to take this into account in their explainer. The author's perspective sounds a bit too positive (and borderline salesmanlike).
The title wordplay dates back to at least Drew McDermott's 1976 essay "Artificial Intelligence Meets Natural Stupidity" [0]. The intro is phenomenal.
---
As a field, artificial intelligence has always been on the border of respectability, and therefore on the border of crackpottery. Many critics <Dreyfus, 1972>, <Lighthill, 1973> have urged that we are over the border. We have been very defensive toward this charge, drawing ourselves up with dignity when it is made and folding the cloak of Science about us. On the other hand, in private, we have been justifiably proud of our willingness to explore weird ideas, because pursuing them is the only way to make progress.
Unfortunately, the necessity for speculation has combined with the culture of the hacker in computer science <Weizenbaum, 1975> to cripple our self-discipline. In a young field, ,self-discipline is not necessarily a virtue, but we are not getting any younger. In the past few years, our tolerance of sloppy thinking has led us to repeat many mistakes over and over. If we are to retain any credibility, this should stop.
This paper is an effort to ridicule some of these mistakes.
---
[0]: Drew McDermott. 1976. Artificial intelligence meets natural stupidity. SIGART Bull., 57 (April 1976), 4–9. https://doi.org/10.1145/1045339.1045340
By your first paragraph's argument, the semantics are in the Transformer, not the tokenizer.
And yes, what they do helps on their two test tasks. I'm not disputing that. It's the fact that there's no scholarship here.
There are so many thousands of knobs to twiddle with in a model these days, and they went after one that's commonly regarded in the NLP community as the 'defect'—the only part of the model that's not end-to-end trained along with the rest. Which would be great, if they acknowledged it! But there's no citation to any tokenization literature beyond BPE or SentencePiece. The literature review is as superficial as what you could find in a blog.
There are certainly byte-level or character-level tokenizers (think about CANINE or ByT5), and we can argue back and forth about their data-hungriness or slow inference. It would be nice to give more helpful units to a Transformer, so it doesn't have to learn syllables (or even characters) all on its own. Rebracketing/incorrect segmentation is a problem! And these authors have clued into that, but so have several hundred (or thousand?) researchers they don't cite.
What I'm having trouble with is the notion that this paper uncovered some exciting, revelatory fact about tokenization. Yes, "Japanese Language" would be a reasonable semantic unit! But these authors didn't discover that fact. Nobody's questioning whether 'good tokenization is better than bad tokenization'. Tokenization has seen ongoing attention in NLP forever.
These authors tried one variant, compared it against a library default option (and nothing else), evaluated on one task, put a bit of marketing around it, and called it a day. In the NLP course I used to TA, this wouldn't even qualify as a complete final project for the course.
I was being generous - stemming is poor man's morphology. Empirically useful (ask the IR folks) but incredibly heuristic.
I'm going to add a contrarian take here: this preprint is not a research paper. While it's nice to see that there is an improvement here on their one task, this is not "semantically" driven tokenization. It's morphologically driven. To be semantically driven, it would be reasonable to expect that synonyms would have similar representations. I got really excited from the title, and the content is a let-down.
The line of research here has been going on for 30+ years, from Michael Brent's work, to Linguistica, to Morfessor, and now several approaches to incorporate morphology into tokenizers. The stand-out example is [0]. This paper doesn't seem to acknowledge any of that intellectual legacy. It's not a _research_ paper.
I'm getting a bit tired of people putting their class projects or quick engineering projects on arXiv. I don't know why they're surfacing so high on HN either.
Yep! Percy Liang in an interview with Chris Potts said he sees BERT and ELMo as foundation models.
This is a God (language)-of-the-gaps argument: we can't figure out this rarely language, but maybe we can figure out an entirely unattested language instead, and also learn the correspondence between it and Linear A.
Deep learning can predict plenty of phenomena in the world, sure, but it needs data, not aspirations.
Is that a pun, because this was done by Pearson?