canjobear

It was a nice way to organize things so you could find stuff serendipitously. I remember clicking around in my neighborhood (where I had my N64 cheat codes website) and finding some cool website with lots of slick looking 3D renders. It's rare to stumble across unplanned-for things like that nowadays.

Personality disorders like BPD tend to attenuate with age, so you would be more likely to become less sociopathic.

Correlation/causation. People whose faculties are still intact are more likely to do and enjoy activities requiring those faculties.

The academic paper is here https://www.researchgate.net/publication/361675439_The_Decip...

It looks convincing to me. It's exactly the same method used to decipher cuneiform, starting with proper names in certain inscriptions and then figuring out the content of other inscriptions somewhat like a crossword puzzle. The fact that one of the tablets turns out to record the syllabary in order is especially compelling.

The problem with Indus script is that the inscriptions are all very short. This gives you very little information to go on.

The Proto-Indo-European language is usually dated to something like 6000 years ago, well before any writing.

This is a booming genre. For example this one popped up in my recommendations yesterday https://www.youtube.com/watch?v=uVKMY-WTrVo

Do you have any writeups of this work?

It seems like the same set of arguments to me.

I’m an author of the paper that this article is about. Ask questions!

There is no compensation for reviewers, and usually no compensation for editors. It’s effectively volunteer work. I agree to review a paper if it seems interesting to me and I want to effectively force myself to read it a lot more carefully than normal. It’s hard work, especially if there is a problem with the paper, because you have to dig out the problem and explain it clearly. An academic could refuse to do any reviews with essentially no formal consequences, although they’d get a reputation as a “bad citizen” of some kind.

This is standard lossless compression. None of the concepts particular to lossy compression (like rate-distortion theory) are used.

Yes, this is an area of active research. For example https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm...

> Instead you mostly created an abstraction of the problem that no one cares about: that there exist certain synthetic language constructions that LLMs have difficulty with. But this is both trivial (consider a language that requires you to factor numbers to decode it) and irrelevant (there is no relation to what humans do except in an abstract sense).

Thanks for your feedback. I think our manipulations do establish that there are nontrivial inductive biases in Transformer language models and that these inductive biases are aligned with human language in important ways. There's no universal a priori sense in which Moro's linear counting languages are "simple" but our deterministically shuffled languages aren't. It seems that GPT language models do favor real language over the perturbed ones, and this shows that they have a simplicity bias which aligns with human language. This is remarkable, considering that the GPT architecture doesn't look like what one would expect based on existing linguistic theory.

Furthermore, this alignment is interesting even if it isn't perfect. I would be shocked in GPT language models happened to have inductive biases that perfectly match the structure of human language---why would they? But it is still worthwhile to probe what those inductive biases are and to compare them with what humans do. As a comparison, context-free grammars turned out to be an imperfect model of syntax, but the field of syntax benefited a lot from exploring them and their limits. Something similar is happening now with neural language models as models of language learning and processing, a very active research field. So I wouldn't say that neural language models can't shed any light on language simply because they're not a perfect match for a particular aspect of language.

As for using languages more directly based on the Moro experiments, we've discussed this extensively. There are nontrivial challenges in scaling those languages up to the point that you can have a realistic training set, where the control condition is a real language instead of a toy language, without introducing confounds of various kinds. We're open to suggestions. We've had very productive conversations with syntacticians about how to formulate new baselines in future work.

More generally our goal was to get formal linguists more interested in defining the impossible vs. possible language distinction more carefully, to the point that they can be used to test the inductive biases of neural models. It's not as simple as hierarchical vs. linear, since there are purely linear phenomena in syntax such as Closest Conjunct Agreement, and also morphophonological processes can act linearly across constituent boundaries, among other complications.

> The one language that you use that is most similar to the linear languages cited by Moro, "Hop", shows very little difference in performance, directly undermining your claimed refutation of Chomsky.

I wouldn't read much into the magnitude of the difference between NoHop and Hop, because the Hop transformation only affects a small number of sentences, and the perplexity metric is an average over sentences.

> A one-to-one mapping is not structure preserving.

That's just the point we were making: if you mess with the structure of natural language, then language models don't learn it as well any more.

The transformations do preserve the entropy in the sense that the lowest achievable perplexity is the same for everything except the nondeterministic shuffle. Since applying a one-to-one function preserves the entropy of a discrete random variable, in this case the random variable over documents in the training and test sets. In principle, a universal approximator should be able to learn to invert any one-to-one transformation we apply, although of course in practice a GPT architecture doesn't achieve this.

The transformations do mess up the local entropy in strings. But I think that's part of the point. Human language seems to be structured in a way so that things are locally predictable. When you screw up that structure, then languages are harder to learn.

We are working on a followup with more transformations including syntactic ones, as you might imagine. It's surprisingly hard to come up with manipulations that target specific aspect of linguistic structure, while fitting the criteria that (1) they don't affect the lowest achievable perplexity, (2) they clearly change the language from "possible" to "impossible" in a way that all or most linguists would agree with, (3) they can actually be implemented using the data we have---for example a transformation that relies on detailed syntactic parses would require that we parse the whole dataset, which is not only time consuming but also introduces possible confounds from errors in the parser, etc. We're talking to a lot of people, if you have ideas we'd be happy to hear them!

Thanks for the feedback! The point about perplexity is totally valid for the nondeterministic shuffle baseline. This seems to have misled a lot of people. But for all the other baselines, we're applying some one-to-one transformation function to the original training set, and so not increasing the inherent entropy of the distribution being learned.

As for tokenization, good point: it's worth retokenizing based on the altered datasets to see if that changes anything. I think it might not, because the tokenizers we use are based on the frequency distribution of "words" identified by whitespaces, but we'll have to check.

That brief moment of thinking he means you should cut off your thumb is a source of humor.

What's your objection to Mission Impossible Language Models?

Obviously no one is claiming that nothing happened before writing. Historians use “history” as a technical term for events recorded in written and other durable records. They use “prehistory” for events recorded in oral traditions and archaeology.

Consider the many cultures that Europeans discovered in the Americas. Of them, only one had a real writing system (the Maya). The Aztecs wrote down history in a sort of comic book form, and the Inca did accounting with qhipu, but only the Maya had a conventional system of written forms that correspond systematically with verbal language.

You can communicate and remember a lot orally. The Polynesian navigators encoded information about how to navigate from island to island in oral poems passed down generation to generation. A lot of these traditions seem to vanish once a culture gets writing, leaving it unclear how they used to pass on so much information before.

I used Sonique for a while because when I downloaded Winamp the demo said “it really whips the llama’s ass!” and I was afraid my parents would take away my internet access if they heard such a scandalous thing.

It did have cool skins.

Once we can reliably induce lucid dreaming, and maybe with the help of some neural prosthesis to capture the output, we can recover all that lost productivity time by having people do immersive vr spreadsheet work in their dreams.

Interestingly this has a rather frequentist flavor. The probabilities end up coming from frequency ratios in very large samples.

The principle of end weight is a special case of a more general principle: sentences are usually easier to understand when syntactically related words are close to each other, all else being equal. A more precise version of Strunk and White's "keep related words together".

The interesting thing is that in English this generally pushes you to put short phrases before long phrases, as described in the article, but in languages with other word orders you can get the opposite effect. For example in Japanese the verb is always the last thing in a sentence, and so the way to keep related words together is actually to put long phrases before short phrases. So you'll often get sentences with structures like [[very long object] [short subject] verb].

It's not hard to find the hall. Grad student offices now, I believe.

I did this once too, and had to pay a “reinstatement fee” to get on the original trip back.

Copernicus used the same circular-orbit-plus-epicycles system as Ptolemy, just the orbits were centered around the sun (kind of---each planet had its own circle, with the sun only approximately in the middle). The system actually had more epicycles than Ptolemy's and was less accurate. It wasn't an advance in any meaningful sense.

The real breakthrough was Kepler, who dropped the idea that planets moved in circles. It was indeed partly a mathematical breakthrough and the reason Kepler's work took a while to catch on is that people couldn't understand his math at first. But it was also empirical, as Kepler had access to new and much more precise observational data collected by his mentor Tycho Brahe.

I understand the frustration, but a lot of Geoff Hinton's foundational work was based on physics principles. For example https://proceedings.neurips.cc/paper/1993/file/9e3cfc48eccf8...

They're not identical but they are related. There's a series of approximations and simplifications you can go through to get from biological neurons to neural nets. Essentially the weights in the neural net end up corresponding to steady-state firing rates of populations of spiking neurons. See for example Chapter 7 of Dayan & Abbott's Theoretical Neuroscience.

Incoming Schmidhuber drama in 3…2…1…