The example sentences generated “only from neural data” at the top of this article seem surprisingly accurate to me, like, not exact matches but much better than what I would expect even from 10k hours:

“the room seemed colder” -> “ there was a breeze even a gentle gust”

Tangential to your point, if you collect 10,000 hours of brain scanning in exactly one damp basement, I wonder if perhaps the model would become very, very specialized for all of the flavors of "this room seems colder."

For the record, it was two basements -- we moved office in the middle -- and a bigger issue was actually overheating. But your point is basically right! The model is a lot better at certain kinds of ideas than others. Particularly concerning was the fact that the first cluster I noticed getting good was all the different variations of 'the headset is uncomfortable/heavy' etc. But this makes sense -- what participants talk about has a lot to do with what kinds of ideas the model can pick up, and this was more or less what we expected

Exactly. And honestly both this example and the one about the woman seemed to be what I would actually think/feel vs what I say.

Very interesting!

Yeah, agreed

This looks cool but I’m confused exactly how the compilation is supposed to work, and the docs[1] don’t make it any clearer. Is it promoting an LLM to try to get it to generate deterministic code to do the task?

[1] https://docs.a1project.org/guide/compilation

Related: LLMs trained on "A is B" fail to learn "B is A"

https://arxiv.org/abs/2309.12288

This website was a part of my education as a computer scientist and seeing it here again I'm curious for the full story of how this site was made. Who made it, what do they do now, is it part of a broader project they have?

There is only very basic info here: https://en.wikipedia.org/wiki/Project_Euler

Back in early 2000s, before hackerrank and similar coding sites, this is what my professors recommended for training programming skills.

Yeah my HS CS teacher recommended it but I don't think it was ever required

This is aesthetically very pleasing but the contents of the sections currently either don't load for me or load very slowly.

Thank you for doing this!

There must now be projects like the one the author describes that are at least 5-10 years old now, I’d love to see a comparison of ones that delivered their promises and ones that keep delusionally lumbering on.

Glad to see this on here again. This book is what bridged the gap between my undergrad degree and feeling like I could contribute to nontrivial real software projects. I originally found it because of HN

A good place to start with understanding the context for Tractatus is reading about Bertrand Russell’s logical atomism https://plato.stanford.edu/entries/logical-atomism/

In this critique have you considered that a way that that X can be logically determined by Y is by X just being Y.

Sure, but that's a vacuous observation. X being determined by Y is only useful when X and Y are somehow distinct. Simply having two different labels for the same object is not very interesting.

I would suggest that if one label is singular (the world) and the other is plural (the totality of facts) and you explain how the latter combine to form the former then the observation should be especially informative.

In general pointing out that two labels point to the same object can be very informative and this is relevant for the context in which Wittgenstein was writing https://en.m.wikipedia.org/wiki/Frege%27s_puzzles