simiones

13 ms·

211.
by simiones 11mo ago
The point I'm trying to make is this: the LLM output is a set of activations. Those are not "hidden" in any way: that is the plain result of running the LLM. Displaying the word "Blue" based on the LLM output is a s
212.
by simiones 11mo ago
The LLM proper will never answer "yes" or "no". It will answer something like "Yes - 99.75%; No - 0.0007%; Blue - 0.0000007%; This - 0.000031%" etc , for all possible tokens. It is this complete response that i
213.
by simiones 11mo ago
Even that is not necessarily true. The output of the LLM is not "Blue". It is something like "probability of 'Blue' is 0.98131". And it may well be 0.98132 for the other question. Certainly they only talk about
214.
by simiones 11mo ago
They can't. Biological neural networks have no resemblance to the artificial neural networks of the kind used in LLMs. The only similarity is based on a vague computational abstraction of the very primitive understanding of how brains
215.
by simiones 11mo ago
Edit: reading the paper, I'm no longer sure about my statement below. The algorithm they introduce claims to do this: "We now show how this property can be used in practice to reconstruct the exact input prompt given hidden state
216.
by simiones 11mo ago
> If I am understanding this paper correctly, they are claiming that the model weights can be inverted in order to produce the original input text. No, that is not the claim at all. They are instead claiming that given an LLM output that
217.
by simiones 11mo ago
I think their claims are limited to the "theoretical" LLM, not to the way we typically use one. The LLM itself has a fixed size input and a fixed size, deterministic output. The input is the initial value for each neuron in the in
218.
by simiones 11mo ago
> You could have several final states have the same unique prompt. They explicitly claim that the function is injective, that is, that each unique input produces a unique output.
219.
by simiones 1y ago
> You are just saying "well ackshually". I dare you to build a cabinet using the Hamiltonian. I double-dog-dare you. The Hamiltoninan (and Lagrangian) are much more amenable to actual physical calculations, at least on a comput
220.
by simiones 1y ago
Most of these things are misunderstandings of quantum mechanics, as we know it today. The main thing that is at the root of all of them is the word "things". In QM, the ground truth of the world is the wavefunction of the system.
221.
by simiones 1y ago
Sold in , not sold to . The GP meant: if you consider it legitimate to sell your product in Myanmar, you should obey the laws of Myanmar. If you consider the government is illegitimate, don't do business there.
222.
by simiones 1y ago
Which would be very relevant if anyone were trying to sue them for this - which no one is. The license establishes the limits of legal requirements and responsibilities. It doesn't shield you from criticisms and people being annoyed wi
223.
by simiones 1y ago
You realize that if relations really broke down to that level, then most likely NATO would simply exclude the USA, it wouldn't be all of the other NATO members leaving it, right? Also, if it gets to the level that the EU / Europ
224.
by simiones 1y ago
The biggest problem with Starlink's proposed solution would be that it would have been B2C - people in Greenland would talk to other people in Greenland through Starlink's satellites. That would put communication inside Greenlan
225.
by simiones 1y ago
Let me dissect the frog: 1. The lumberjack song (I'm a lumberjack and that's ok) is a Monty Python sketch [0] 2. The song goes from the singer being proud he's a lumberjack to being proud that he puts on women's clothing
226.
by simiones 1y ago
Yes, diacritics are very helpful if you want to understand the language. But they aren't any help if you're trying to read a travel journal about a place whose language you neither speak nor want to learn, which will be the case f
227.
by simiones 1y ago
A jungle is generally stable and doesn't grow in the sense that we say that GDP grows - definitely for large periods of time, with occasional exceptions. The fact that individuals grow in this jungle doesn't mean that the jungle i
228.
by simiones 1y ago
Especially if you haven't done this before, you start experiencing very strong hunger about 8-12 hours after your last meal. This is very, very much in advance of any kind of threat to your life or health from starvation. In fact, the
229.
by simiones 1y ago
The point, I believe, was that their laptop battery lasts longer than many people's, with any OS or battery, because they shut it down much more often than most.
230.
by simiones 1y ago
> Writing proper is a correspondence between marks and sounds to represent speech. I don't think this is the proper definition, since by this standard, Egyptian hieroglyphs, Chinese ideograms, Norse runes and many others would not b
231.
by simiones 1y ago
> The types say exactly what is proved. To whom? Fully formalized mathematics is extremely dense and difficult, and few mathematicians actually practice it, especially for complex problems. Just because the symbols are theoretically unam
232.
by simiones 1y ago
This is true in the sense that such a proof can't have "implementation bugs". But it can very much still have "specification bugs", where the formal specification doesn't actually match the fuzzy intention of w
233.
by simiones 1y ago
It's absolutely illegal in the USA to download a movie (or game, or song, or book, etc) that you haven't bought [0]. It could be argued that if you bought a movie, say on DVD, downloading another copy of it from an online source c
234.
by simiones 1y ago
You're mixing up several things, all of which actually boil down to the fair use exceptions I was mentioning. Making copies of a book you legally own for personal use is an established fair use exception to copyright. However, making c
235.
by simiones 1y ago
A download is a copy of a work. So, downloading a movie is making a copy of a work that you are not a copyright holder of - in other words, either you or the site you are downloading from are infringing on the copyright holder's exclus
236.
by simiones 1y ago
Not in any country that is part of the big international IP agreements (Berne convention, Paris Act). The only exception (sort of) is Switzerland. And the reason downloading copyrighted content you haven't bought for personal use is le
237.
by simiones 1y ago
No, this is just wrong. It might indeed use package-lock.json if it matches your node_modules (so that running `npm install` multiple times won't download new versions). But if you're cloning a repo off of GitHub and running npm i
238.
by simiones 1y ago
The point is still different. In PyPI, if I put `requests` in my requirements.txt, and I run `pip install -r requirements.txt` every time I do `make build`, I will still only get one version of requests - the latest available the first time
239.
by simiones 1y ago
The solution is not to go back to vanilla JS, it's for people to form a foundation and build a more complete utilities library for JS that doesn't have 1000 different dependencies, and can be trusted. Something like Boost for C++,
240.
by simiones 1y ago
Most people have addressed the package registry side of NPM. But NPM has a much, much bigger problem on the client side, that makes many of these mitigations almost moot. And that is that `npm install` will upgrade every single package you

More ›