grungegun

5 ms·

1.
by grungegun 2y ago
In a sense, I spend most of my time on material after quals because it's so much harder to understand.
2.
by grungegun 3y ago
So there's no performance gain for quantization enabled by the transformer architecture? It seems very strange that quantization works so well since in most of my experiments, the internal model weights of mlps look random.
3.
by grungegun 3y ago
Does anyone know if this works on vanilla deep networks? These quantization articles always seem to target LLM's which leads me to wonder if there's something special about the LLM architecture vs a vanilla deep architecture.
4.
Ask HN: API for Custom Cards?
2 points by grungegun 3y ago|0 comments
5.
by grungegun 3y ago
I taught one of my younger brothers a simple sign language when I was younger. His speech was delayed by like 8 months compared to the rest of us. Can't say for sure, but there's a need to emphasize spoken communication with the s
6.
by grungegun 3y ago
This is somewhat overblown, and only really an issue for some particular theorems and obscure corners of mathematics. And this isn't a recent problem, Hilbert's proof of Nullenschtatz in the early 20th century had logical errors,
7.
by grungegun 3y ago
the infinite density is actually after a finite amount of time, since black hole collapse has already happened, which is why it matters that they are predicted to have infinitely dense points at their centers. The key is that this isn'
8.
by grungegun 3y ago
would be interesting if you could used deepfloyd first for image composition, then apply stable diffusion after for purely stylistic modifications
9.
by grungegun 3y ago
What does this mean? Isn't the quality of a model determined by how easy it is to get a good picture?
10.
by grungegun 3y ago
Loved this book. When I first started programing, I was convinced that text applications weren't real programs, and the visual aspect of the book pulled me in to be a life-time programmer :)
11.
by grungegun 4y ago
For diversity, Dalle 2 has a random chance of injecting "women" or "black" after a prompt. When this happens, at least for me, it generally destroyed the quality of the images. Probably "King" was identified as
12.
by grungegun 4y ago
I'd argue that the incredibly fast rise in quality over the last year alone is what most people are interested in. DALLE and GPT-2 were always able to make heaps of trash. The trajectory of the quality of DALLE-N and GPT-N is what inte
13.
by grungegun 4y ago
yep, that sounds right to me. also I see you're the author of the Noether article from Ars Technica. I enjoyed the read.
14.
by grungegun 4y ago
Yes, the quantumness of a differential equation is not a property of the differential equation, but a statement about one possible taxonomy of differential equation. Then, whether quantum-type diff eq's have unique properties pertainin
15.
by grungegun 4y ago
> Differential equations don’t give rise to physical systems. Sure, but differential equations describe physical systems, and there is a canonical way to derive a quantum differential equation from a classical one by quantifying the clas
16.
by grungegun 4y ago
The differential equations which give rise to physical chaotic systems operate on quantum states interacting with each other at the atomic level, so there has to be a link of some kind.
17.
by grungegun 4y ago
It looks like in the results section eqs (8)-(13), they give the main thrust of their idea. To demonstrate that not all machines are reversible they have to take the limit of the number of qubits of their machine as n goes to infinity, so t
18.
by grungegun 4y ago
Actually, it doesn't increase much faster than linear. It's quasilinear, and so in O(n^(1+\epsilon) for all \epsilon > 0. There are a variety of such results, so most objections don't hold weight. The paper you linked to n
19.
by grungegun 4y ago
False, except for a naive implementation, see: https://en.wikipedia.org/wiki/Quantum_threshold_theorem If you look at the theorem, it is definitely sub-exponential in the number of qubits needed to correct for error.
20.
by grungegun 5y ago
For reference, see the book High Dimensional Probability by Vershynin. It's free online. See Theorem 3.1.1. It proves that a sub-gaussian random vector is in some sense close in norm to sqrt(n) where n is the number of dimensions. Most
21.
by grungegun 5y ago
>It didn't have very many interesting properties I could find Usually notation is defined for higher order functionals when the need arises - when a new result is found to have interesting properties. Topological proofs on function
22.
by grungegun 5y ago
The doctors don't believe the story. The woman does. The doctors believe that the woman's belief caused her early death, but that's independent of whether they accept that story as true - people (the woman) can convince thems
23.
by grungegun 5y ago
Is there a reference for this to verify that this isn't just some historical reenactment? I'm trying to be more careful after I was once pranked by a friend who showed me a black and white video, claiming it was real, of people pl
24.
by grungegun 5y ago
For those interested, there is a fairly basic WYSIWYG editor for this: http://www.tikzedt.org/ It's useful to throw together basic 2d diagrams, which can then be refined later
25.
Ask HN: Small, open source, C based, file encryption?
2 points by grungegun 5y ago|1 comments
26.
by grungegun 5y ago
Plus, when you double click a spot in a pdf you generated with latex, Sumatra opens up your default .txt application with the line highlighted.
27.
by grungegun 5y ago
You're right. I've looked at quantum codes before, but apparently my understanding was lacking - thanks for the info.
28.
by grungegun 5y ago
There's one big issue with QC currently. There is no mathematical demonstration that QC is scalable. All the speed-ups need some form of error correction to scale the way digital computation has. All hardware advances are just increase