pizza

20 ms·

301.
by pizza 2y ago
This is the kind of thing I love to come across on a holiday break that seems ripe for side projects... :)
302.
llms.txt directory(directory.llmstxt.cloud)
106 points by pizza 2y ago|52 comments
303.
Improvement of visual temporal processing during training of light detection(frontiersin.org)
1 points by pizza 2y ago|0 comments
304.
llm-plugin-generator: `llm` plugin to generate plugins for `llm`(github.com)
2 points by pizza 2y ago|0 comments
305.
by pizza 2y ago
> They are there more as a political org so that facebook can say: "We employ 140 moderators in Kenya alone!" while they do indifferent work that facebook already can filter out. Why assume they're just token diversity hi
306.
by pizza 2y ago
just read the literature arxiv.org/abs/2410.01720
307.
A Simple Undo System(rxi.github.io)
1 points by pizza 2y ago|0 comments
308.
by pizza 2y ago
You're making a fundamental error in your reasoning. An LLM's training corpus being fixed doesn't limit the system's total information processing capability when used as a tool by a human researcher. While the LLM itself
309.
by pizza 2y ago
Ok, done. I can report to you that it helped me cut down my personal search space. Imagine what such a tool could do in the hands of a subject matter expert with rudimentary critical thinking ability and the faintest hint of a grasp of usin
310.
by pizza 2y ago
Ok, I can accept that that kind of downweighting or delisting of inconvenient posts is definitely suboptimal, sure. Does that mean we’re all going to sleepwalk into quoting from The Governance of China ? Unless our congress sees us all as
311.
by pizza 2y ago
The hysteria is reaching HUAC levels. You can read the statements from the congresspersons in support of the bill [0] and they all read like one variant or another of General Jack D Ripper in Dr. Strangelove quavering about how the communis
312.
Undecidability of Underfitting in Learning Algorithms(arxiv.org)
2 points by pizza 2y ago|0 comments
313.
by pizza 2y ago
> There is no risk of "authoritarianism" because the overwhelming political plurality is in consensus. Nor is there any concrete political opposition or protests. You could write three books on those two sentences
314.
by pizza 2y ago
> TikTok said it wasn't exfiltrating data and turned out to be lying. Is there a source for that? The Wikipedia article [0] states "There has been no public evidence of American TikTok user data being accessed by the Chinese go
315.
by pizza 2y ago
Would you like that with or without tokens?
316.
by pizza 2y ago
It's for KV caching. In most conversations that will mean inference. But you can do reinforcement learning using sampled sequences, and you could use KV caching to speed that up too, so that would be an instance where training coul
317.
by pizza 2y ago
Funny, this is one of those topics that has repeatedly given me great pessimism. If you take the conclusion that social media is corrosive to society then it should lead to a game of whack-a-mole against all social media. Yet I don't
318.
by pizza 2y ago
I do with Claude: https://news.ycombinator.com/item?id=42351796#42355665
319.
Stereochemistry-aware string-based molecular generation(chemrxiv.org)
1 points by pizza 2y ago|0 comments
320.
by pizza 2y ago
They’ve seen it all before because it’s all that they do so they know what to do when they’re doing it
321.
by pizza 2y ago
If I'm not mistaken, there's a great need for high quality, open source COT-style reasoning datasets. You might as well try to make one of those :)
322.
Forking Paths in Neural Text Generation(arxiv.org)
2 points by pizza 2y ago|0 comments
323.
by pizza 2y ago
If I’m not mistaken this seems like something that would be possible in Bend/HVM
324.
by pizza 2y ago
The internal mechanisms by which a model achieves a low negative log likelihood become irrelevant as it approaches perfect simulation of the true data distribution.
325.
by pizza 2y ago
Yes: for eg BPE, due to how it progressively pushes compound tokens of already seen - hence more common - subtokens to the ‘top’ of the vocab), you can train a model to do regression over vocabulary index for the next token from the curre
326.
by pizza 2y ago
It seems to me that it would make sense to just include more <BOS>-like meta tokens at the beginning in such cases, and have them as a prefixed scratch space that can be suppressed by treating them as non-output tokens.
327.
by pizza 2y ago
A visitor to physicist Niels Bohr's country cottage, noticing a horseshoe hanging on the wall, teased the eminent scientist about this ancient superstition. "Can it be true that you, of all people, believe it will bring you luck?&
328.
by pizza 2y ago
Just random musings, I have a somewhat visual imagination when it comes to math stuff so it is quite intriguing to see if I can get Claude to reproduce what I see in my 'mind's eye'. Here's an example w/o really get
329.
by pizza 2y ago
I've found a really nice workflow with Claude for generating high quality diagrams. Say you want to come up with, eg, a diagram depicting clustering with k-nearest neighbors: 1. Ask Claude: 'please generate an SVG diagram depictin
330.
by pizza 2y ago
If true, that reminds me a lot of "Anti-Soviet Warrior puts his Army on the Road"

More ›