qwertyforce

As I understand, the problem of continual learning is that if you continue training, it will "overwrite" previous knowledge and this CPU will be lost

We need a graph of empathy vs number of people in organization

Maybe intercept firmware update and then reverse engineer it and find rce?

Me too, hacked into my old fitness tracker, but using grok, because gpt6 refused. Interesting times we live in. https://news.ycombinator.com/item?id=49714479

maybe it is about staring at thinking traces?

but my job is also my hobby, it is the center of my life i cant pretend i dont care

thanks!

I think the worst part is that some companies may be okay with it. They will pay attention only when it's messed up beyond repair. And it’s possible that fixing it will be a task for more advanced models, not humans. Do they win or lose at the end of this martingale? I am not sure

fwiw pangram says it is 100% generated

I blame the ML engineers who work on these recommendation systems. They chase simplistic objectives like CTR, time spent, and so on, which can be gamed by this kind of content. This creates huge positive feedback loops in which popular content becomes even more popular and forms “metas,” while models train on clickstream data they themselves have influenced. They could try to fix this, but they won’t, because no one is asking them to

Original title: How I (Spiritually) Won comma.ai's Compression Challenge

It will probably suffer the same fate as the most-upvoted discussion of all time in the GitHub Community repo: https://github.com/orgs/community/discussions/66188

no reaction

thats exaclty why i prefer codex

noticed that firefox gives 2x kHashes/s more than chrome (1000 vs 500)

yep, this is the only moat they will have against chinese AI labs

That is a great question. IMO, major LLM players currently have a large enough user base to generate training data from their users (questions and user provided answers, corrections, etc). So, if StackOverflow dies, it will become harder to keep up with closed source models

I think there should be some graph algorithm for this, to find a bottleneck in a graph

it's sad that startups become corps and decay. this article is the perfect illustration, from the bio, to the llm slop content of the article. Just sad it has to be this way

Looks fake

dx1*=2; dy1*=2; dx2*=2; dy2*=2;

used 3090 is the best option imo

https://www.youtube.com/@animatedai/videos for visibility

you can also take a few steps back from the screen, it works too

Yep. Most of my google queries are prepended with "reddit" or "stackexchange". I particularly like stackexchange, lots of useful information there.

That's why people should consider hacktivism (at least DDoS)

highlight.js is for syntax highlighting (code looks pretty). I've just checked, in Chrome site works with js disabled

Thanks for clicking! Clicking 'use IPFS' makes all the images point to the local IPFS gateway on 127.0.0.1:8080, so browser can get the images from people who pin (distribute) them, like in a bittorrent. But even if you installed it, it wouldn't work because currently I don't pin image files on IPFS. IPFS daemon consumes a lot of ram and cpu resources, so I decided to shut it down.

It seems it was introduced after 1.9.0

https://pytorch.org/docs/1.9.0/generated/torch.nn.functional...

I read similar chapters of several books on the topic simultaneously, when tired of reading, watch several youtube videos on the same topic.

More points of view --> better understanding