gwern

35 ms·

541.
by gwern 2y ago
I think there may also be a lack of specification there. When you get more demanding and require more, the creative writing seems to be better. Like it does much better at things like sestinas. For all of those questions, there's proba
542.
by gwern 2y ago
No. It's bad news, because you can't see the rationale/search process that led to the final answer, just the final answer, and if training on the final answer were really that adequate, we wouldn't be here. It also is pr
543.
by gwern 2y ago
> Also... telling that they choose to benchmark against CodeForces rather than SWE-bench. They also worked with Devin to benchmark it on Devin's internal benchmarks, where it's twice as good as GPT-4o: https://x.com&
544.
by gwern 2y ago
Probably should try more base models, given that the weaknesses sound characteristically mode-collapsed, like ChatGPTese. (Prompt engineering tweets with base models is near trivial - just include a bunch of tweets.) It would also be more r
545.
by gwern 2y ago
> So why quit? Three reasons. First, I’ve felt for a few years now that the startups I’m seeing don’t seem so much like progress as just shopkeeping. This isn’t a dig: there’s money to be made razoring a thin slice off a huge market, and
546.
by gwern 2y ago
"Attacks only get better." It's 2m... now .
547.
by gwern 2y ago
> Some of the issues are probably HTML5 limitations, unlikely to be fixed any time soon (unless one wants formulas to become graphics). You can convert a lot of formulas into either Mathjax/Katex-style fonts or MathML, or even just
548.
by gwern 2y ago
You mean Discworld!
549.
by gwern 2y ago
I increasingly recommend against the Arxiv HTML version. I thought it had an acceptable start and they would fix the remaining problems and rapidly become on par with the PDF, but that seems to not be happening. The HTML version is serious
550.
by gwern 2y ago
Mirror: https://gwern.net/doc/math/1972-davis-2.pdf
551.
by gwern 2y ago
The Acton quote is given more extensively at http://www.strangehistory.net/2013/03/23/lord-actons-lost-wo... (the original book, _On the Writing of History_, Oman 1939, is hard to get).
552.
by gwern 2y ago
So, because you disagree with one nitpicky detail, you believe a blog post laying out the history and features and uses of howm is inferior to a landing page which contains nothing but 2 links, which themselves contain nothing useful except
553.
by gwern 2y ago
> Really? Selling goods online (Amazon) is not AGI. It didn’t take a huge leap to think that bookstores on the web could scale. Nobody knew if it would be Amazon to pull it off, sure, but I mean ostensibly why not? (Yes, yes hindsight be
554.
by gwern 2y ago
The Neukirchen link would be a much better HN submission than OP.
555.
by gwern 2y ago
The discussion of the SD Early Access Program starting on pg34 ( https://www.plainsite.org/dockets/download.html?id=332879335... ) is interesting. I didn't realize those Tesla-driving Twitter/YT accounts were i
556.
by gwern 2y ago
Today, for example, on Twitter, people are pointing out that Runway, which made a big deal out of how they were releasing DL models and not keeping them proprietary, has quietly deleted all its stuff on HuggingFace. (Much like how, say, Mis
557.
by gwern 2y ago
No. I think they use that web & ML software to help run their $1.3 trillion marketcap online social network company, on which, I am given to understand from US Congressional hearings, they sell ads.
558.
by gwern 2y ago
It absolutely is released like that. Please note that 'such as' does not mean 'all', 'a majority', or anything like that. It simply means 'one example'. (I eat many things 'such as' honey, y
559.
by gwern 2y ago
There is nothing weirdly cynical about it. This is a fact of life in Silicon Valley - that a lot of FLOSS is released for strategic reasons (such as building up a community before enclosing it to extract a profit), and not because the Grinc
560.
by gwern 2y ago
People may recall GameGAN from May 2020: https://arxiv.org/abs/2005.12126#nvidia https://nv-tlabs.github.io/gameGAN/#nvidia https://github.com/nv-tlabs/GameGAN_code
561.
by gwern 2y ago
> 'd say we decide on some arbitrary rules (like "only monospace fonts, no external javascript, only static content", I think that's a great idea. You can go a long way with just Unicode these days: https:/&#x
562.
by gwern 2y ago
This looks like it'd be quite useful for faking documents more convincingly. Existing handwriting fonts always have tell-tale regularities and there's so few that forensics analysis exposes them easily.
563.
by gwern 2y ago
"Most notably was Gauss at Göttingen, of whose style Abel once wrote: “He is like the fox, who effaces his tracks in the sand with his tail.”" This is incorrect. Abel was reporting someone else's comment, who remains
564.
by gwern 2y ago
> GPT-3 was terrible at jokes. Many LLMs are because it's hard to model puns in a stochastic parrot GPT-3 was great at jokes. The Navy Seals were hilarious ( https://gwern.net/gpt-3#navy-seals ). And the difficulty of
565.
by gwern 2y ago
You'd need a wide screen to make more than 3 articles pleasant, I think. The more you open, the more you'll want to rearrange or be able to increase/decrease or tile vertically. For example, on gwern.net, we have a very nice
566.
by gwern 2y ago
> So, no, pigs are not resistant but more resistant than cattle and humans. ...what does 'resistant' mean if not 'less likely to be infected than something else is'? Is it not inherently a comparison?
567.
by gwern 2y ago
This is common in cases of plagiarism, fabulists, & serial fabricators: "there's never just one cockroach in the kitchen". The same way Elisabeth Bik will spot one falsified image in a scientific paper and then immediat
568.
by gwern 2y ago
It's an unfortunate use of Youtube embeds. Aside from letting people uncharitably dunk on them, the Youtube here isn't even doing anything. The YT video in question is https://www.youtube.com/watch?v=_1Y8PwD5XDs ,
569.
by gwern 2y ago
When it comes to R&D, those are all very similar. Sora is another way to predict tokens, if you will. It's not like you have to choose that much: Sora is not using up all the compute that should've went to GPT-5. (It might n
570.
by gwern 2y ago
Marvin Minsky in 1970 (54 years ago) on how you can't just "turn off the X" when it is a powerful economically-valuable pervasive computer system: "Many computer scientists believe that people who talk about computer aut

More ›