gwern

26 ms·

391.
by gwern 2y ago
The narrow emphasis on autobiographical content we see today in art is really quite strange. You'll go on Reddit or Twitter and you'll see people insisting that poetry exists only to convey autobiographical content, and so good
392.
by gwern 2y ago
The Gemini models have become notorious for being rather uncensored and unrestricted, with an oddly base-model-like feel, since at least last November.
393.
by gwern 2y ago
https://pagedout.institute/download/PagedOut_006.pdf#page=3 to skip to the ToC.
394.
by gwern 2y ago
Sounds like they only evaluated GPT-4o and weaker LLMs like mid-last year?
395.
by gwern 2y ago
> It should be noted that the other xAI shareholders are/were supportive of the X takeover and all. No xAI shareholder has any incentive to be honest about that. If they criticize it, Musk will cut them out of all future investments
396.
by gwern 2y ago
If nothing else, you want to do this merger after the check clears from some new investors, for two reasons: it provides a high valuation you can use to minimize the wasted equity, and because it is highly dilutive of X.ai shareholders to
397.
by gwern 2y ago
On the bright side, Google's new Gemini-2.5-pro comes with no Holocaust-denying issues, and is definitely a lot better than Grok, and is also free and fast. (I've been using it the past 2 days and am pretty impressed - the first
398.
by gwern 2y ago
> Exactly. So then why draw conclusions about how SolarCity was a bad deal for shareholders (see parent comments)? You can absolutely draw adverse inferences from Tesla's refusal to report certain straightforward, useful numbers. Fo
399.
by gwern 2y ago
Yeah, looks like the subreddit has been taken over completely. You can see that the original founder of the subreddit had resigned shortly before... and his farewell-post has been deleted by the new mods. The other mods all seem to be inact
400.
by gwern 2y ago
> All user facing LLMs go through Reinforcement Learning. Contrary to popular belief, RL's _primary_ purpose isn't to "align" them to make them "safe." It's to make them actually usable. Are you claimin
401.
by gwern 2y ago
It sounds like their unit-tests cover abs(), but they weren't covering all of abs(), and were not reliably triggering the optimized codepath: > When doing the refactoring, they needed to provide new implementations for every opcod
402.
by gwern 2y ago
I recently learned about https://en.wikipedia.org/wiki/Self-verifying_theories which gives you most of multiplication while still being decidable, which is pretty crazy.
403.
by gwern 2y ago
I think "birds migrate" is a good contender for one of the most startling "humanity discovered that when ? and they thought what instead? And it took what to disprove it?" factoids: https://en.wikipedia.o
404.
by gwern 2y ago
You have discovered the Law of Fives: https://en.wikipedia.org/wiki/The_Illuminatus!_Trilogy#Numer... ; as it is written: > In the Erisian Archives is an old memo from Omar to Mal-2: "I find the Law of Fives to
405.
by gwern 2y ago
"Bye Bye Nerdie", S12E16, 2001 https://en.wikipedia.org/wiki/Bye_Bye_Nerdie : https://www.youtube.com/watch?v=u4rbW96R5GM https://www.reddit.com/r/TheSimpsons/commen
406.
by gwern 2y ago
That's a pretty major deal breaker for OP to leave out of his post touting it as something to build everything in your site on (especially for a tech blog)! Does it at least have a polyfill story? I see no mention of how to make it wor
407.
by gwern 2y ago
I never thought I'd see the day that LessWrong would be accused of being biased against near-term AGI forecasts (and for none of the 5 replies to question this description either). But here we are. Indeed do many things come to pass.
408.
by gwern 2y ago
As I always point out, all of the anecdotes claiming microdosing works are usually using street blotters as well, including the research papers which draw on crowdsourcing to circumvent the obvious problems with using schedule I psychedelic
409.
by gwern 2y ago
A MLP must be compilable to some arrangement of logic gates, so you could always try a tack like initializing everything as randomly-wired/connected MLPs, and perhaps doing some pretraining, before compiling to the logic gate version a
410.
by gwern 2y ago
And still is disappointing: > Reached for comment, LastPass said it has seen no definitive proof — from federal investigators or others — that the cyberheists in question were linked to the LastPass breaches. “Since we initially disclose
411.
by gwern 2y ago
> The MPV patch of v0.1 is without a doubt build 36 (e16bab8). The "Cripy optimization" turns status bar percentage rendering into a noop if they have not changed. This prevents rendering to a scrap buffer and blitting to the s
412.
by gwern 2y ago
One interesting thing about your archive is that I notice the site got hacked and used to distribute spam. The line " A special shout to our limo service San Diego for that one! " is a spammer insertion in https://web.a
413.
by gwern 2y ago
> They also do not, of themselves approach AGI without the addition of numerous other ingredients, particularly training data, alignment, and prompt engineering. That's also not true. A base LLM like GPT-3, without any (specialize
414.
by gwern 2y ago
This is a terrible analogy. Stone soup is disanalogous to generative AI models in almost every way that could matter. This analogy offers no insight, and at best comes off as a pretext for redistributive policies: "you mean our gener
415.
by gwern 2y ago
Humans are trained to follow directions too, and you usually don't have to explicitly tell a human you're playing chess against, "by the way, don't cheat or do any of the other things which could be validly put after the
416.
by gwern 2y ago
> Having visited the hermit state around 200 times through working as a tour guide He's a regime supporter who brings in desperately needed foreign currency to help pay for the nuke & missile program and elite luxuries. Hence, g
417.
by gwern 2y ago
> Not many people have Eliot down as an influential player in the history of AI — and they would be right. Her contribution has been for most part overlooked, even as folks begin to unpick the role of fiction shaping the trajectory of th
418.
by gwern 2y ago
> After he demonstrated this, I decided to try to help. I followed the wires that were wrapped. Two of them. To my surprise they were not connected at either end. This was insane, and yet his study - and my own observation - demonstrated
419.
by gwern 2y ago
No. The collapse is not as simple as simply shifting down most of the logits, so ramping up the temperature does little until outputs start degenerating.
420.
by gwern 2y ago
> Another important finding: Terry is by far the most popular name for a tortoise, followed by Turbo and Toby. Harry is a favorite for hares. All LLMs are loving alliteration. Mode-collapse. One reason that the tuned (or tuning-contamina

More ›