gwern

51 ms·

781.
by gwern 3y ago
OP is echoing, with a Catholic-centric viewpoint, a point that's been made in much greater depth by others: I suggest https://gwern.net/doc/economics/2014-robbins.pdf and then more briefly, https:/&#x2F
782.
by gwern 3y ago
You can see this as an endgame of 'commoditize your complement' ( https://gwern.net/complement ): you're happy to contribute to commodification while you are the small scrappy player, but at some point, if yo
783.
by gwern 3y ago
> For better or worse, the quality of output LLMs can produce based on such publicly available data appears to be peaking. Recent work has documented that GPT, for example, has begun to exhibit declines in the quality of its output.3 It
784.
by gwern 3y ago
I'm not really following your point. If the land/space is not the bottleneck but the grid, then that makes the idea of leaving old panels in place all the better. As the electrical output falls by the aging panels and underutili
785.
by gwern 3y ago
No, they shouldn't know how. Most of what there is to know about fixing a coffee machine is useless tedious bullshit. No one's life is enriched by knowing which error messages mean what, and which are misleading, or where the part
786.
by gwern 3y ago
You're really going to make excuses for Slack here? 'o'/'о' is just about the easiest possible homograph attack ( https://en.wikipedia.org/wiki/IDN_homograph_attack ) that there is: > Whe
787.
by gwern 3y ago
I wasn't able to find that quote myself, and I was suspicious because I've skimmed _The Publisher_ in the past (trying to verify a quote about why _Time_ magazine picked red) and the Gemini anecdote doesn't sound like the aut
788.
by gwern 3y ago
> For any employee, the soul-stealing complexity of office machinery such as fax machines, copiers, PCs, voicemail and even coffee makers gives everyone ample cover to studiously never learn how to use them. But the same blank stare acco
789.
by gwern 3y ago
The RNN in question is over-parameterized, so most loss functions wouldn't be expected to coincide with MDL (unless maybe you added in heavier regularization than just L1/L2). There is no incentive towards simplicity. But it is al
790.
by gwern 3y ago
If by design elements you include vector images, you could try https://www.recraft.ai/ or Adobe Firefly 2 - there's not a lot of vector work right now, so your choices are either the handful of vector generators, or ju
791.
by gwern 3y ago
This doesn't seem all that impressive when you compare it to earlier work like 'g.pt' https://arxiv.org/abs/2209.12892 Peebles et al 2022. They cite it in passing, but do no comparison or discussion, and
792.
by gwern 3y ago
Back in 2020, Google was saying 16x16=256 words: https://arxiv.org/abs/2010.11929#google :)
793.
by gwern 3y ago
Disappointing to see no GPT-4 here.
794.
by gwern 3y ago
Ah. A little surprising, MJ's tend to be messier and more textured. But yeah, I too know the sorrow of getting MJ to generate specific text in images - funnily enough, I too wanted a question-mark for my cat image (for https:/&
795.
by gwern 3y ago
DALL-E 3? I've seen some similar images out of it while trying for more pointilist views. (DALL-E 3 has a certain obsession with regularity & centering.)
796.
by gwern 3y ago
Was it these dudes? https://www.thebeliever.net/aristocrat-inc/ There were a lot of armed-robberies involving RAM and chip deals done over the phone back in the 90s.
797.
by gwern 3y ago
Doesn't have to be 'trivial'. It could be a trillion-parameter MoE LLM spitting out 10,000-large embeddings. But a lot of that will act linearly and can, in theory, be put into your linear model easily & profitably.
798.
by gwern 3y ago
> When did the Tasmanian pre colonial people have crops? My 'they' in "they lost tech like crops" referred to the Canary Islanders, as discussed by OP where the botanical evidence shows crops disappearing gradually ov
799.
by gwern 3y ago
> I wrote this story [publisher December 2023] using only my own human cogitation, with no computational input, except for the excerpts of Good Stories’ machine-formed fictions, which were generated by ChatGPT 3.5 (May 3, 2023 version).
800.
by gwern 3y ago
Sounds like a parallel to Tasmania: the tech level kept ratcheting down, as they lost tech like crops and had no way to regain them, so every loss was permanent. Slowly, by simple drift in small populations, and occasionally abruptly, in en
801.
by gwern 3y ago
I wonder how many of these are https://www.hyrumslaw.com/ in action? Like the long-channel name seems like something you'd lock down as soon as you notice you made the classic error of trusting client-side validation o
802.
by gwern 3y ago
Cats are like hands: they are hilariously hard for generative models and then after thinking about it, you realize that cats/hands really are hard. I mean, look at photos of a black cat curled up where it might have its paws sticking
803.
by gwern 3y ago
Yes, the 'special effects' effect will kick in. Within a year or so, you'll spot this easily, quite aside from the more obvious issues. (That Landrover captioned 'DANDOVER' - is this still using BPEs?!) Aside from
804.
by gwern 3y ago
This seems like a rather misleading submission title.
805.
by gwern 3y ago
I don't understand your confusion. I thought the Quanta article did a perfectly reasonable job of explaining the difference between stabilizing selection for a single fixed, permanent, unaltered optimal phenotype, and having simply a b
806.
by gwern 3y ago
You sometimes read about them if their hustling ends in a particularly messy fashion: https://www.newyorker.com/magazine/2024/02/12/a-teens-fatal-... RIP Zoomer Musk.
807.
by gwern 3y ago
You can do a lot. 'Digital sundials' come to mind: https://en.wikipedia.org/wiki/Digital_sundial https://gwern.net/doc/math/1991-stewart.pdf But yes, you are restricted to linear th
808.
by gwern 3y ago
Indeed. As I said, "The effect of the selection is permanent and has already 'bred true' and will persist indefinitely unless undone by some other force."
809.
by gwern 3y ago
> The reason this happens is that the recessives that influence the one parent's trait are often different from the recessives for the other parent. Therefore the child winds up expressing fewer recessives than either parent. That&#
810.
by gwern 3y ago
Tell me that when Nigeria gets as rich per capita as countries which do have good rule of law. (I'll wait.)

More ›