krackers

I don't understand Gruber's points either, I wonder if there is some fundamental technical misunderstanding. Does he think that the logits should be sampled from in a "pure" manner without introducing any other bias? Does he know that there's already a sampling temperature, and that most providers have probably moved on to sampling strategies other than top-k? Does he know that the word choices have already been altered irreversibly during RLHF which is how you get the obvious Claudism like "load bearing" and "seams"?

Perhaps it would be useful to publish examples of samples with/without watermark. I'd suspect that the variability from simply sampling repeated times would dwarf any semantic differences you'd detect with the watermark.

A similar project (LLM trained only on vintage material): https://talkie-lm.com/introducing-talkie

What would be your test to determine that?

There's a better explanation of it at https://alexkritchevsky.com/2024/12/18/triple-products.html

>maybe some other biological organisms can experience

I don't think this even matters much in terms of moral patienthood. Many people might agree that cows can "feel pleasure" or "have fun", but yet they are slaughtered nonetheless. Dogs are only offered this not necessarily on account of their intelligence or the strength of their qualia but more because they're cute.

I think it's limited to using base models (non post-trained), because the post-training would skew the logit distribution. There are ways to "coax" post-trained models back into behaving "like" a base model, I wonder if the benchmark could be unofficially updated with those somehow.

I think ARC-AGI-3 specifically forbids harnesses. This means that you're basically limited by the context window, so it's no wonder that LLMs can't do that well. Unofficial versions that use a harness seem to be doing fine on it.

There's a krazam video satirizing that https://www.youtube.com/watch?v=CJwPb_76jqU

>I think it's a really clever idea that you could measure an LLM's prediction abilities and language understanding by some sort of held-out compression metric

This is the premise of https://huggingface.co/spaces/Jellyfish042/UncheatableEval

They had it as a hidden feature up till the nexus 5 or 6 I believe. Someone bothered to include hardware for an LED even if the stock software never supported it.

>I’ve just gone through a butt load of resumes

Perhaps she said "bulk load of resumes" and you misheard it as the funnier version?

The "14 Corporate Flavors" had me rolling. This seems less like encouragement than the stick though. I wonder if you took the same principles and rewrote it to be more compassionate instead (maybe lines encouraging it to meditate a bit or something, I don't know) you'd get much better results.

>roughly an order of magnitude cheaper to cook at home

Why should this be the case in principle though? Do economies of scale not apply to food preparation?

This deserves its own hn post!

Probably someone installed a custom theme (you could do it on osx back in the day, see e.g. "Aqua Extreme").

I think most of the complexity is around the concept of disk images. If the app was just distributed in a zip file (like many in fact are) then you just unzip it, and the application is right there. You can move it to /Applications if you want (and many apps offer this) but it's not necessary.

Disk images seem to be much rarer on other platforms, and are usually only used for images of optical disks. Even the concept of "ejection" can be confusing, especially since when mounted disk images appear almost identical to actual physical removable media. I'd like to hear the reasoning behind why mounting disk images isn't relegated to a vestigal "power user" feature by now, since it seems like zip files (especially in appledouble format) can serve the same needs for most normal user flows.

Why is providing a mechanism to view your own data a security risk? You cannot view application data (even in a read only fashion) without root on android. Why should the app be able to read its own data, but the human not able to?

Tulip mania!

>It’s an information theory question similar

How would you approach it from an information theoretic sense?

That's the whole point of the article, that regular expressions you use day-to-day are not strictly regular.

Terminal.app is amongst the best, if these old danluu benchmarks still hold https://danluu.com/term-latency/

The nature of preference-ranking in LLM post training means that "slop" is something that the average person actually prefers. Maybe there's an argument that it's not intentional but just emerges as a form of reward-hacking (e.g. people like a little bit of slop or sycophancy but the model dials it to 11) but I imagine if the end result wasn't rewarded ultimately that would get penalized.

The fact that a clearly sloptacular image won first place seems to corroborate that indeed people do like it. Until people's tastes change and they learn to recognize and immunize themselves against slop, the models will continue to generate it.

I guess maybe the reasoning is that because it's a watched event it's useful to have libre/public domain photos of it?

The old (now legacy) apple documentation roughly seemed to follow this pattern I think.

Amazon is the best example of this. They can surely implement proper search or filtering, but it's not in their interest to do so.

See also https://news.ycombinator.com/item?id=49135245

Author summary thread on twitter https://x.com/AlexiGlad/status/2083230922196107288

The idea seems mind-bogglingly simple, instead of

  y    = model(sample_latent())   # generate one output (from noise, a mask, …)
  loss = recon_loss(y, x)         # score it against the data target x
  loss.backward()

you do

  losses = []
  for _ in range(K):                   # explore K candidate outputs
      y = model(sample_latent())       # generate one candidate
      losses.append(recon_loss(y, x))  # score each against x
  min(losses).backward()               # train only the closest candidate

I guess the intuition is that if you just generate one output and score it, your loss function forces it to split the difference between all samples and converge to the average (even if it produces _a_ valid output, if it's not the exact target being trained on it gets penalized). Whereas if you do basically best-of-n, it's not penalized for generating other valid samples as well.

They have become cheaper and cut costs over time. LED bulbs bought around the l-prize time when the technology was new are heavy, you can feel their metal heatsink. New LEDs from the same reputable brand are comparatively cheap.

There's two separate issues where LEDs still have a hard time competing with incandescent bulbs: CRI (really R9 or other deep-red measures, because CRI stats are sort of gamed) and flicker. Most LED bulbs also have poor heatsinks and are more likely to die if kept inverted or in enclosed fixtures.

All of these are individually solved problems, you just can't buy a bulb that's an e26 retrofit that satisfies all of them for whatever reason. If LED bulbs actually lasted 10 years, then there'd be 10 year warranties. Yet even the most expensive boutique LEDs (e.g. from YujiLED) are only 3 years at best.

Modern browsers should no longer score 100 on Acid 3 though.

>By April 2017, the updated specifications had diverged from the test such that the latest versions of Google Chrome, Safari and Mozilla Firefox no longer pass the test as written