krackers

I sure hope not: https://en.wikipedia.org/wiki/Train_to_the_End_of_the_World

>This spills over into other processes wanting to use the GPU, namely the WindowServer.

Why does this spill over? Unlike CPU which is multiplexed by the kernel's scheduler (so infinite loops can't lock out other programs), is the GPU not multiplexed in the same fashion?

How do you know that what is being proved in the lean code is the same as the millennium prize criteria though?

>For all trains to be self-driving

Given that we have self-driving cars, isn't this easier if someone really wanted? I guess compared to cars the marginal savings is not worth it though.

I've never quite understood the hate either. But the way I see it is that it's not fundamentally inferior to explorer, but rather just has annoyances and warts that Apple has bizarrely not fixed for decades which makes tedious to use for anything other than casual browsing.

You can still navigate by path via cmd+shift+g (this works in the file picker too), but I feel like the popup animation used to be a lot quicker on older versions. You can customize views, but to get them to actually stick you have to use an arcane incantation (delete all existing ds_store, then go to view options menu and hit "make default").

You can cut-paste but you have to use cmd+c and cmd+option+v instead of cmd+x. There is an option to merge, it only appears under conditions: only when copy/pasting and only when "alt" is held in the dialog I think.

It also is really bad at browsing networked volumes, partly due to poor defaults as it tries to litter ds_store and generate thumbnails.

Opening a new tab always takes you back to the root directory instead of cloning your existing view. All of it is stuff that could be added in a simple point update rather than a ground-up rewrite.

Elliot Glazer (FrontierMath lead) traces how it snowballed over time

https://x.com/ElliotGlazer/status/2096298696438906934

For it to be memory safe, do you have to disable the JIT?

It mentions a type confusion in V8. Is it possible to trigger that without JS enabled?

The "all chromium versions" part of the title is also misleading, most browser CVEs do not distinguish between "untested lower bound" vs "affects all" (even though it seems like it'd be trivial to bisect).

Automatic Captions are not ADA compliant.

https://www.speechpad.com/blog/ada-compliant-captions-vs-aut...

If you have a risk of getting sued because someone notices a transcription error versus just not uploading videos, people will choose the latter.

How? Can't apps link you to the app store where you have to make an in app purchase?

>almost nothing is safe

That's a bit of an overstatement? There's a list of things that you _can_ call, and fairly useful ones too like `write`

https://man7.org/linux/man-pages/man7/signal-safety.7.html

Feels weird that after their attempts to shut down corellium they themselves shipped the necessary bits to allow this

https://github.com/wh1te4ever/super-tart-vphone-writeup

They laughed at objective-c back then, now the shoe's on the other foot

>It's crazy, because 15.ai could have had the venture-scale outcomes if he'd fundraised and turned it into a SaaS product

Or equivalently could have gone down as a "deepseek" moment in TTS if the model was open-sourced (I don't think even a paper was published). It's a bit strange, he never made any profit but also never ended up publishing anything even after it was shut down.

Another benefit of open models is that you can mask out out words from logits directly when sampling. I wonder if anyone has put together a "desloppifier" for e.g. DeepSeek

There might be precedent for being able to detect changes in magnetic field

https://www.pbs.org/newshour/science/dogs-poop-in-alignment-...

but then again maybe not

https://www.sciencedirect.com/science/article/abs/pii/S15587...

Or take the ribbonfarm-pill and jump straight to Gervais Principle

https://news.ycombinator.com/item?id=46657678

ref: https://www.youtube.com/watch?v=IU4ByUbDKNc

>fails to understand the gumbel softmax technique

I think in this case it doesn't help that there are multiple watermarking schemes, and the easiest for people to understand is the red/green scheme by Kirchenbauer et al. (https://arxiv.org/pdf/2301.10226), which does technically distort the logits (but I'd argue only in cases where you wouldn't notice it anyway).

I wasn't aware of this gumbel softmax scheme, it seems you're referring to https://simons.berkeley.edu/talks/scott-aaronson-ut-austin-o... ? That's really clever as it doesn't even distort the logits, basically cryptographically indistinguishable from a "real" random sample unless you have the key.

The actual scheme Claude uses seems to be neither of those two though, they say it is SynthId-text which seems to be tournament sampling based.

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