Viewing profile — porridgeraisin

The training here is RL training, the rollouts there are not different from inference and have access to the same tools as regular inference.

There is some additional check, not sure what it entails, some people I know and I passed it, others failed.

There were multiple paths by multiple agents, not all of them led to the final exploit of hugging face. So its a bit confusing, but here's my reading anyways. Setup: the agent was …

You do a KYC and you can get access. It may depend on country's quality of KYC.

And also the most important few thousand accounts will get custom support and all sorts of patchwork and MS will put the effort to solve all their problems. So the customers that m…

For the longest time, I didn't appreciate the "AI-induced traffic" excuse. But seriously, I checked the rough github egress for our lab versus an old log from 2024, and there's an …

I dont pay for a chatgpt subscription, but sometimes I did use the web app for throwaway questions. GPT 5.5 Instant or whatever it was that they had was absolutely horrendous. Neve…

Did they do away with think? I think now you have to do it with /think

Yes. They have grad students too. This is just like having more grad students that don't need to be trained so the work you can get done is not bottlenecked by the number of people…

Yep. A communications professor where I did my MS says a 200usd/mo claude sub (which ant gives for free) does as much work as 5 grad students. It's mostly like you said, trying out…

The fact that by hand is emphasized so often and so often (see src/handgpt.h comments as well above backward()) makes me think it's AI. Anyways, it looks like gradients derived by …

Pretty funny that this is the thing Europe model is sota in.

The ground floor of apartment buildings in my country (not europe, india) is almost always reserved for parking. It is mandated though not uniformly followed in my city that there …

Yes, see https://news.ycombinator.com/item?id=49161078

That is because there is human annotated data there. Every session you or I used, then of course paid human feedback on repos (such as the recently famous example of meta forcing t…

Today's models depend on inference time compute to get these results. The inference time compute available on any claude subscription is not comparable to the ones used to get some…

stty is a very important existing command, but yeah sterm works.

Yep, I want really fast startup too. Of the two I use: xfce4-terminal starts utterly instantly without single instance feature. Kitty has the single instance feature and starts ins…

No reason, I always thought launching pinentry directly chose the right one out of pinentry, pinentry-curses, pinentry-emacs, pinentry-gnome3, pinentry-gtk, pinentry-qt, pinentry-t…

What? it is mostly literal nonsense. Moreover the last paragraph where they say it only reproduced on one machine just does not justify the pseudo-deep analysis in the whole docume…

I wrap `pinentry` (which is in the dependency trees of gnupg, pacman, gvfs, udisks2 and other software that you can de-facto depend on on desktops) inside my own simple askpass. $ …

Blackwell. You could do it with wgmma on hooper but then you'll only be able to run 1 CTA per SM. There are cases where this is OK, but most commonly 128x128 mmas are primitive in …

Most things weren't really reviewed that well as it was. It was always a small fixed set of people that actually played their part in the process in good faith. Those people contin…