Ariel
@redtachyon
Followers
2K
Following
2K
Media
135
Statuses
2K
p/hd | Big RL energy | 0.71 |research⟩ + 0.71 |engineer⟩ for now still at Meta | Soon to be MoE elsewhere | Prev. lead maintainer of Gym(nasium)
Paris
Joined November 2011
The correct form factor for robots is a humanoid. The correct form factor for digital agents is desktop/keyboard/mouse emulation. There's nothing about either of those that makes them innately optimal - there are way better designs that can milk a spherical cow in a vacuum.
3
0
6
Ok so the fp16 paper is nice and all, just one question - it seems to show that bf16 GRPO runs pretty consistently collapse, and fp16 is the savior who fixes it. ...is this something that actually happens? I get the occasional collapse, but it's not super common, even in bf16.
8
2
85
"The best way to resolve authorship order disputes, is to simply order the authors alphabetically." ~Aaron Aardvark
0
0
9
The funny thing about the job market in France is that if the recruiter messages me in French, I just know it's gonna be a huge lowball. Not to toot my own GPU, but what about my profile suggests that "up to 100k" is a super attractive offer?
3
1
74
No, that sounds pretty dumb, why on earth would you do that? Just because you can do something, doesn't mean you should do it.
Have you considered paying some of that money to the artists, authors, and programmers who created 99% of your product, but who as of yet have not received a single cent from you?
1
0
9
In all my shitposting about Meta I'm not too worried about breaking any NDAs - it's so silo'd that I learn more from @suchenzang's gossip than from any internal channels.
0
0
49
e/acc in sf: haha funny meme e/acc in europe: lol what a bunch of weirdos e/acc in india: WE WILL FIGHT AND DIE FOR YOU MY LORD BEFF JEZOS
2
0
32
It's only fine-tuning if it comes from the low-rank region of the parameter space. Otherwise it's just spicy retraining.
0
4
28
In other news, I just trained a powerful foundation model at the 14B - completely for free! You can find it at huggingface as Qwen/Qwen3-14B
The end of the transformer era marches slowly closer: we trained a completely attention-free foundation model at the 14B scale for only $4,000. The performance matches other models of similar scale, including transformers and hybrid models.
7
3
222
Fun fact of the day: France has a (surprisingly good) tax calculator that helps you navigate the (surprisingly bad) income tax system. If your gross annual salary is above 120k€ (~$140k), it warns you that it's so high, you probably got the pay period wrong (monthly/annual).
8
1
123
Hi guys, I just wanted to say that Meta is still an incredible company. MSL has a tremendous potential, and I am confident they will ship huge models, the best models. They are the best guys around. I have full faith in Zuck's and Wang's leadership. Pic unrelated.
9
3
111
Just submitted my resignation from Meta. Some might say that ~one year at the company is not that long, but I survived like three rounds of layoffs, so I think it’s fair play. For what it's worth, my team is great and it was an overall positive experience. But there are other,
154
65
4K
I'm not doing LLMs because I want funding lmao, I'm doing LLMs because I want a magic superintelligence in the sky, and nothing else comes even remotely close right now.
6
4
85
Worth noting that neither will actually get you a job
If you can only learn two languages, they should be: 1. one of (Rust, Zig) - will teach you to program at a lower level of abstraction, more aligned with the underlying hardware 2. one of (Haskell, Scala, OCaml) - will teach you to program at a higher level of abstraction
6
0
93