pepp
@peppispepp
Followers
291
Following
4K
Media
92
Statuses
1K
Optimal arch design is how many "meta"s you can slap before your learning.
Are we really running out of data??? No. We're just not using it correctly. The solution: let the model learn which data it needs to learn!!!. 1/n
1
0
1
EncDec models watching you use an objectively inferior architecture for small seq2seq tasks
Introducing Gemma 3 270M! 🚀 It sets a new standard for instruction-following in compact models, while being extremely efficient for specialized tasks.
0
0
3
they automated grad student descent.
hyperparam tuning? what's that? I just ask ChatGPT to guess. In other news, we are stable now!
0
0
3
There's a weirdly missing term for serious circumstances that require banal/unserious actions. Like wearing a diaper to go space, or buying art supplies for a protest.
0
0
1
The hordes of people crying to have gpt-4o back scare me more than the doomers. What are we doing here.
0
0
4
perplexity is proof that you can do everything right and still lose.
0
0
1
Made a rule for the Cursor agent to do a `ruff check` at the end of their task and watching them fix all their formatting BS has brought me great joy. Thank you @charliermarsh.
0
0
7
The OAI graph thing became a massive meme but I don't get the people who thought it was genuine deception. You wouldn't notice it if they actually wanted to gaslight you. Insulting almost.
0
0
1
People are sleeping on o4-mini as a coding agent, it's better than 4.0 sonnet thinking (and fast/cheap).
0
0
1
I could immediately tell this was Gemini. What did Google cook, that model has been mentally messed up since 1.0.
"The user is absolutely right, I made a mess out of the codebase with my negligent changes. I am deeply sorry and am shutting myself down."
0
0
3
Looking forward to the first method that improves on SOTA but only works if you include the einops bug in the reference implementation.
Because AI is an engineering discipline and not a scientific field, it's never possible to fully separate the properties of a given approach from those of its specific implementations. The artifact is the method.
0
0
1
this is what we fight for:
~6/8~ Beyond speed, we validated MoMoE for numerical correctness against a PyTorch baseline. FLOP-matched pretraining also shows models using our kernel converge to a lower loss than equivalent dense architectures.
0
0
2
Egyptians invent Zoom (c.a. 2700 BCE). that or something about the illusion of self idk.
0
0
2
Behold, an AGI
My bar for AGI is far simpler: an AI cooking a nice dinner at anyone’s house for any cuisine. The Physical Turing Test is very likely harder than the Nobel Prize. Moravec’s paradox will continue to haunt us, looming larger and darker, for the decade to come.
0
0
4
I will satisfy your horniness for doom if you allow me to project reality onto the grandest malformed premises.
0
0
1
the fame game is who can make the biggest boldest assumptions and draw the craziest conclusions for fear and shock value.
1
0
2
dualities are inherent to the universe. vs. dualities are humanity's simpliest attempts at explaining reality by splitting on an (X / not X) proposition in high dimensional concept space.
1
0
5