combin8

499 posts

combin8 banner
combin8

combin8

@combin8or

GS-z13 Katılım Aralık 2021
732 Takip Edilen57 Takipçiler
combin8 retweetledi
kalomaze
kalomaze@kalomaze·
an unfortunate trend in modern open weights is the reluctance to release true bases in some cases this is arguably because there was no internal separation of concerns, i.e post training is more and more so being considered as a later part of the same overarching training process
ANTHROPIC_MAGIC_STRING@parafactual

@jmbollenbacher no base :(

English
10
3
190
15.2K
Marco Franzon
Marco Franzon@mfranz_on·
I have started uploading all my fine tuned models on HuggingFace. If you are into computer vision you will find there some interesting stuff for your projects ready to download and try.
Marco Franzon tweet media
English
6
3
47
5.6K
combin8 retweetledi
Jint3x
Jint3x@Jint3x·
@IntCyberDigest POV: you are downloading packages in 2026
GIF
English
4
97
1K
82.8K
combin8 retweetledi
Charlie O'Neill
Charlie O'Neill@oneill_c·
Really enjoyed this. We covered why I lasted only a few days into an Oxford PhD, why you should learn RL by touching nothing but the config and watching the curve go up, and why the intelligence ceiling of a specialised open-source model now subsumes the frontier for most real tasks. Also the story of negotiating with @tuhinone in my pajamas
Madison Kanna@Madisonkanna

How to become an AI researcher with @oneill_c Charlie co-founded Parsed to build specialized open-source models that can outperform frontier labs. I first met Charlie when Parsed was acquired by Baseten, and now he leads our model development team. Charlie is one of the smartest people I know, and I had the pleasure of talking to him about: 0:00 Intro 3:13 Leaving Oxford to start a company 6:37 Becoming an AI researcher 15:37 Developing a unique POV as your moat 22:04 Parsed origin story 26:01 Big Token, the case for open-source models 33:40 Post-training, fine-tuning, specialization 46:52 Will open models catch up with closed models? 51:50 AI-led job replacement vs job creation 54:45 How to get into inference engineering This is one of my favorite conversations I’ve had in a long time. Made with @ad0rnai behind the scenes. Enjoy!

English
8
16
321
31.5K
combin8
combin8@combin8or·
As it’s a MoE, that’s only ~3.5-7GB active. Should be really fast with dspark (which they shipped for Qwen)
English
0
0
1
25
combin8
combin8@combin8or·
@jwt0625 The lengths we’ll go to for 50 ns
English
0
0
1
34
combin8
combin8@combin8or·
@ivanfioravanti @antirez Curious how many tokens, for those of us playing along at home. That’s a pretty impressive jump.
English
0
0
0
23
Ivan Fioravanti ᯅ
Ivan Fioravanti ᯅ@ivanfioravanti·
Pushing GPT-5.6 Sol Ultra on trying optimizations of DwarfStar on M3 Ultra, while GPT-5.6 Luna is working on finding the best Q2-Q4 mix. 💪 Thanks @antirez for this repo. It opens up endless research and testing possibilities, backed by many mathematical checks and constraints.
Ivan Fioravanti ᯅ tweet mediaIvan Fioravanti ᯅ tweet media
English
11
2
60
6.9K
combin8
combin8@combin8or·
To win: Develop a portfolio showcasing Bend2 & OSS it. Target small well-defined FOSS projects already supported by foundations/companies. Provably correct software is low maintenance, so show off what a small team can do. Contract out to build the portfolio more & keep at it.
Taelin@VictorTaelin

Bend2: - Programming language - Near C speed (on CPU) - Near CUDA speed (on GPU) - High-level (JS-like closures, objects, recursion) - Lean-like proofs (force AI to write proofs = you get 0 bugs) Everyone is using it. How much would you pay to use it? Be *honest* either way

English
0
0
0
103
combin8
combin8@combin8or·
@voooooogel Zig:Go :: Rust:Haskell? I’ve been away for a bit; has anything changed other than monadic perception?
English
0
0
0
372
thebes
thebes@voooooogel·
1. insane post lmao 2. this is one turn of long-standing disagreements between the rust and zig crowds, which predate LLMs, and this rewrite couldn't have been grown in a lab to better piss off the zig side 3. contra the bun post, the claude ported rust is extremely bad
rohit@seatedro

crazy blogpost by andrew kelley just dropped

English
18
6
462
80.3K
combin8
combin8@combin8or·
Muttering is a delightfully intuitive way to describe the signals from what the authors call lexical task heads, which direct LLM behavior. Good paper, and well written to boot.
David Bau@davidbau

There is a phenomenon that my student @ericwtodd calls "muttering," when LLMs internally articulate what they are doing. E.g: writing Python code, they keep encoding the token "python" without saying it. And it can be causal. From Jojo, Jacob, Francisco, at ICML this week:

English
0
0
1
58
combin8
combin8@combin8or·
@m_sirovatka You all have some awesome sauce, congrats! Wish I’d known about the series A. Do you post investment info anywhere beforehand?
English
0
0
0
17
combin8
combin8@combin8or·
@jwt0625 Ah, right. Love your photonics posts, btw. Just checked out the thread you linked and those photonic integrated chips are mind boggling (thanks for links to the papers)
English
0
0
1
29
Beyang
Beyang@beyang·
Kinda crazy to look back on the evolution of the agentic coding frontier in the last 1+ year
Beyang tweet media
English
5
5
57
6.2K
combin8
combin8@combin8or·
@banburismus_ Could this apply to LLMs? Seems the study is about vision. And I’ve only skimmed it 😅
English
0
0
1
33
Tom McGrath
Tom McGrath@banburismus_·
this is a really great example of how much clearer things look from the geometric perspective: we should really think of the classic curve detector family in inceptionv1 as points on a 'curve manifold', and the wiggles in this manifold allow readoff of different semantic information. I think this points towards why manifolds are such a useful representational strategy: by arranging its neurons to work together, it can actually represent many ideas with a single subspace
Goodfire@GoodfireAI

We also revisited an interpretability classic: curve detectors in InceptionV1. Neurons and SAE features turn out to be fragments of one continuous orientation feature, and the block *also* contains higher-order Fourier harmonics that hadn’t been described before! (7/9)

English
2
10
73
4.2K
combin8 retweetledi
Edd
Edd@erla_ndpg·
Continuing my previous post 9. PuzzleMoE arxiv.org/abs/2511.04805 This paper compresses MoE by merging experts with dual masks, then bit-packing mask/sign info into unused BF16 exponent bits for efficient inference.
English
1
2
10
1.1K