←
Artur Chakhvadze47,437 posts
Artur Chakhvadze avatar✓

Artur Chakhvadze ✓

@norpadon

intensifying the arithmetic @trymirai

◷ Joined October 2015▧ 3,133 Photos and videos
Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·8h

Like some systems papers for example use very shitty vLLM setups as their baselines. If you try to implement their thing in a properly tuned engine it is a net negative Average reviewer doesn’t know anything about vLLM configuration so they see big numbers and get excited

Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·8h

Well, “a lot” here is like 3 or 4. And by “glaring gaps” I mean the thing they are proposing doesn’t work in real life

Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·8h

I see a lot of very shitty papers with glaring gaps which got high scores at ICLR (334/344!). I feel like lazy reviewers tend to give high scores to papers which they don’t understand well enough. Should I become *that guy* and bring up issues in comments?

Jessica Hullman avatar
⟳ Artur Chakhvadze retweeted
Jessica Hullman@JessicaHullman·21h

I recently became aware of how confused even many academics are about Pangram. Please, please, look at the audits that have been done before you deliberately sow doubt. (link in thread)

Post attachment
Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·9h

There are of course many ways around it: you can parallelize RNNs using various fixed-point iteration methods. But it is all very ugly

Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·9h

I hate this line of work because it fucks up my beloved speculative decoding Please don’t fuck up my beloved speculative decoding

Post attachment
Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·11h

We had a short moment of clarity when things like scaling laws and muP were discovered and deep learning was starting to somewhat resemble proper science But then RLmaxxing happened and it all went to shit again

Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·11h

I think the level of reproducibility and generalisability of results in the post-training literature right now is significantly below that of psychology

Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·11h

Over the last couple of years I have been working across the full AI stack, from kernel engineering to post-training and harness design. I feel like the quality of papers in the most hyped-up subfields is so poor, and the level of rigor so low, that reading them is not productive

Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·Nov 17

> The professors could never academically refute me regarding genetics, archeology, mythology, or the fossil record Oh, right… Those woke professors simply can’t face the evidence of ancient subterranean civilisations

Post attachmentPost attachment
Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·Nov 17

And here are benchmarks for our upcoming quantization pipeline from @MNakhodnov (tested on Qwen3.5 9B) Mirai-S prioritizes quality and size Our new Dash maximizes raw decode speed I beleive we now have the most advanced LLM quantization pipeline in existence

Post attachment
Artur Chakhvadze avatar
Artur Chakhvadze@norpadon·Nov 17

Maybe I have some mutuals who can somehow connect me to the Qwen team? We make the best quantization and the fastest inference for Qwen on Mac, yet we don’t have early access to their models, while lots of other teams do. Lets fix it

Post attachment