Xavier Belda

Approximate knowledge of many things

[Xavier Belda @XavBelda]

@sun_onyour_back @vikhyatk It's what remains after burning too many tokens

[elie @eliebakouch]

mythos will be bad ON PURPOSE on ai "frontier llm research" tasks, this is very very sad for the research community

also the fact that this is un purpose not visible to the user is crazy

[Florian Brand @xeophon]

if a model refuses, it should score as 0 on that task

[Xavier Belda @XavBelda]

@TrelisResearch This makes a lot of sense. I have the intuition Diffusion LMs should probably shift to denoising the internal token embedding instead, progressively unmasking the meaning/representation of each token instead.

[Xavier Belda @XavBelda]

@EIFY @papers_anon Absolutely, why not just test it on NanoGPT? Muon and NanoGPT always go hand by hand.

[François Chollet @fchollet]

LLM adoption among US workers is closing in on 50%. Meanwhile labor productivity growth is lower than in 2020.

Many counter-arguments can be made here, e.g. "they don't know yet how to be productive with it, they've only been using for 1-2 years", "50% is still too low to see impact", "models next year will be unbelievably better", etc.

But I think we now have enough evidence to say that the 2023 talking point that "LLMs will make workers 10x more productive" (some folks even quoted 100x) is probably not accurate.

[will brown @willccbb]

remember when yann lecun said this and everyone was like "shut up dummy" and scared him off this website

[Shangshang Wang @UpupWang]

Sparse autoencoders (SAEs) can be used to elicit strong reasoning abilities with remarkable efficiency.

Using only 1 hour of training at $2 cost without any reasoning traces, we find a way to train 1.5B models via SAEs to score 43.33% Pass@1 on AIME24 and 90% Pass@1 on AMC23.

[Xavier Belda @XavBelda]

@vikhyatk This morning it took me 15s to find something on my notebook that would have taken me 10 on an app.
You can clearly see the great value there!

[Javi M @info_libertas]

🚀 Meet F Lite 7B — Freepik’s new diffusion model distilled from F Lite. 30 % fewer params (7B vs 10B) with similar visual quality and same licensing and SFW properties as the original model. See thread for tech and method details! 👇

[Xavier Belda @XavBelda]

@cuenca @hckinz @freepik Precisely. Sometimes changing the prompt to add this color correction changes the final generation noticeably. This way we can keep the original image with minor changes.

[Xavier Belda @XavBelda]

1/7 Ever noticed GPT‑4o images looking warmer than expected? You’re not alone. At @freepik, we've explored a simple scaling hack to neutralize that yellow cast - no heavy lifting required. ⬇️

[Xavier Belda @XavBelda]

7/7 To learn more, read the full blog post:
https://t.co/bYpAvo0DDA

And if you want to try it out, head up to our Freepik Assistant and start generating:
https://t.co/FebDTzxt3V

Or play around the standalone fix tool:
https://t.co/gJ91JfjMca

Enjoy! 🎨✨

[Xavier Belda @XavBelda]

6/7 Remember those videos passing the same reconstructed image to ChatGPT again and again? Well, notice the yellow tint gone between the original and the scaled version 👭

[Carlos Cabrera Jimenez @Ccabrerajimenez]

Excited to announce the release of a fast, lightweight NSFW image detector! It delivers impressive metrics in evaluation. Check it out on Hugging Face https://t.co/lUw96LmGTn

[Xavier Belda @XavBelda]

@QuentinWach I would have killed to be able to play with something like this when I was doing my Master's Thesis. Awesome!