Salma Mayorquin PRO

salma-remyx

27 22 483

https://remyx.ai

smellslikeml

AI & ML interests

None yet

Recent Activity

commentedon a paper about 18 hours ago

TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment

liked a model about 20 hours ago

Qwen/Qwen3-VL-4B-Thinking

reacted to wop's post with 🔥 3 days ago

# PixelModel v1 Last month we released PixelModel — a neural network whose weights are literally the pixels of a PNG. It was a toy: 202,752 parameters, welded to 32×32 output, trained on six solid-color swatches. It scored FID 566.84 on the Tiny-T2I-Leaderboard, mostly by producing the same yellow noise for every prompt. Today we're releasing PixelModel v1. It is 8.5× smaller — 23,747 parameters — and it beats v0 on both benchmark metrics while being trained on 20,000 real MS-COCO caption/image pairs instead of six color swatches. The entire model now fits in a 160×149 PNG. That image is not a visualization of the model. It is the model. All 23,747 weights, one per pixel. ## links https://huggingface.co/bench-labs Blog post [read it here ⇗](https://huggingface.co/spaces/bench-labs/blog?post=pixelmodel-v1.html) See us on the [Leaderboard ⇗](https://huggingface.co/spaces/FlameF0X/Tiny-T2I-Leaderboard) Model card [here](https://huggingface.co/bench-labs/pixelmodel-v1) ## The catch A 23K-parameter model does not draw sandwiches. With ~1 parameter per training image, the loss-minimizing behavior is to output the average of all plausible images for a caption — caption-conditioned color, light, and layout statistics. Food prompts come out warm and brown; sky prompts come out cool and bright. That is the ceiling for this size class, and we'd rather show it than crop around it. # cherry on top 🍒 The model generates 600 images (cpu) in 5 (five) seconds. Thats 5000 images in 24 seconds on cpu. The model trained on cpu for just 30 minutes.

View all activity

Organizations

Posts 35

Post

It's like Christmas in July for frontier coding models.

There's a lot of talk about AI costs, but I've managed to increase the rate of code production by orders of magnitude while reducing the average cost per feature branch over the past month.

Models like GLM-5.2 and Kimi-k3 have been a big part of that, but so has better automation.

I don't keep confirming prompts or redirecting my agent to check out the CONTRIBUTING.md to apply repo conventions, that's baked in. Boris Cherny was just posting about how to build the abstractions to push the limits to your attention in AI coding: https://x.com/bcherny/status/2077929390806073807

If that agent isn't proactive, it's just a chatbot!

I already know what repos I'm building on today, I know I'm going to check out what's new in the arXiv, look for code and licenses, smoke-test the repo, adapt the method to my own application, apply my own org's conventions to shape up a PR.

Now I orchestrate agents to work this out, when an idea makes it through the gates, I'm notified about a PR to review.

And I'm seeing more throughput by scheduling AI coding runs in Github Actions, get outrider: https://github.com/remyxai/outrider

It's a bit like moneyballing vibecoding.

We're using AI to filter the firehose of methods published to the arXiv, matching against your repo and development history, to meet the challenges to improving your application.

Way beyond AI chat with arXiv, the coding agents ground the recommendations and implementations on actual code, using verification and testing to gate which proposals get ramped up to PRs for review.

Check out this Outrider contribution, making its way upstream into the libraries we all use: https://github.com/huggingface/peft/pull/3382

Post

4043

It's conference season, so you'll find an uptick in chatter around the research reproducibility crisis. Consider this a PSA on where the real challenges in working with research actually live.

After all, AI has made it way easier to release code and model artifacts alongside the preprints. And how many times do you really need to replicate the authors' exact configuration?

Downstream of that, as engineers evaluate candidate methods for improving THEIR systems, they rarely find a drop-in solution. More often, they're making tough tradeoffs in fidelity to the documented technique and the constraints of their deployment scenario.

They're swapping models or data indexing strategies. They have their own benchmarks to measure changes against. They're making principled reductions of a technique to respect some resource limit not considered in the source paper.

AI coding has made replication cheap when a paper provides starting point for your own experiments. But the work of adoption requires validation grounded in real-world outcomes.

So put these techniques to the test in your own system, and you'll understand a method's impact well before the survey paper drops in six months.

At Remyx AI, we're helping teams discover, implement, and validate what's next for their systems.

Get Outrider: https://github.com/remyxai/outrider