Guatemalan 馃嚞馃嚬 physicist, Postdoc @CVC_UAB, researching autonomous driving.
Barcelona
Joined October 2011
- Tweets6.9K
- Following366
- Followers1.1K
- Likes41.4K
Pinned Tweet
Fun weekend project: explore the latent space with your hand! Using SD-XL turbo and some pre-defined prompts, this is a proof of concept, but we plan to do so much more. Stay tuned :)
Diego Porres retweeted
Ok my turn to be brave for the good of the community
Not sure who needs to hear it, but usually models with JEPA in their title are bad. Like LeJepa is much worse than DINO in practice
Pretty papers, fancy math, bad results
V-JEPA 2.1 is probably the only exception [1/2]
Can we please have more posts like this one?
Too many people are afraid to openly criticize posts and papers, and that really hurts the community and lets bad ideas spread fast
Thank you @giffmana for saving me (and thousands other people) an hour of my time
been working on my own data cleaning and benchmark viz tools and it has made my life so much easier, ty claude
Diego Porres retweeted
Claude struggled to write my kind of music (theory-heavy music for nerds lol) until Opus 5.5. I asked for a motif passed from one player to the next and reinterpreted, with a solo from each instrument demonstrating specific techniques, and it made this piece, Ouroboros. Then I had it animate the video using tools I've built over the past year while reverse-engineering music to study theory.
people are going to discover via first principles that predicting future pixels doesn't help for the downstream task of actually driving, driving me in turn, insane
Diego Porres retweeted
These are all SVG!
Something really cool is coming to @QuiverAI around icons, asset families, and styles.
Reply for early access.
> 1% of context left
> claude code: I'veidentified two more simplifications to the code...
> me: fml
Diego Porres retweeted
馃獥 Your Transformer Is Secretly Elastic 馃У馃憞
Transformer families come in a few fixed sizes, so you commit to one for your hardware and latency constraints, even as query loads vary. What if a single pre-trained model could serve every budget?
We learn one importance ranking over every MLP neuron and attention head, once, with frozen weights. At inference time, the model can then be pruned to any computational budget by dropping its least important units, without retraining.
Across 27 vision and text encoders spanning 22M to 8B parameters, performance degrades smoothly up to 60% sparsity. For example, pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters, for only a 0.8% drop in ImageNet-1k linear probing.
Super fun project in collaboration with Dheeraj Varghese (@dhevarghese), supervised by Gertjan J. Burghouts (@gjburghouts) and Cees Snoek (@cgmsnoek)
馃寪 Project page: secretly-elastic.ashita.nl
predicting future latents/views in autonomous driving makes no sense, you are in an ever-changing environment in contrast to robotics where you hardly move at all (at this stage, i imagine they will realize these things as they want to move between rooms or walk outside)
turns out it means there is a catastrophic error and no driver works on the server so we have to reinstall everything, fml
"the center of the latent space"
the trippiest ending, finding another version of yourself waiting to accompany you; this can only happen in real life if the universe is *infinite* and we go 10^10^128 m away
"At the edge of a generative cultural precipice" (also the title of my CVPR workshop paper)
I like how we can tell the training data has a lot of recordings where the cameraman can be seen, as after all, it is another actor in the scene
prompt: " " led me to a never-ending blanket folding hell, loved it
all the actions I could make, besides jumping, were folding those blankets for some reason lol