I never really considered how dangerous QK-norm actually is before working on RBF Attention. While solving some obvious issues, it can be the cause of some much less obvious ones.🧵

Neural networks have a fundamental problem. Feed them garbage data and instead of admitting that they are confused, they will confidently hallucinate. I just open-sourced the HALO-Loss to try and fix this. It give the model a mathematically sound *I don't know!* button.🧵

I dove deeper into the rabbit hole of RBF-Attention. I refined the Triton kernel, added register-tokens and developed SuSiE positional embedding as a replacement for RoPE in Euclidean space. Go have a look at the repo or the blogpost in the comments if you're interested! :)

For some reason I decided to swap out standard dot-product attention for a scaled-rbf kernel. Pretty much expected it to fail to converge or be impossibly slow but the scaled-rbf-attention is getting unexpectedly good results?? 👇

Playing with a new kind of attention. Plots are for the same setup with standard vs. modified attention on one epoch of tiny-stories. Speed is roughly the same as flash-attention. Looking good!🤞

AI isn't coming for your creativity. It's coming for your lack of diligence. People talk a lot about #AGI and "super-intelligence," but the immediate disruption is much simpler: AI is killing "vibe-based" decision-making.

Over the past year Michał Lewandowski and I published a series of papers on Space Folding , and while Michał went to #AAAI to present the latest one, I worked on a blog-post explaining some the central ideas behind the papers. Let me know what you think! www.pisoni.ai/posts/space-...

After a long hiatus I decided to update my blog and write about some of the things I did over the last few years. Come have a look! pisoni.ai

Currently heading to #EurIPS in Copenhagen to present our work on space folding and model interpretability. If you're attending and would like to discuss Representation Learning, SSL, Multimodal LLMs, CV, or other topics that YOU are excited about, feel free to reach out.

The US government should subsidize Open AI rather than OpenAI

On the occasion of the 1000th citation of our Sinkhorn-Knopp self-supervised representation learning paper, I've written a whole post about the history and the key bits of this method that powers the state-of-the-art SSL vision models. Read it here :): docs.google.com/document/d/1...

We're ready!

The single most undervalued property of neural networks is self-consistency. We should change that!

You've been researching for a while! Time to have some SOTA! #aislop

You and Adam keep beating Sota? Stop doing that! Poor Sota!

Have some cool idea but only evaluate it on small models? Tough luck buddy. You only get your paper accepted if your experimental results are 0.2% above SOTA and too expensive to falsify! Is academic publishing pay to win yet?

Is there a reason why none of the recent models use RBF-kernel Attention to get rid of the softmax-bottleneck for long context? I tried replacing dot-product attention with the negative squared KQ-distance and was able to remove the softmax without issues and loss in performance!

NeurIPS is endorsing EurIPS, an independently-organized meeting which will offer researchers an opportunity to additionally present NeurIPS work in Europe concurrently with NeurIPS. Read more in our blog post and on the EurIPS website: blog.neurips.cc/2025/07/16/n... eurips.cc

Has anyone experimented with "conditional gradients"? Thinking about a setup where, within a specific activation range (e.g., right before a ReLU), you'd only permit positive or negative gradients.

Quick question to the SSL experts out there: Usually you evaluate an ssl-model by freezing it and training a linear probing layer. Would it be fair to somehow learn a final layer with more dimensions than classes and do a nearest-neighbor evaluation?

There is an oak forest in central France that was planted 400 years ago by Colbert so that France would have quality hard wood by the 2000s to build ships for its navy. This is the type of long term planning that Seldonian predictions can help improving.

New anti-censorship jailbreak just dropped ;)

Currently on my way to #ICLR in Singapore where we'll present our latest paper on space folding in neural networks. Would be happy to meet some people there so if you're at ICLR as well and want to hang out feel free to pm!🙂

Grok this! What a roller-coaster of emotions...🤪

ModernBERT or DeBERTaV3? What's driving performance: architecture or data? To find out we pretrained ModernBERT on the same dataset as CamemBERTaV2 (a DeBERTaV3 model) to isolate architecture effects. Here are our findings:

Just assembled a slide about local feature training time/dataset size. Anything wrong/missing?

Is the project even still worth doing when wandb runs out of funny names or am I cooked?🫠

Meta introduced Llama 4 models and added this section near the very bottom of the announcement 😬 “[LLMs] historically have leaned left when it comes to debated political and social topics.” ai.meta.com/blog/llama-4...

🚀Hello, world! We are now live on Bluesky. This is the official account of the Department of Computer Science at ETH Zurich. Follow us for cutting-edge research, the latest innovations, event updates and insights into the future of technology. inf.ethz.ch @csateth.bsky.social @ethzurich.bsky.social

Recently had the pleasure of helping @miclew.bsky.social with a couple of his papers in exchange for him helping me with a couple of mine! This is the first fruit of our common work. We quantify space folding in relu neural networks with a range based measure. Lots of fun to write and read!😉

x''= 0

🚀 Paper Release! 🚀 Curious about image retrieval and contrastive learning? We present: 📄 "All You Need to Know About Training Image Retrieval Models" 🔍 The most comprehensive retrieval benchmark—thousands of experiments across 4 datasets, dozens of losses, batch sizes, LRs, data labeling, and more!

anybody else a fan of "three body problem" - remember that part where the aliens attack earth by shutting down our ability to do science? what a crazy, fictional idea, good thing nothing like that could happen in real life.

"no b-but deepseek c-can't tiannamen" Here's Grok for you:

The fact that Deepseek R1 was released three days /before/ Stargate means these guys stood in front of Trump and said they needed half a trillion dollars while they knew R1 was open source and trained for $5M. Beautiful.

Super interesting new CLIP-loss that takes cross-sample similarities into account to learn consistent representations. Also makes pretraining very data efficient. But i think there is a catch...👇

Not that I'm super active on social media recently but I still feel like I need a break... 🫣

Another nail in the coffin of cosine similarity! I started disliking cossim some years ago due to multiple reasons such as the non-linearity around 0.0 and the loss of certainty-information due to the normalization of feature vectors but this study seems to give another good reason to abandon it.

I'll get straight to the point. We trained 2 new models. Like BERT, but modern. ModernBERT. Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff. It's much faster, more accurate, longer context, and more useful. 🧵

She said YES!🥰

O3 is costly. These numbers are for a single ARC benchmark task

Fantastic Muse quote on a fantastic NeurIPS poster! Doesn't get much better than that!😅

HOT 🔥 fastest, most precise, and most capable hand control setup ever... Less than $450 and fully open-source 🤯 by @huggingface, @therobotstudio, @NepYope This tendon-driven technology will disrupt robotics! Retweet to accelerate its democratization 🚀 A thread 🧵

The global rise of anti-intellectualism and anti-science is directly related to the global rise of fascism and right-wing authoritarianism. The defense of truth is inherently anti-fascist.

Stand by while annual NeurIPS FOMO is loading...

I mean no disrespect but the timing of publishing this before moving to OpenAI, who has changed its ethical standpoint quite frequently and is by now openly deploying its tech on the battlefield is a bit unfortunate. I wish all involved parties the best though so let's hope nobody gets burned.🤞

Whaat?

It's so funny that this movie was so inspiring to so many people. I also did a project to detect and decode Arrival glyph codes in university but our solution wasn't quite as revolutionary...😅