The validation submission deadline is Aug 31, 2025, and the paper deadline is very soon (if you'd like to publish a paper) - on Jul 22, 2025. See all dates at

For all those interested in image/video compression, a gentle reminder that #CLIC2025 is ongoing. You can participate at We're using human evaluations and Elo scores to determine the best methods!

RT @adityakusupati: Pocket powerhouse admist I/O awesomeness!. Gemma 3n E4B & E2B are insane models, optimized for on-device while rivaling….

I'll post more details as we update the website, but it's very exciting that this is happening!.

🚀 The 2025 Challenge on Learned Image Compression is live!. Join us at the Picture Coding Symposium in Aachen for #CLIC2025 — where humans, not just metrics, judge image/video compression quality. 🏆 Leaderboards open June 1. Let the coding begin!. Web:

🌌 New solar-system demo live! Fully procedural planets, sun & star-field—now gleaming with silky ray-traced soft shadows. ☄️ Comets obey gravity while planets & moons cruise on rails for this showcase. Powered by Gemini Pro 2.5. Hit play and enjoy!✨.

For now, the most productive use I've found is delegating background coding chores to an AI agent while I tackle the main logic. Beats me doing the grunt work, and avoids the frustration of micromanaging it on complex, attention-hungry tasks.

The current state of agentic AI coding is wild: sometimes it's a genius, conjuring solutions I'd never conceive. Other times? Utterly baffled by minor details. Context is key. 😅 The potential is HUGE, but the journey is. interesting.

You'll need a desktop browser most likely because I also asked Gemini to use WebGPU to compute the physics. If you like this demo, I have much more advanced versions in the works (where I spent more time to fine tune things).

I love astronomy, physics simulations, and procedural generation. Gemini 2.5 Pro let me combine all three into a quick demo I developed over a couple of hours:. (fixed link)

I went all-in on coding with AI for a few months. Here's what I learned - the good, the bad, and the unexpected. Enjoy this snapshot of our AI moment.

I am fairly sure that the "infinite loop" where the orchestrator got stuck cost a good $20 out of the total sum, and it was something trivial. In any case, this is the worst LLMs will ever be.

Primarily doing this for fun (not as part of my job), the main concern is of course the cost. The entire app ended up costing around $100 (it's an estimate) and it eneded up having around 6k lines of javascript, most of which were tests. Continued⬇️.

Overall, I am quite impressed with what can be done semi-automatically. It would have been even nicer if the code had used selenium to fully not need me at all, but at the same time I am scared about what may happen if we leave LLMs unsupervised. Continued⬇️.

It wasn't 100% smooth sailing. The orchestrator did wonderful things but at some point it was stuck in a rather costly loop. I had to explain the error and then it continued. Another time there was some confusion on the deployment environment vs dev. Same deal. Continued⬇️.

I finally gave this vibe coding a go. Using Roo Code + Gemini Pro 2.5 (preview) last night I spent a couple of hours "coding" a web app. The end result not only worked, but it also came with ~100 unit tests, UI test, and integration tests + depolyment script. Continued⬇️.

RT @learn_to_cmpres: Less than 24 hours left to submit to the Machine Learning and Compression Workshop 🗜️🗜️ at #NeurIPS2024!. 👉Workshop we….

RT @alirezafathi: Our team at Google DeepMind is seeking a Research Scientist with a strong publication record (multiple first-author paper….

RT @Official_CLIC: We’re happy to announce that CLIC 2025 will be co-located with Picture Coding Symposium in December 2025. More details t….

RT @JosefDeanGC: Sure matplotlib is cool, but what if I want to load my loss curves into the 2006 hit Flash game LineRider? .