NeuroReview for the Week (#73)
Since I am currently at CVPR 2025, I am a bit delayed with this issue.
CVPR 2025
– Announcing — it’s the biggest exhibition in my favorite area — computer vision.
– And what exactly is CVPR and why does everyone get so excited about it? I’ll explain a little for those who might not be in the loop.
– One of the authors of FlashAttention is a legend in the industry, sharing fresh ideas about more efficient approaches to attention mechanisms. You could see him live during the presentation!
– The vibes from the oral presentations were very energetic.
– My intern took the stage and presented our paper — well done! A real professional, no doubt.
LLM
– Mistral unveiled their first conceptual project — Magistral. Unfortunately, it’s not very successful so far, but there’s hope it will be improved through methods like SFT.
– They released o3 pro — traditionally available via subscription for professionals. Plus, it’s several times cheaper via API compared to the previous o1 pro version.
– A new language model was launched by the team at Avito — they stayed sharp and created a fast, Russian-language LLM.
– Additionally, the limits for o3 have been doubled, so now they can be used to their fullest.
Other News
– At the Apple WWDC, a developer presentation took place. Now it’s possible to integrate Apple Intelligence directly into applications.
– AMD prepared its response to Nvidia — introduced new solutions with great price-performance ratio and a strong focus on inference optimization.
And if you’re interested, you can check out issue #72.