I vibe coded my personal website over the weekend. One of the best co-pilot AI experiences I had so far. Its a simple static website but it feels so nice to be able to build something in a language you have 0 exeperience with while being able to control the fine details

This is so cool ✨

I am working on a new app that scrapes ML papers from different sources and lets me easily sweep through and curate them. I will share more details along the way!

Our MIT class “6.S184: Introduction to Flow Matching and Diffusion Models” is now available on YouTube! We teach state-of-the-art generative AI algorithms for images, videos, proteins, etc. together with the mathematical tools to understand them. diffusion.csail.mit.edu (1/4)

I just watched the two recent videos from Andrej Karpathy and they are truly great. I really like how he dives into the LLM's "psychology", showing very well how brilliant and dumb these models are at the same time. Would highly recommend to watch them. youtube.com/watch?v=7xTGNN…

Finally finished porting mast3r-slam to Rerun and adding a Gradio interface. Really cool to see it running on any video I throw at it, I've included the code at the end

As one of the people who popularized the field of diffusion models, I am excited to share something that might be the “beginning of the end” of it. IMM has a single stable training stage, a single objective, and a single network — all are what make diffusion so popular today.

More progress on developing a straightforward method to collect first-person (ego) and third-person (exo) data for robotic training with Rerun . I’ve been using the HO-cap dataset to establish a baseline, and here are some updates I’ve made: * added in MANO parameters from

I just got a VisionPro and man, watching movies in there is such an incredible experience! Also the eye and hand tracking is just phenomenal.

This is the most sci-fi like news I have heard in a long time. This is soo cool!

Pretty smart, you can achieve 30% LOSSLESS compression on most LLMs by entropy-coding the exponent component of their BFloat16 weight values. I like that they also implemented custom GPU kernels and showed actual throughput gains with this representation arxiv.org/pdf/2504.11651

This was such a good and comprehensive read about all the progress in diffusion models over the past years. Sander Dieleman explains really well the rationale behind the different approaches, why they made/make sense and how they came about. sander.ai/2025/04/15/lat…

Found this approach for compressing LLMs also pretty clever. By approximating the weights of trained models with linear combinations of random matrices, you only need to store the seed numbers that generated the random matrices and the coefficients. arxiv.org/pdf/2410.10714