Erik Brynjolfsson Too often, we think a task is easy because some animal can do it. But the reality is that the task is fiendishly complex and the animal is much smarter than we think. Conversely, we think tasks like playing chess, calculating an integral, or producing grammatically correct text

KV caching in LLMs, clearly explained (with visuals):

Magnus Carlsen claims that one or two signals per match from a chess AI indicating when he should think hardest would make him “almost invincible”. IMO you could boost researchers similarly with just one or two signals a year saying “think hard about the paper you just read”.

Looks amazing! 😍

Hello! If you like pretty images and videos and want a rec for CVPR oral session, you should def go to Image/Video Gen, Friday at 9am: I'll be presenting "Motion Prompting" Ryan Burgert will be presenting "Go with the Flow" and Pascal CHANG will be presenting "LookingGlass"

DINOv2 meets text at #CVPR 2025! Why choose between high-quality DINO features and CLIP-style vision-language alignment? Pick both with dino.txt 🦖📖 We align frozen DINOv2 features with text captions, obtaining both image-level and patch-level alignment at a minimal cost. [1/N]

Check out our AI Research Lab - Explained episode on Multimodal AI. Had a blast creating this episode with the team! Salesforce AI Research

NVIDIA wants to sell you NVL72 rack ($3M) so you can do real-time video generation 😅 Good thing: you don't need it. Self Forcing does the job with one 4090, and with better quality 😊 self-forcing.github.io

Wow, checks out! 🤯

Wow

Introducing Shortcut — the first superhuman Excel agent. Shortcut one-shots most knowledge work tasks on Excel. It even scores >80% on Excel World Championship Cases in ~10 minutes. That's 10x faster than humans. Our early preview is live. Just comment for an invite code.

Oh wow, did you guys know that torch.compile can compile numpy code? And even run it on GPU? This is pretty neat for all kinds of "surrounding" code besides the model (like evals and fancy metrics) that I used to do with numba/numexpr (cuz CPU-XLA was pretty meh). Poll below

reciprocal reviewing is a terrible idea that unfortunately more conferences are adopting, as if we didn't already have enough problems with the review quality from people who are willing to review.

A beautiful paper that goes through Diffusion step by step, explaining the entire math of it from the beginning.

🚨 Did you know that small-batch vanilla SGD without momentum (i.e. the first optimizer you learn about in intro ML) is virtually as fast as AdamW for LLM pretraining on a per-FLOP basis? 📜 1/n

Just like tokenization is a necessary evil in LLMs (at least for now), time discretization is a necessary evil in robotics/RL. I think there must be a better way to handle continuous time than via naive discretization...

Transformers haven't changed much since 2017, but there have been some innovations over the years. This is an excellent summary of architectural differences in recent LLMs. Nice diagrams too! 👏 It would be great to see something like this for diffusion Transformers as well 🤔

🌟 Happy National Intern Day! Today we celebrate the brilliant minds and diverse perspectives that our interns bring to Salesforce AI Research. Our interns contribute to groundbreaking AI research from day one, bringing fresh ideas that drive innovation and solve complex problems for

Introducing Genie 3, our state-of-the-art world model that generates interactive worlds from text, enabling real-time interaction at 24 fps with minutes-long consistency at 720p. 🧵👇