HUGE shoutout to our #CVPR2024 Outstanding Reviewers 🫡

It's very easy to make things seem like they work/don't work at smaller scales. Think about how many efficient attention papers were published in the last 5 years. Do we use any of them?

Training at a larger scale has fundamentally different training dynamics/properties than the

I probably have some credibility as a person who has worked on TensorFlow and PyTorch both (also, caffe / ONNX / distbelief / a few others that never saw the light), so here are my two cents:

(1) Speed doesn't really matter today as long as it is not particularly bad.

WOW.

Spotted earlier from our I-90 Bridge traffic camera. We have no words.

Just wow.

#NoFilter #Bellevue #Seattle

Welcome David Merrill as an advisor to Scaled Foundations! David is the CEO and co-founder of Elroy Air where he is building large autonomous delivery aircraft systems, and shares our mission to build safe AI-powered robots.

A fellow MIT alum, David is also a TED speaker and an avid

If you’re at #GTC2024 this week, come hear Scaled Foundations’ CTO Sai Vemprala discuss 'The Impact of Generative AI on Robotics'.

Look forward to seeing you there!

Wednesday, 3/20 at 2:00 pm
SJCC 230B (L2)

nvidia.com/gtc/session-ca…

#ai #robotics #generativeai

Inspiring!

phys.org/news/2024-03-i…

🗣Exploring beyond interpreting actions that directly involve the camera wearer, how can egocentric audio-visual signals aid in understanding the natural social behaviors among all partners, ultimately enhancing our daily communication? #CVPR24
Project: vjwq.github.io/AV-CONV/

For folks getting excited about 'Next Token Prediction' in robotics, I present 2-year-old work on Perception-Action Causal Transformer (PACT). 🧵

To be clear, we are very happy that OpenAI adopted it & now even more people will continue to innovate on it. Products should use academic research!

However, from the pov of a grad student, it would have meant so much to us if there was attribution.

Thanks for all the love 🩷

🤯WOW🪆Matryoshka Representation Learning enables 'native support for shortening embs' &'very flexible usage'

Jokes aside, excited that OpenAI serves MRL by default in v3 embedding API for retrieval & RAG!

Other models & services should catch-up soon😄

arxiv.org/abs/2205.13147

A Seattle icon and role model. I will always remember his infectious energy and leadership at a young age of 70. Thanks for the memories!

Aditya is one of the best young researchers and an incredible mentor. He has been leading high impact work in representation learning for a while. You should absolutely hire him!

World meet #emuvideo
For the past year, our team has been pushing on video generation. The result? Emu Video that generates high quality videos from text or images. SOTA performance vs. commercial products and academic papers. Check it out emu-video.metademolab.com

keveman Lucas Beyer (bl16) This is roughly right. Basically wanted to send fewer bytes over the network for our distributed neural network training system, and easiest way on a CPU was to lop off the low 16 bits of mantissa, and fill with 0s on other side. Turns out it was fine for training.

Aditya Kusupati is on the job market this year. If you are looking for a super smart, driven, and passionate researcher+leader, you should definitely contact him!
cc: Sham Kakade, UW RAIVN Lab, Inderjit Dhillon, Manish Gupta Jeff Dean (@🏡)

This is one of my favorite works in a long time. A simple, elegant, and extremely impactful idea that you can use in as little as 5-6 lines of code in existing Transformer architectures. Well done Aditya Kusupati, Sneha Kudugunta and co!

Announcing MatFormer - a nested🪆(Matryoshka) Transformer that offers elasticity across deployment constraints.

MatFormer is an architecture that lets us use 100s of accurate smaller models that we never actually trained for!

arxiv.org/abs/2310.07707 1/9

Presenting AirGen, the next-gen aerial robotics simulator and evolution of #AirSim ! AirGen adds several new features to synthesize data for aerial intelligence. Currently available in preview and free/unrestricted for academic purposes. Key features include: 🧵👇 (1/5)