🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Drew Linsley nice! relatedly, we find that humans beat deep networks at recognising objects in unusual poses, although our latest unpublished analyses seem to indicate that Gemini reaches human performance (but not GPT4 nor Claude3), possibly due to data contamination arxiv.org/html/2402.0397…

Super excited to announce our cool project, Trace, on optimizing general AI systems, using LLMs.😎 Trace is a new AutoDiff-like tool for training AI systems end-to-end with general feedback (like numerical rewards, natural language text, compiler errors). microsoft.github.io/Trace/

Super excited to share our recent work on privileged representational axes in biological and artificial neural networks! w/ Nancy Kanwisher @[email protected] Josh McDermott Alex Williams biorxiv.org/content/10.110…

Excited to share what I've been up to since joining Google DeepMind this year! We developed "JEST" a method for jointly selecting the best *batches* of data to train on. We use JEST to accelerate multimodal pretraining, surpassing previous SoTA with just 10% of the FLOP budget!

I’m thrilled to share our paper published online today in Science Magazine ! I appreciate my mentor, Dr.Fujisawa, for running through this project with me. We discovered the new type of grid cells that have grid field for future projected locations. science.org/doi/10.1126/sc…

François Fleuret There's three parts. 1. Fitting as large of a network and as large of a batch-size as possible onto the 10k/100k/1m H100s -- parallelizing and using memory-saving tricks. 2. Communicating state between these GPUs as quickly as possible 3. Recovering from failures (hardware,

Here it is! A huge milestone in neuroscience – the complete connectome of the fly brain from the FlyWire team (FlyWire). In this special issue, we have 9 research articles. I’ll try to summarize them in a couple threads starting here 🧵

What can you do a map of all connectivity in the brain? Using very simple biophysical modeling, we “ran” the fruit fly connectome. For any given set of input neurons, we can predict what neurons will respond.

Stop worrying about polls and do something Early voting is open in most states, and Kamala Harris needs YOU to talk to the voters who will decide this election. This weekend, #TeamKamala is hosting a Dial-a-Thon to call 5 million voters! share.kamalaharris.com/s/5Safydmd

🎭Recent work shows that models’ inductive biases for 'simpler' features may lead to shortcut learning. What do 'simple' vs 'complex' features look like? What roles do they play in generalization? Our new paper explores these questions. arxiv.org/pdf/2407.06076 #Neurips2024

Thanks Carney Institute for Brain Science for running this story on our work with Drew Linsley Thomas Fel Sohini Ramachandran and twitter-less Sahar Shahamatdar.

Now in Nature! 🚀 Our method learns a controllable 3D model of any robot from vision, enabling single-camera closed-loop control at test time! This includes robots previously uncontrollable, soft, and bio-inspired, potentially lowering the barrier of entry to automation! Paper:

Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵