Lessons learned debugging ML models:

How can it be that Germany was so naive for so long? it's disgusting. We need to stop finance Putin's war. Immediately. #StopRussia #SWIFTban

Temporal Difference Learning for Model Predictive Control nicklashansen.github.io/td-mpc/

Robots in flow state.

An amazing week at the International Computer Vision Summer School comes to an end. A week full of fantastic talks, interesting discussions, and a lot of fun. Thanks to everyone, and especially to the organisers ICVSS, for making this such a memorable experience!

Wayve presents Model-based Imitation LEarning (MILE): a method to jointly learn a world model and a driving policy from an offline corpus of driving data. This model is another step in achieving a more generalisable driver - paper to appear at #NeurIPS2022 wayve.ai/blog/learning-…

Are you an autonomous driving or AI researcher? Wayve is organising this year's "Scalable Autonomous Driving" workshop at ICRA. Paper submission is open now until 20th of March. x.com/SinavskiOleg/s…

Another great example showing that robotics is mostly a software problem! The "Egg Peeling Challenge" could be considered the Turing Test for embodied AI.

LLaMA has been fine-tuned by stanford, "We performed a blind pairwise comparison between text-davinci-003 and Alpaca 7B, and we found that these two models have very similar performance: Alpaca wins 90 versus 89 comparisons against text-davinci-003."

I’m a car guy who enjoys the meditative (and fun) aspects of driving–but I’m also eagerly anticipating the day when I can hand over the keys to a machine. gatesnot.es/3Km3MBc

Super excited about the progress we've made with world models!

With LINGO-1 we have a built an AI system that explains its own actions in natural language. One step closer towards embodied AI that reasons about the world like humans 💬

I’m in Paris this week for ICCV. Drop me a message if you want to meetup!

New paper demonstrating large language models with 1bit weights (+/-) scale just as well as FP16. This potentially opens up 100B+ params open-source models running on consumer hardware arxiv.org/abs/2310.11453

Backdoor in liblzma through a mix of social engineering and clever obfuscation. It modifies the RSA decrypt function on a linux system, potentially allowing unauthorised access over ssh! openwall.com/lists/oss-secu…

End to end AI is the way towards generally capable robots. Great progress on robot dexterity for real world applications!

LINGO-2 is the first vision-language-action model that can drive a car on public roads. Excited for the future work that this opens up!

The ultimate application of AI will be sorting your shopping list by supermarket layout.

It blows my mind that a 8B model is able to outperform GPT4o/Sonnet here.

Excited to be at NeurIPS Conference in San Diego next week! If you want to chat large scale RL, agents, synthetic data, reward hacking, or want to know more about what we’re up to poolside, I’d love to connect/grab a coffee. Our booth is #913, come say hi! ⛱️