4 Friday Stories in AI

Robots, catalysts, nuclear, and decentralized AI

Happy Friday, everyone. Today we’re highlighting just 4 stories that rose to the top of our weighted analysis of AI community engagement, but as a collection, they really paint striking picture.

This is all much bigger than just ChatGPT, or any other single AI technology. These four stories point to much, much more.

We’ve posted these as a Twitter thread here, if you want to help get these out to a larger audience. I’d be thankful.

Thanks for your time,

Marshall Kirkpatrick

DeepMind's Robotic Transformer 2 Levels Up Robotic Control
The new Robotic Transformer 2 (RT-2) language-vision-action model combines computer vision, web scale language capabilities, and control of physical robotics to enable robots to develop general and emergent skills, interpret new commands, and performing basic reasoning. For example, robots trained with the model decided to use a rock as an improvised hammer and to give a tired person an energy drink. In over 6,000 trials, RT-2 surpassed previous models, tripling generalisation performance and managing tasks with previously unseen objects or scenarios. Emergent skills and basic reasoning with previously unseen objects? Using a rock as a hammer and giving people energy drinks? Whoa. [DeepMind: RT-2: New model translates vision and language into action]

Meta Says Its AI Can Accelerate Materials Science Catalyst Reactivity Experiments by 1000X
The Open Catalyst Project (OCP), a collaboration between Meta AI's FAIR and Carnegie Mellon University, has introduced a demo that identifies ideal binding sites for adsorbate molecules on catalyst surfaces. Supporting over 11,000 catalyst materials and 86 adsorbates, the demo enables the study of around 100 million surface-adsorbate combinations. Although the models have limitations compared to Density Functional Theory simulations and physical experiments, the OCP demo is set to speed up catalyst discovery via machine learning, simulating catalyst reactivity about 1000 times faster than current methods. [Meta AI: Accelerating catalyst discovery with AI]

Andreessen and Horowitz to Discuss Nuclear Tech, AI in Podcast
Marc Andreessen and Ben Horowitz are planning a podcast on Oppenheimer, nuclear tech, Communism, and AI. They're inviting their Twitter followers to submit questions. No one has asked yet about robots with emergent skills with previously unseen objects. [via @pmarca]

5 Questions of Timnit Gebru, Decentralized AI Advocate
"Timnit Gebru, PhD is the founder and executive director of the Distributed Artificial Intelligence Research I-nstitute (DAIR). DAIR is an interdisciplinary and globally distributed research institute exploring the harms and potential benefits of AI. What’s Next Health spoke with Dr. Gebru about her organization, which is supported in part by the Robert Wood Johnson Foundation (RWJF)." Gebru’s interview concludes with these words: “To me, our work boils down to two questions: First, ‘What are the issues people are facing with AI?’ and second ‘What kind of AI systems can we create or how do we use our expertise in this space to actually build a healthier technological future’” [What's Next Health: 5 Questions For Timnit Gebru]

That's it! Have a good weekend.