. . AI: OpenAI and Figure AI Show Off Their Human-Like Robot (3.13.24)

Friends, in today's top story, OpenAI and Figure share a video of their humanoid robot that moves eerily like an actual human. Will there be one of these in many homes someday? In other news, we see stories evaluating algorithmic progress in AI and the successful development of generalist AI skills by DeepMind through gaming, once again all pointing to faster, efficient and capable models.

These are the stories the AI community is talking about most today. We hope you find the summaries so useful you’re inclined to share them.

Now here’s today’s news…

-Marshall Kirkpatrick, Editor

OpenAI and Figure Share Video of Robot Talking And Showing Human Level Dexterity

First impacted: Robotics Engineers, AI Researchers
Time to impact: Short

OpenAI and Figure's 2 week old collaboration towards building humanoid robots is off to a fast start. Figure released an impressive video demonstrating the capabilities and the robot appears to have human-like dexterity, with OpenAI contributing visual reasoning and language understanding capabilities, and Figure's neural networks enabling fast and precise robot movement. The company says this partnership has significantly increased the robot's capabilities, and that Figure is aiming to operate humanoid robots on a billion-unit scale. The demo video is likely to evoke some emotions. [via @adcock_brett] Explore more of our coverage of: OpenAI, Humanoid Robots, Neural Networks. Share this story by email

DeepMind's SIMA Masters Multiple Video Games and Shows Potential for Generalization

First impacted: Video game developers, AI researchers
Time to impact: Medium

Google DeepMind has introduced SIMA (Scalable Instructable Multiworld Agent), an AI that understands instructions in everyday language across various video game environments. It relies solely on screen visuals and user directives. Trained on nine 3D games, SIMA outperformed agents specialized in single games, showcasing its adaptability beyond its initial training. However, further research is essential for SIMA to match human performance in both previously trained and new games. This development hints at the potential for creating generalist agents capable of learning in data-absent environments, a domain previously dominated by humans. [Introducing SIMA, a Scalable Instructable Multiworld Agent] Explore more of our coverage of: Google DeepMind, AI Gaming, Natural Language Processing. Share this story by email

Language Model Compute Halving Every Eight Months and Outpaces Moore's Law

First impacted: AI researchers, Data scientists
Time to impact: Medium

Researchers have found that the computational power required to pre-train language models is halving approximately every eight months, a rate faster than Moore's Law, according to a study that analyzed over 200 language model evaluations from 2012 to 2023. [Algorithmic progress in language models] Explore more of our coverage of: Language Models, Computational Power, Algorithmic Progress. Share this story by email

Physical Intelligence Raises $70 Million for AI Robotics and to Develop Universal Model

First impacted: Robotics Engineers, AI Researchers
Time to impact: Medium

Physical Intelligence, a startup co-founded by Karol Hausman, formerly of Google, and Stanford University's Chelsea Finn, has secured $70 million in funding to develop a universal AI model for robots. The company plans to purchase a variety of robots to train their AI models on, with the goal of creating the largest robotics data collection to date, positioning them as competitors to companies like Figure AI and Tesla Inc. Hausman: "We aim to bring AI to the physical world with a universal model that can power any robot or any physical device basically for any application.” [Physical Intelligence Is Building a Brain for Robots] Explore more of our coverage of: AI Robotics, Startup Funding, Universal AI Model. Share this story by email

Open-Source Project Aims to Replicate Autonomous Agent Devin

First impacted: Software Engineers, AI Developers
Time to impact: Medium

Yesterday we shared a story on Devin, an AI agent designed for software engineering tasks and built by Cognition AI. Today, an open source project is kicking off to replicate Devin and it is called OpenDevin. The project, which operates under the motto "Code Less, Make More," is being led by Junyang Lin, a member of the Alibaba AI team. [GitHub - OpenDevin/OpenDevin] Explore more of our coverage of: AI Software Engineering, Language Learning Models, Coding Efficiency. Share this story by email

That’s it! More AI news tomorrow!