Hi! 👋 Welcome to The Big Y!
GenAI continues to steal the spotlight, but while we wait for GPT-5 I want to take a moment and highlight something cool happening in other AI.
Google DeepMind released new research around generalist AI agents executing tasks in 3D virtual environments (aka video games). The Scalable Instructable Multiworld Agent (SIMA) was trained on nine different video games learning how natural language instructions are connected to in-game behavior. Through a language interface, the agent was able to complete simple tasks in game.
Imitation learning for agents can be extremely beneficial as AI agents learn from humans and can generalize from recurring patterns and behaviors exhibited in games. They learn how to handle generic, recognizable patterns that exist across different games and virtual environments. The researchers also found that SIMA agents trained across all nine games outperformed an agent trained on only one game, even in the game the specialized agent was trained on.
It’ll be interesting to watch as AI agents like this become more generalizable and better at performing multi-step tasks. Many AI systems struggle with transfer learning, bringing knowledge gained in one training scenario to another environment, but this generalized agent is a big step in that direction. This is something that will be helpful for robot training and utilization.
To bring this back to LLMs… Large language models started taking off when they could learn from the contents of the internet, what if they could also learn from human users in the same way this agent system does?
Following up from the main topic last week… The US is facing a major power crunch and data centers are just one piece of the puzzle. I recommend this read from the NYT on what is going on in the US regarding the power grid and the impact of 24/7 data centers.
Know someone who might enjoy this newsletter? Share it with them and help spread the word!
Thanks for reading! Have a great week! 😁
🎙 The Big Y Podcast: Listen on Spotify, Apple Podcasts, Stitcher, Substack