PaschalIfeanyi🕵️

🚨 A new MIT paper investigates "machine play." What happens when AI agents are left to explore on their own? Here's why it matters from an AI governance perspective: A new paper, written by MIT researchers @nacloos, @noranta4, @jacobandreas, @phillip_isola, Daniel Durbin, and Daniela Rus, investigates what happens when no specific instructions are given to AI agents. They placed a coding agent powered by Claude on a small simulated island and let it explore freely. Over 30 hours, the agent climbed hills, stacked blocks, drew mandalas, and tried sports like bowling. They associated this type of setting with human play: “We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress. Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows. Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.” In the paper, the authors show that: - an embodied AI agent architecture they named ‘Eko’ is able to spontaneously generate its own activities in an unknown world; - these activities exhibit behavioral signatures of play; - engaging in them triggers a form of learning that expands the agent’s later capabilities. On the project's website, you can see replays of what happened over these 30 hours, read the agent's quotes, explore specific behaviors, and learn more about the experiment. I found this study particularly interesting for AI governance research and practice, even though this was not the paper's main goal. As I shared a few days ago, frontier AI labs are currently focused on automating AI R&D, which, in practice, means that AI will soon conduct its own experiments and develop other AIs. Some say that this will lead to an “explosion of intelligence.” It remains unclear whether these “agentic researchers” will align with human values and goals. Moreover, recent agentic safety incidents, particularly the OpenAI/Hugging Face hack, have made it clear that AI labs have overestimated how much they know about agentic AI behavior and safety, and it is still unclear if they can guarantee that agents will remain under human control. The two paragraphs above are grim reminders of the urgent need for AI regulation and governance, and the risk of severe, multi-casualty events if we do nothing. They also remind us of the importance of continuous research on agentic AI behavior, including experimental studies like the one above. 👉 This is a very interesting paper with great graphics and carefully organized data; don't miss it (link below). 👉 Next week, at our 11th AI Ethics & Governance Workshop, we will discuss emerging AI security threats and practical measures to address them. Join us! 👉 To learn more about AI ethics, safety & governance, join my newsletter's 102,500+ subscribers below.