Elon Musk pokes fun at empowering AI agents

“People give OpenClaw ultimate access to their entire lives,” Musk wrote on X on February 24, posting an image of a monkey being given a rifle.

This image was originally created by user Devour used in response to a post by Summer Yue, Meta’s Director of AI Safety. Yue ran into trouble while trying out OpenClaw, an open source AI agent that can work 24/7 on behalf of users. Yue allows OpenClaw to check for overloaded mailboxes and suggest emails to delete or archive, but requires the agent to confirm before taking action. However, it became out of control and immediately deleted a series of emails, ignoring her order to stop on the phone.

“I couldn’t stop it from my phone. I had to run to my Mac mini computer like I was defusing a bomb,” Yue told X on February 24.

Peter Steinberger, the “father” of OpenClaw, responded to Yue’s post saying that simply entering the “stop” command would solve everything. Meanwhile, Musk quipped at Yue: “Someone who was defeated by OpenClaw will surely be able to handle the issue of AI safety.”

Yue said she tested her virtual assistant with a smaller “toy” mailbox. It worked well with less important emails and built trust, leading her to switch to applying it to her real mailbox.

Yue assessed that the sheer amount of data in the real mailbox triggered the compression process, which occurs when the context window – a continuous record of everything the AI ​​records and does during a session – becomes too large, causing the agent to summarize, compress and re-manage the conversation. The AI ​​can then ignore instructions that the user considers important, such as Yue’s “no action” request.

OpenClaw and Moltbook, social networks just for AI agents, have been attracting major attention in recent weeks as programmers have shared videos of AIs themselves posting, debating and networking.

Last week, Sam Altman, CEO of OpenAI, announced that he had recruited the “father” of OpenClaw to “lead the next generation of personal assistants”. Steinberger and Altman affirmed that OpenClaw will continue to exist as an open source project and be supported by OpenAI. “The future is multi-agent, and supporting open source is important to us in that context,” Altman said.