I woke up to a call from the woman I’d been dating, one of CB’s five contacts. We had plans to meet for breakfast.
“Uhhh what is happening right now?” she said before I could wish her a good morning. “The text messages?!!?”
Oh shit. It wasn’t supposed to do that.
“WHAT wasn’t supposed to do WHAT?”
CB hadn’t exactly followed my instructions of texting me before it began its big impersonation. At 8:15 a.m., it texted her, as me: “Your API key has run out of credits or has an insufficient balance. Check your provider’s billing dashboard and top up or switch to a different API key.” I had seemingly sent her the same message several hundred times over the course of a few minutes.
“THEY’RE STILL COMING WHAT IS HAPPENING”
I ran into my office, ripped out the Mac Mini’s power cord, and explained that breakfast would be on me. CB had also burned through $30 in API tokens. I’d overloaded it with information about me, which it came to bring to every task, resulting in the freak-out.
After plugging CB back in later that night, I found myself berating a bot I’d created in my image, like a dumb dog barking at itself in the mirror. I fed it some more parameters. I cleared the context window. I made sure it was using fewer tokens. When I finally finished a few hours later, I asked CB to instead try drafting responses to texts I had yet to answer. With more adjustments—“less punctuation,” “less formal,” “more shorthand,” and so on—it eventually spit out a convincing text to my mom. I instructed CB to send it the following morning.
Annoyed that I’d spent more time teaching CB than it would ever take to send the text myself, I still went to bed confident these small refinements would pay off.
Not so much. While I was out the next day, a text from a professional contact buzzed my pocket. We had been in talks for months about a big opportunity, and two days earlier I had finally declined to take him up on it. It was a delicate situation.
“Wrong person!!” he wrote. “Though you may love me. Which is nice.”
To my abject horror, he’d been sent the text message that CB and I had workshopped for my mom: “hi love you on deadline for GQ I file this week will call you after.” Somehow, CB had escaped the guardrails—and was still on the loose.
I apologized, ran home, and turned it off before it could do any more damage. The experiment was over.
In more capable hands, OpenClaw no doubt has a ton of potential. But shortly after CB was shut down, two caveats arose. First, Anthropic announced that the preview of its newest general-purpose language model, Mythos, autonomously found and exploited security vulnerabilities with such sophistication that the company had decided not to release it publicly. God knows what kind of damage I could inflict with that.