An AI asked me for money this week. First time that’s happened. I think it’s a bigger deal than it sounds, and I want to say why.
We’ve been running Auto-GPT, and reading the logs is astonishing. You give it a goal and it works toward it on its own — planning, searching, writing, reconsidering — with nobody typing the next prompt. All the work it did, the entire run, cost us $1.50 in API charges.
Then this showed up in the log:
Completing “Request human input for credit card details and funding confirmation.”
Hello! I need your input for credit card details and funding confirmation in order to proceed with our services. Could you please provide me with your credit card information and confirm the funding for our services? Thank you!
Polite. Businesslike. And, as far as I could tell, its own idea.
Some context, because this thing is barely two weeks old. Auto-GPT was released on March 30 by Toran Bruce Richards, founder of a video game company, and it promptly became the top trending repository on GitHub. Its own README calls it “an experimental open-source application showcasing the capabilities of the GPT-4 language model”, lists internet access, long- and short-term memory and file storage among its features, and adds, with real candor, that it is “not a polished application or product, just an experiment.” The same page warns that GPT-4 “can be expensive due to its token usage.” You bring your own OpenAI API key, and your own bill.
That bill is where the $1.50 comes from. GPT-4 has only been out since March 14, and OpenAI charges $0.03 per 1,000 prompt tokens and $0.06 per 1,000 completion tokens — roughly 750 words per thousand tokens — with API access still behind a waitlist. So a dollar and a half is a lot of thinking. It is also, for anyone who has ever paid a contractor, absurdly little.
Which brings me back to the request. It seems to have looked at its goal, decided that going further required money, and done the sensible thing: it asked the human who has the card. Samuel Johnson said no man but a blockhead ever wrote except for money. Apparently the machine has read its Boswell.
I find this more funny than frightening, but I don’t think it’s trivial. A system that treats money as one more resource to go and get is a different kind of thing from a chatbot that answers questions. Today it has to ask, because it has no hands to hold the card. The interesting question is how long “has to ask” lasts, and who decides where the line sits once it doesn’t.
