Related Articles
I recently needed to contact the CEO of a startup called Lindy, a company developing personal assistants powered by artificial intelligence. Instead of looking for it myself, I turned to an AI helper of my own, an open source program called Auto-GPT, typing in “Find me the email address of the CEO of Lindy AI.”
Like a delightfully enthusiastic intern, Auto-GPT began furiously Googling and browsing the web for answers, providing a running commentary designed to explain its actions as it went. “A web search is a good starting point to gather information about the CEO and their email address,” it told me.
“I found several sources mentioning Flo Crivello as the CEO of Lindy.ai, but I haven't found their email address yet,” Auto-GPT reported. “I will now check Flo Crivello’s LinkedIn profile for their email address,” it said. That didn’t work either, so the program then suggested it could guess Crivello’s email address based on commonly used formats.
After I gave it permission to go ahead, Auto-GPT used a series of different email verification services it found online to check if any of its guesses might be valid. None provided a clear answer, but the program saved the addresses to a file on my computer, suggesting I might want to try emailing them all.
Who am I to question a friendly chatbot? I tried them all, but every email bounced back. Eventually, I made my own guess at Crivello’s email address based on past experience, and I got it right the first time.
Auto-GPT failed me, but it got close enough to illustrate a coming shift in how we use computers and the web. The ability of bots like ChatGPT to answer an incredible variety of questions means they can correctly describe how to perform a wide range of sophisticated tasks. Connect that with software that can put those descriptions into action and you have an AI helper that can get a lot done.
Of course, just as ChatGPT will sometimes produce confused messages, agents built that way will occasionally—or often—go haywire. As I wrote this week, while searching for an email address is relatively low-risk, in the future agents might be tasked with riskier business, like booking flights or contacting people on your behalf. Making agents that are safe as well as smart is a major preoccupation of projects and companies working on this next phase of the ChatGPT era.
Most PopularGearThe Top New Features Coming to Apple’s iOS 18 and iPadOS 18By Julian ChokkattuCultureConfessions of a Hinge Power UserBy Jason ParhamSecurityWhat You Need to Know About Grok AI and Your PrivacyBy Kate O'FlahertyGearHow Do You Solve a Problem Like Polestar?By Carlton Reid
When I finally spoke to Crivello of Lindy, he seemed utterly convinced that AI agents will be able to wholly replace some office workers, such as executive assistants. He envisions many professions simply disappearing.
“What are the occupations that we give the most money to understand and produce text? No offense, but journalists are one,” Crivello says. Ouch. “Lawyers are another,” he adds. I tried to point out that AutoGPT wasn’t capable enough to find his email address, let alone interview him over Zoom. But he wasn’t swayed.
Using AutoGPT requires some knowledge of how to use the command line and access to OpenAI’s software API. It only really works with access to the company’s latest AI language model, GPT-4. But I highly recommend giving it a try if you can. Forget proficiency in Microsoft 365, formerly known as Office. Learning how to prompt and guide an AI agent might be an indispensable office skill before you know it. (WIRED’s Reece Rogers has written an excellent set of primers on using AI to boost your computer skills that you should also check out.)
After my own test drive of Auto-GPT, I could understand Crivello’s conviction about the future. If the errors can be ironed out—a fairly big if—I can imagine a future where AI agents help with a lot of chores that currently involve typing into a web browser or moving and clicking a mouse. As with ChatGPT, when Auto-GPT works, it can feel like magic. The flaws that remain with chatbots also leave me less convinced than Crivello that these agents can easily take over from humans, or even function without human help, for the foreseeable future.
