Ordering a pizza online is wildly more interesting when AI does it.
When OpenAI announced ChatGPT Agent, an AI-powered tool that handles complex tasks from start to finish, I didn’t fully understand its capabilities. OpenAI says it can be your personal assistant, but that means little to someone who’s never had one. I imagine if I did have a personal assistant, I might ask them to order me a pizza. So, that’s what I did.
Near CNET’s New York office is a relatively new spot named Cello’s Pizzeria that’s become our go-to place. It’s a slightly elevated slice shop that offers a naturally acidic sauce on top of a snappy crust with a fresh shaving of Parmigiano-Reggiano just as it comes out of the oven.
It took 6 minutes for ChatGPT Agent to order my Margherita pizza. I watched as it processed my request, opened up a virtual desktop where it navigated to Cello’s website, found the pie I wanted, added it to the cart and proceeded to checkout. For security reasons, ChatGPT Agent had me enter the address and credit card information. I saw ChatGPT Agent move a virtual cursor on a web page being processed at some unknown data center. Seeing it comprehend and move through an interface made for humans shows how far AI has come from simply being able to generate funny poems or essays on King Tut.
A GIF of ChatGPT Agent ordering a pizza.
The internet is about to be flooded with AI agents. These AI bots can do more than simply scrub through websites to siphon data. Instead, AI agents can understand complex queries and can accomplish multistep tasks. And since AI agents are built on large language models, like ChatGPT and Gemini, users can use plain language to have these agents run out and get tasks done, kind of like the Cell Jrs. from Dragon Ball Z. So, if you need to find hotels near the San Diego Convention Center, ChatGPT Agent can spend up to 20 minutes scrubbing through Expedia and other travel sites to find you exactly what you’re looking for.
OpenAI CEO Sam Altman believes that AI agents will join the workforce this year.
At the same time, people with malicious intent can deploy AI agents to go after companies and individuals, seeking vulnerabilities and sensitive information.
AI agents aren’t intelligent in the same way humans are, and can be manipulated via tactics like prompt injection. This is when an AI reads a piece of text that makes it do things it wasn’t meant to do. If your AI agent is tasked with helping balance your bank account, a compromised agent may give your login info to bad actors. This is all still very early and speculative, and companies are working to patch vulnerabilities as they arise. But since AIs interact via human language, there will always be some degree of susceptibility from “vibe hackers.”
About 30 minutes after the ChatGPT Agent placed my order, the delivery driver called. Obviously, he had no idea that my new automated “personal assistant” placed this order. Should I be concerned that AI can now interact with online systems and make humans do things?
I’m not sure. At the very least, the pizza was delicious.
Read the full article here