Here is the result in plain text:
In the past week, OpenAI’s Operator has done the following things for me:
Ordered me a new ice cream scoop on Amazon.
Bought me a new domain name and configured its settings.
Booked a Valentine’s Day date for me and my wife.
Scheduled a haircut.
It did these tasks mostly autonomously, although I did have to nudge it along from time to time and occasionally rescue it from a loop of failed attempts.
If you’re just catching up — or if you’ve been distracted by the DeepSeek news this week, which has overshadowed all other A.I. news — Operator is a new so-called A.I. agent released last week by OpenAI.
The tool, which was billed as a “research preview,” is only available to people who pay $200 a month for the company’s highest subscription tier, ChatGPT Pro. It gives users the ability to direct an A.I. agent that can use a web browser, fill out forms and take other actions on a user’s behalf.
A.I. agents are all the rage in Silicon Valley right now. Some industry insiders think they’re the next big step in A.I. capabilities, because an A.I. agent that can use a computer can actually accomplish valuable real-world tasks, rather than just provide assistance. Many of the leading A.I. companies, including Google and Anthropic, are testing autonomous agents that they claim that companies will eventually be able to “hire” as full-fledged workers.
I upgraded my ChatGPT subscription to put Operator through its paces and see what an A.I. agent could do for me.
On the surface, Operator looks a bit like regular ChatGPT, except that when you give it a job — “Buy me a 30-pound bag of dog food on Amazon,” for example — Operator opens a miniature browser window, types “Amazon.com” into the address bar and starts clicking around, trying to follow your instructions.
It might ask a few clarifying questions. (Do you want chicken-flavored or beef-flavored food? Overnight shipping or two-day?) Then, once it’s feeling confident it has made the right choice, Operator prompts you for a final confirmation, puts the dog food in your cart and places the order.
The whole point of Operator is that you don’t have to supervise it — it can carry out tasks in the background while you’re doing other things. But I found myself glued to the window, mesmerized by the sight of a self-driving web browser clicking on buttons, typing words into boxes and selecting from drop-down menus, all on its own.
Operator also did impressively well on a few relatively simple tasks I gave it:
* It successfully ordered lunch on DoorDash for my colleague Mike and sent it to his house.
* It responded to hundreds of unread LinkedIn messages for me, after I gave it control of my LinkedIn profile.
* It made $1.20 for me by setting up accounts on websites that offer small cash rewards for filling out surveys.
But Operator also failed at a bunch of other tasks and revealed its limitations:
* It couldn’t scan my recent columns and add them to my personal website, because Operator’s browser was blocked from entering the Times’s website.
* It wouldn’t play online poker for me.
* And it was prevented from logging into a number of sites by CAPTCHA tests.
In all, I found that using Operator was usually more trouble than it was worth. Most of what it did for me I could have done faster myself, with fewer headaches. Even when it worked, it asked for so many confirmations and reassurances before acting that I felt less like I had a virtual assistant and more like I was supervising the world’s most insecure intern.
This is, of course, early days for A.I. agents. A.I. products tend to improve from version to version, and it’s a good bet that the next iterations of Operator will be better. But in its current form, Operator is more an intriguing demo than a product I’d recommend using — and definitely not something most people need to spend $200 a month on.
That said, I think it’s a mistake to write off A.I. agents. When they become more capable, they could start to substitute for human workers in some occupations. And some experts worry that more powerful, unrestrained A.I. agents could pose safety risks, if they learn to carry out commands like “drain a bank account” or “execute a cyberattack.”
Right now, A.I. agents are too incompetent to be much of a threat. But it doesn’t take much imagination to envision a near future where most of the web will consist of robots talking to robots, buying things from robots and writing emails that only other robots will read.
The self-driving internet is almost here, in other words — get your clicks in while you can.
Source link
