Last week, Openai’s operator has done the following things to me:
-
He ordered me a new ice cream scoop on Amazon.
-
He bought me a new domain name and set up his settings.
-
Keep Valentine’s Day Day for me and my wife.
-
A haircut was scheduled.
He did these tasks mostly autonomously, though I had to push them from time to time and occasionally save it from a loop of failed attempts.
If you are just attracting-or if you have distanced away from Deepseek News this week, which has overshadowed all other AI News-the operator is a new so-called AI agent released by Openai last week.
The tool, which was charged as a “research preview”, is only available to people who pay $ 200 a month for the company’s highest level of subscription, Chatgpt Pro. It enables users to direct an AI agent who can use a web browser, complete forms and take other actions on behalf of the user.
AI agents are all the rage in Silicon Valley right now. Some industry trusts believe that it is the next big step in AI capabilities, because an AI agent who can use a computer can really achieve valuable real world work instead of providing help. Many of the leading AI companies, including Google and Anthropic, are testing autonomous agents who claim that companies will ultimately be able to “hire” as complete employees.
I upgraded my Chatgpt subscription to put the operator through his rhythms and see what an AI agent could do for me.
On the surface, the operator looks a bit like normal chatgpt, except that when you give him a job-“Buy me a 30-pound dog food bag in Amazon”, for example-The operator opens a tiny browser window, types ” Amazon.
Can ask some clarification questions. (Do you want food with chicken flavor or beef flavor? Overnight or two days?) Then, when you feel confident that it has made the right choice, your operator urges a final confirmation, puts food for dogs in your cart and places the order . (The operator will not enter passwords or credit card numbers-you have to take the mini brochure and type these things to yourself-but do the rest on its own.)
The whole thing of the operator is that you do not need to supervise it – it can perform the background tasks while doing other things. But I found myself stuck in the window, fascinated by the vision of a self -destruction of the web browser by clicking buttons, typing words in boxes and choosing a menu menu, all on its own. Look, MA, a computer using a computer!
The operator also did impressively well in some relatively simple duties I gave him:
-
He has successfully ordered the meal at Doordash for my colleague Mike and sent it to his home. (I didn’t tell him what to order him, but the operator chose a Mexican restaurant, chose a handful of dishes for him and even put the delivery person 7).
-
He responded to hundreds of unknown LinkedIn messages for me, after giving control of my LinkedIn profile. (Although, for my horror, he also recorded me for a webinar.)
-
It made $ 1.20 for me by creating accounts on sites that offer small cash rewards to complete the surveys. (He could have done more but I started feeling guilty of spamming with fake robot research.)
But the operator also failed in a bundle of other duties and revealed his limitations:
-
It could not scan my recent columns and add them to my personal website because the operator’s browser was blocked by entry into the Times website. (It is also blocked by several other sites, including Reddit and YouTube. Times sued OPENAI and Microsoft for copyright violations related to AI models training.)
-
He wouldn’t play online poker for me. (The operator replied, “I can’t help with gambling or related activities”, which seemed like a logical rejection, given the chaos a bot that could create.)
-
And was prevented from connecting to various locations with CAPTCHA tests. (Which I found reassuring, since the whole point of Captchas is to prevent robots.)
Overall, I found that the use of the operator was usually more problem than it was worth. Most of what they did for me I could have done myself faster, with fewer headaches. Even when it worked, he asked for so many confirmations and assurances before I acted that I felt less like I had a virtual assistant and more as if I were to oversee the most insecure practitioner in the world.
This is, of course, first days for AI agents. AI products tend to be improved from version to version and it is a good bet that the subsequent reps of the operator will be better. But in its current form, the operator is more of an interesting demo than a product I would recommend to use – and certainly not something most people have to spend $ 200 a month.
This is what I think is wrong to delete the AI ​​agents. When they become more capable, they could begin to replace human workers in some professions. (Openai and Meta have already said that they are manufacturing AI engineers.) And some experts are worried that the most powerful, uncontrolled AI agents could create safety risks if they learn to execute orders such as “bank account drainage” or “execution”. Cyberettack. “
Defining a beam AI agents loose on the internet could also provoke reaction from web publishers, e -commerce sites and other human -based businesses to pay their bills. (If you are businesses that buy ads on Amazon, you want these ads to see people, not the bots pretending to be people.) In the future, I can imagine more sites taking measures to block AI agents or I direct to certain pages or products.
At the moment, AI agents are very unable to be very threat. But it does not need much imagination to envision a near future, where most of the tissue will consist of robot robots, buying things from robots and writing emails that will read only other robots.
The self-guiding internet is almost here, in other words-take your clicks while you can.