Last week, Openai operators did the following for me:
I ordered a new ice cream scoop on Amazon.
I purchased a new domain name and configured the settings.
I booked the date of Valentine’s Day for me and my wife.
I schedule a hair cut.
These tasks were mostly autonomous, but I had to make fine adjustments from time to time and rescue them from the failed attempt.
If you are catching up, or if you are concerned with this week’s Deepseek News, which is covering all other AI news, the operator is a new, so -called AI agent released last week by Openai.
This tool, which was charged as a “research preview”, can only be used by those who pay $ 200 per month to Chatgpt Pro, the company’s best subscription layer. This allows users to use a web browser, fill out the form, and instruct the AI agent that can perform other actions on behalf of the user.
The AI agent is now very popular with Silicon Valley. Some industry officials believe that AI agents, which can use computers, are not simply providing support, but can actually achieve valuable real world tasks, so the next big step of the AI function. 。 Many major AI companies, including Google and humanity, are testing autonomous agents that claim that companies can ultimately hire as full -fledged workers.
I upgraded the Chatgpt subscription, placed the operator at that pace, and checked what AI agents could do for me.
On the surface, the operator is a bit like a normal Chatgpt, but for example, “Buy a 30 -pound dog food on Amazon” (for example, buy a 30 -pound dog food for me “-The operator is a miniature browser. Open the window, enter the “Amazon”, enter the address bar, and follow the instructions.
You may ask some clear questions. (Do you want chicken flavor or beef flavored food? Is it a night shipping or two days?) Then, if you are convinced that it has made the right choice, the operator urges you to finalize and carts the dog food. Put in and place your order. (Operator does not enter a password or credit card number. You need to inherit the mini browser and enter them in yourself, but the rest will be done alone.)
The overall point of the operator is that there is no need to supervise. You can run the task in the background while doing other things. However, I realize that I am glued to the window, click the button, enter the word in the box, select all from the drop -down menu, and click the button. I was fascinated by. Look at Massachusetts, a computer using a computer!
The operator was also impressive about some relatively simple tasks that I gave:
I ordered lunch at DOORDASH for my colleague’s microphone and sent it to his house. (I didn’t tell him what to order, but the operator chose a Mexico restaurant, chose a handful of dishes for him, and turned $ 7 to the delivery person.)
After controlling the Linkedin profile, I responded to hundreds of unread Linkedin messages. (However, it was registered in my fear and webinar.)
I got $ 1.20 by setting an account on a website that provides small cash remuneration to fill out the survey. (I may have made more, but I started to feel guilty by spaming the investigation with the answer with a fake robot.)
However, the operator also failed in other tasks and revealed the restrictions.
Since the operator’s browser was blocked to enter the Times website, it was not possible to scan recent rows and add it to a personal website. (It has been blocked from many other sites, including Reddit and YouTube. Times complains Openai and Microsoft due to copyright infringement related to AI model training).
It doesn’t play online poker for me. (The operator said, “I can’t support gambling and related activities.” This seemed to be a reasonable refusal considering the chaos that gamble bots can be created.)
Also, the Captcha test makes it impossible to log in to many sites. (I was relieved to think that the overall point of CaptChas is to prevent robots.)
As a whole, using an operator was usually more troublesome than worth it. Most of it for me had less headaches and could do it myself faster. Even if it went well, I asked for a lot of confirmation and security before acting, so I didn’t feel like having a virtual assistant and seemed to supervise the world’s most unstable internship. I felt.
Of course, this is the early days for AI agents. AI products tend to improve from version to version, and it is good that the next repetition of the operator is improved. However, in the current format, operators are more interesting demonstrations than I recommend. And most people don’t need to spend $ 200 a month.
However, I think it is wrong to amortize AI agents. When they became more capable, they began to be a substitute for human workers in several occupations. (Openai and Meta say they are already building AI engineer aners.) And some experts will execute commands such as “exhaust bank accounts” or “execute”. If you learn, you are worried that AI agents that are more powerful and untreated can cause safety risks. Cyber attack. “
If you set a lot of AI agents on the Internet to loosen a lot of AI agents, you may cause a rebound from other businesses that paid an invoice, relying on the web publisher, e -commerce site, and traffic generated by humans. (If you are a business to buy an ad on Amazon, you want to be a human being, not a bot that pretends to be a human.) In the future, I have more websites. You can imagine a page or product that you can imagine to take measures to block the AI agent or to control it to identify.
Currently, the AI agent is too incompetent and is too incompetent to become a lot of threats. However, most webs are not so imaginable to imagine the near future consisting of a robot, buying things from a robot, or writing e -mails read only by other robots.
In other words, most autonomous driving Internet is here. Click as much as possible.
