Calvin Wankheed / Android Authority
Last week, Openi released, which may be the most expensive (and potentially controversial) updated for Chat GPT since its inception: agent. Unlike the standard conversation interface, the agent mode gives GPT to control the chrome -run virtual machine, which can allow it to interact with websites like humans. These websites can log into elements, scrolls, buttons, fill out forms, and, if can approve the credentials, logged into your online accounts.
For the first time since the launch of the AI, it can work instead of spitting some text on ways to do it. At the surface, the ability seems to be endless. Chatboat can respond to your emails, purchase grocery, book flight, or perform more complex tasks spread on multiple websites. The best thing is that you can “move” to the Chat GPT around the Virtual Web browser and visit the Internet (as you can see in the video below). Confession, it’s so much that a toddler is struggling to walk in times, but it is very interesting.
So what can the Chattgat’s Agent Mode actually do with all these abilities? To answer this question, I tested this feature with a couple of real-world tasks-the way you really want to offload AI assistant. Here is how he handled them, and when he faced an obstacle, Chattgat did.
Chat GPT Agent Keeping Working: A Grocery Run
Amazon’s Alexa Sound Command can add toilet paper to your cart, but Chat GPT agent mode can be entrusted to do much more. In particular, it can buy your entire grocery list on any platform of your choice. For example: I gave the agent an easy task: buy everything I need for home pizza from Walmart. I did not provide any special ingredients, items, or even guidance at the price just to see what it chooses.
The agent booted the virtual computer and went to Walmart without time. But it almost Immediately immediately went into a road block-Walmart threw the anti-boot verification screen that requires a human being pressed and holding the button. The shocking thing is that the agent recognized the screen and asked me to briefly control the browser and complete the task. I took control and after about ten seconds, we were inside. I returned the control, and the agent immediately. He came to work. It seems that if they want to keep the boats out in the future, Captchas will need to be reproduced.
Chat GPT called me when it needed human contact, which means that CAPTHS only has to be solved.
Going forward, I searched the agent with the procedure “pizza dough,” “pizza sauce,” “Mozarala cheese,” and “Pepperni”. But my surprise is that the agent did not just take the first result. Instead, he preferred a familiar and good price alternative as I would personally do. In multiple examples, I saw it selected the third or fourth item in the results or call the competitive product at a higher price. The agent also correctly transferred the results of the past inaccurate search results, such as a completely pre -prepared paper pizza when it was simply purchasing the Pepperoni, a component.
Within four minutes, my virtual cart was filled with everything I needed to make pizza. The agent went to the checkout page and then returned the control to me to complete another captcha, login, and secure my payment details. Chat GPT says that when you are under the control of its virtual machine, they can’t see your inputs, probably means that it cannot save your login or credit card information. Nevertheless, I did not choose to enter the details of my login and for that reason, I spent the night without pizza.
No need to say, it was a very impressive exhibition, even if it was a straightforward work with a clear continuity of the process. But can the agent handle something more vague? I decided to give it a difficult challenge: Find a list of car candidates used in the Facebook market.
Can Chat GPT get the perfect used car?
I like the Facebook market because there is always a big deal around the corner, but I am afraid to scroll in hundreds of listings to find the right one. So I put the Chattgat’s agent into the task, fully expecting that he would be stumbled at some point. In particular, I asked her to log in to my Facebook account, to visit the Market Place tab and to compile a list of used Honda Foot samples. As it was not so difficult, I asked him to find only 2015 to 2020 models in Toronto’s 50km circle.
My surprise is that the agent made a strong start and logged into my Facebook account without hiccup. I had to take steps to provide a two -element verification code, but in addition, it was a completely painful process. Chat GPT revolves in the marketplace section, changing the place I provided, and even adjusted the maximum distance filter. He then started going through individual lists and recorded important details about each car.
Fairly, the agent mode of the chatigpt is not faster than humans – I may at least go through an agent, at least in its current state, more individual lists. But like any computer, chatting also contradicts. After looking at the listing for a few minutes, I went away from my computer to catch coffee. When I returned about ten minutes later, the agent was still carefully clicking on individual listings.
Chat GPT is strictly permanent, and has been scrubbed through more than a hundred listings on the Facebook marketplace.
The final result was a comprehensive report of two to three dozen cars that meets my standards. As the above photo shows, the agent had set a detailed brief for each vehicle. It organized key details in columns: year and model, price and mileage, transmission, and location. Even he created a note section where he summarized the seller’s details, in which important information such as “dealer listing,” “Salvation’s title,” or “open to negotiation”. Finally, each line item had a screenshot of the marketplace if I was interested in this particular vehicle to use for me.
If it looks impressive, it’s exactly. Manually collecting such a list otherwise it will take at least a couple of hours. And the best thing is that I might go a step further and I can ask the Chat GPT agent to contact me shortlist sellers from me. It is not a responsible task, but it is definitely a possibility.
Not perfect, but impressive… and scary
Calvin Wankheed / Android Authority
Chat GPT’s agent mode is in no way perfect. I saw him countless times in the endless loop unless I was forced to intervene. It almost Almost can follow the instructions very closely, close to a mistake. For example, it refused to accept the search radius of 51 km instead of 50 km, in which I suggested on various used auto website. Result? The agent used his virtual mouse to “drag” a slider backwards for several seconds until it was needed at 50 km. The general understanding does not naturally come into AI, even in agent mode.
I also saw a huge delay in clicking on a link between Chat GPT or pressing the browser back button into practice. This causes AI action to resume, which is a major problem as it inevitably ends on the wrong page. Sometimes, the agent clicks the back button twice and ends on the homepage. After that, the process will resume the entire search, waste several minutes.
Chat GPT agent can sometimes be a stupid fool, but it’s just going to get better.
But for all the clumsy moments, the real importance of agent mode is not in its current speed or ability. The fact is that I can run to make coffee, while AI has compiled a detailed report about autonomous used cars, there is a significant price suggestion. The possibility of increasing productivity is widespread. But at the same time, I am not sure how I feel about the AI making decisions on my own. So now the question is: How much we are ready to hand over AI?
If you want to use an agent mode, keep in mind that you will need Chat GPT Plus subscription. This will make you return for 20 months, and currently this feature has been closed on only 40 messages each month. But if you are patient, I am sure it will eventually reach free levels like the features of other new chat GPT in the past.
Thank you for being part of our community. Read our comment policy before posting.
