a week ago I tried the new ChatGPT Atlasthe new OpenAI browser and, although it has a lot to improve, it seemed like a threat to Google’s dominance with Chrome. Today I put it to the test again, this time with a Plus subscription, and I wanted to check if agent mode is capable of hmake the purchase at Mercadona.
Posing the situation
It was the first time I used ChatGPT for something like this and I didn’t want to just give you a list. of the purchase, so first I asked him for ideas to make healthy recipes that are delicious. He offered me several options and when I decided on one of them, I activated agent mode and asked him to buy the ingredients at Mercadona.
We have already talked about AI browsers are vulnerable to prompt injection attacks and OpenAI knows it. Before starting, a message appeared alerting me that using agent mode carries risks and I could use it with or without the session logged in. In my case I have chosen the logged in session because I wanted to see it work more easily, but as a precaution I have first deleted my payment details on the Mercadona website.
Making the purchase
Once the risks have been accepted, agent mode has been activated and the mouse has started to move through the Mercadona website interface. The sidebar shows the model’s entire thought and decision-making process while buying the ingredients to make a chickpea curry. In the video you can see the entire purchase process.
The agent has been making decisions when he has found several items to choose from. For example, the recipe required an onion, but decided that it was more practical to buy a 1kg package. However, when choosing spinach, he decided that a package of baby spinach was better than the large package that is much cheaper. When he finished choosing ingredients he asked me to check it and I asked him to change the spinach. He has done it without question.
The process has stopped when it has run into an insurmountable obstacle: it only had 10.28 euros and the minimum order on the Mercadona website is 50 euros, so I asked it to also include the ingredients of another of the recipes that it had suggested to me at the beginning, one for baked salmon. Since that one didn’t reach the minimum order either, I told him that I wanted to make it for four people and please don’t give me frozen salmon, but fresh ones.
The agent adjusted the quantities and changed the salmon for a fresh one, but it still didn’t reach 50 euros, so I asked for something more creative.: to look for the most viral Mercadona products recently and add them to the basket.
The purchase is made for you, but there is a problem
When he was done, it was time to check the basket. I found that I had added garlic and also purple garlic. The normal garlic was fine, but the purple ones? I have reviewed the chain of thought and he was confused looking for purple onion. Mercadona calls it “red onion” and the agent has decided that it was better to add purple garlic because the color matched, even though they were a different ingredient.
Regarding the viral products, I have chosen an advent calendar with makeup, smoked raclette cheese, cookie nougat and pistachio cake. The total amount was 66 eurosit is true that I have not expressly told you to adjust to 50 euros, but it seems to me that you have gone a little overboard.
The agent has taken control of the browser and done exactly what he wanted: make the purchase for me. However, there is a problem and that is It’s very slow. I haven’t helped much either. Not having anticipated that there was a minimum order and the additional requests that I have been making, such as changing the quantities or choosing products by itself, has made it even slower.
In total he has been thinking for almost 15 minutes, but if we take into account only the first part of the purchase for the chickpea curry, it has taken 2:14 minutes. More than two minutes to add eight items to cart. All the time I had the feeling that I would already have the order finished and paid for.
Regarding reliability, I have to say that He has made fewer mistakes than I expected, but it is still necessary to check what you have added to the basket at the end because you can sneak in some garlic instead of onions, and I already have enough garlic in the pantry.
Much more practical in other scenarios
One of the use scenarios that OpenAI gave in the presentation of its new browser was precisely to make the purchase. After trying it, it is clear to me that the ChatGPT Atlas agent mode has a lot of potential, but not for making the purchase, that’s why I have tried another scenario where it can be much more useful: organize a trip.
I asked him to find places for me to go on a getaway over the December long weekend, that were less than 2 hours by car from Valencia, with a specific budget and to look for them on Booking and Airbnb. In six minutes he gave me options for two different destinationsorganized in a table with price per night and highlights. Once I have decided, I only had to give him the personal information to complete the reservation.
To organize a trip it is practical.
Making the purchase is simply adding things to the cart, a much more mechanical process that we can do manually in a very short time. If we also encounter obstacles such as the minimum order or we are not completely clear about what we want, we end up losing more time than gaining it.
Where the agent does offer more potential is for more complex tasks that require more steps to filter information. For example, when organizing a trip you have to do many more things; choose a destination, plan routes, search for accommodation, check that they fit the budget and much more. In this case We can leave agent mode thinking while we do other things And it does save us a lot of time.
Images | Amparo Babiloni, Xataka
GIPHY App Key not set. Please check settings