a fascinating demo to understand how AI moves between different websites – Explica .co


You never tire of watching GPT-3 demos. After we had taught ourselves a generator in which we only had to describe the web we wanted and the AI ​​gave us the HTML / CSS code, Sharif Shameem, founder of Debuild, integrated GPT-3 language model in Chrome and asked you to “buy some AirPods”.

The suggestion is simple. The AI ​​has access to the browser as we could. Namely, it has a cursor and it moves through the different webs like we would with the mouse. A very simple demo that will allow you to watch how AI understands the internet, what elements it stops in and where it gets stuck on the different web pages.

A privacy policy prevents GPT-3 from completing the task

We assume that unfortunately GPT-3 not yet advanced enough to be able to shop by yourself. The internet is complex enough to bundle up and lose the AI.

Here is the demo of GPT-3 posted on Twitter by the same author.

The first step in AI, like most of us, is googling. Enter the address. AI also understands where to write on the Google web, in this central box. “Buying AirPods” as a keyword seems like a good strategy for getting them.

Instead of clicking on the first search option, AI tries to find “the best deals” from AirPods that you can buy right now. With that in mind, the cursor moves down within Google and decides to enter an article from TheVerge. (It’s a shame that GPT-3 does not rely on the AirPods offers from Xataka Selección).

As soon as the article is in the article, the cursor moves and GPT-3 can understand what the purchase link in the article is. In the special case of the demo, this link leads to Walmart. We don’t know what criteria the AI ​​follows, but the fact that Walmart is a recognized store certainly helps.

Once you are on the Walmart product page, before clicking buy, the AI ​​goes to the conditions and the privacy policy. At this point, the complexity is great, as the AI ​​could have been mistaken for an advertisement or bought via a false link.

Somehow GPT-3 is trying to know what kind of buying process is being carried out and ends with you entering a privacy policy that completely derives you from your original mission. There’s no element on this privacy page that the AI ​​can relate to the task of buying an AirPod and then it fails.

A first step in creating wizards that navigate the web

The demo creator explains that the order could be split off in several steps, where each of these tasks would generate its own AI interaction. It could then be programmed to select the best interactions, a kind of “tournament group”. The possibilities of GPT-3 are enormous, but this is where the skill of the programmer comes in to make the information more precise.

The ability for an AI to perform such actions with such a simple command would open the door to thousands of applications, but for now, GPT-3 is still a long way from achieving that. It is fascinating to see GPT-3 in action in a task that is akin to what we could be doing.

For now, the demo will be used to attract engineers with React skills to help the author create a tool prototype that takes advantage of GPT-3 to move around using web interfaces only. Business tools based on this language model should not be long in coming. And they promise a much higher degree of automation than today.

In Engadget | Chatting with your own self as a future option: Artificial intelligence and systems like GPT-3 make it possible

Leave A Reply

Your email address will not be published.