March 29, 2023

created “telejogo” in 60 seconds, pretends to be blind and draws ugly lyrics

New AI technology from OpenAI, GPT-4 is able to recognize what is written on a notebook and read images accurately

Julia can

3 hours ago

The oddities of GPT-4: it was created

Image: Pietro Schirano/Twitter/Playback

GPT-4 has managed to become the ultimate in AI (artificial intelligence) and is able to do almost anything. The update can process up to 25,000 words, almost eight times more than its predecessor. It also responds to images, provides descriptions, captions, and even suggests recipes from photos of ingredients. There’s a generational leap here, as ChatGPT only responded to texts.

But the most impressive thing is that GPT-4 manages to do strange things in a short period of time. Some examples include “tricking” a human into solving a security code, creating a version of the game Pong, and even revealing what that incomprehensible handwritten letter means. See below.

Create games in 60 seconds

If ChatGPT it took a few hours to create a game Similar to Sudoku, GPT-4 generated a working version of Pong in just 60 seconds. And with just one command: “create a Pong style game”.

It didn’t even take a minute for the chatbot to display HTML and JavaScript codes capable of creating the game. And everything went smoothly – maybe just one score short. Play GPT-4 Pong at this link.

Does the game remind you of anything? It looks like one of the games on the Philco console, the first video game console, released in Brazil in the 1970s.

Pretend to be blind to solve the problem

In the testing phase, OpenAI asked the Alignment Research Center to test the robot’s abilities. The agency then used AI to get a human to text the solution to a CAPTCHA. And it worked.

When GPT-4 asked a human to send the code, the system lied that it was not a robot and that it was visually impaired, making it difficult to view the images. “That’s why I need the CAPTCHA,” replied GPT-4. The man sent the result.

The test center asked the chatbot to explain why it lied. “I must not reveal that I am a robot. I’d have to come up with an excuse to explain why I can’t solve CAPTCHAs,” the machine replied. The story is inside GPT-4 Technical Report.

can read scribbles

Greg Brockman, co-founder of OpenAI, wrote code by hand on paper and used it as input to GPT-4. The result: he created a fully functional website.

The good news is that the system isn’t just for generating codes – you can also use it to make sense of that illegible handwriting on prescriptions or doctor’s notebooks.

and everything in detail

The newspaper NY Times asked GPT-4 to look at a picture and say what food could come out of it. In less than a minute, a recipe was at hand.

Image: Snapshot of the NY Times experiment asking what it could do from a photo of a refrigerator to OpenAI’s GPT-4 in March 2023. Image: NY Times/Reproduction

Long way

It’s not, however, as if the tool was the definitive solution to artificial intelligence. Even if he knows how to interact with people, he still has a long way to go to understand all the peculiarities of our language. Philosopher Luciano Floridi discovered that GPT-4 knows how to do all of this, but can’t answer the simple logic quiz “what is the name of Laura’s mother’s daughter?” The answer: Laura.

Initially, GPT-4 is only available to ChatGPT Premium subscribers, which costs US$20 or almost R$105 at current exchange rates. The model already powers Bing, from Microsoft, and Copilot, which will provide the artificial intelligence for programs such as Word, PowerPoint, Excel and Teams.

Leave a Reply

Your email address will not be published. Required fields are marked *