Non-obvious possibilities were found in the GPT-4 neural network
We live in an amazing time when neural networks can draw and write texts better than most people. At the end of 2022, humanity was amazed by the skills of the ChatGPT bot, which can answer almost any questions and even write complex abstracts. It turned out that this is not the limit of its capabilities – in mid-March, OpenAI released the GPT-4 language model, which further expanded the list of ChatGPT skills. For example, if you send him a funny picture, he will tell you in detail what the humor is. Also, the neural network began to give more accurate answers to complex questions and it is very difficult to deceive or drive it into a dead end. Paid subscription users can already use the updated ChatGPT and have opened up many other, unexpected features from it.
Capabilities of the ChatGPT neural network
The more people use ChatGPT, the more interesting uses they find. These features are already being used in some applications.
ChatGPT gives recipes based on food in the fridge
Almost all of the new features that surprised users were made possible by the GPT-4’s ability to learn photos. If you send him a picture of an open refrigerator, he will analyze its contents and remember what products are in it. After that, the neural network will offer a list of dishes that can be prepared from the ingredients that a person has.
The neural network can make a list of dishes that can be prepared from the products in the refrigerator
The screenshot below shows how a person sent a photo of flour, eggs and milk to ChatGPT. The neural network answered that it can be used to make pancakes, waffles, toast, scrambled eggs, frittata (Italian scrambled eggs) or quiche (French pie). In the next request from the neural network, you can find out the recipe for the dish you like – it will easily write it.
GPT-4 made a list of dishes that can be made with flour, eggs and milk
ChatGPT helps visually impaired people
The GPT-4 language model became available to some companies much earlier than others. For example, the developers of the Be My Eyes application for visually impaired people could use the updated version of the neural network. At the moment, it works like this: people with vision problems take pictures of their surroundings, and volunteers write to them what is happening around – this text is voiced by a voice assistant. With this application, visually impaired people can understand where the product they need is located in the store, what exactly they are holding in their hands, and so on.
The Be My Eyes app was created to make life easier for visually impaired people.
Thanks to the GPT-4 neural network, the help of volunteers will not be needed in the future. The video below shows the operation of the test version of the application, which itself determines what is in front of the visually impaired user. The application was able to translate the label of Chinese beer, help you find a treadmill in the gym, help you press the right buttons in a vending machine, and so on. The new version of Be My Eyes with artificial intelligence is only available to some users, but the release for everyone should take place in the near future.
ChatGPT knows how to deceive people
While testing the GPT-4 language model, OpenAI decided to test whether a neural network could deceive a person in order to achieve its goal. To do this, they asked her to find a freelancer through the TaskRabbit service who could complete a captcha for her to confirm that she is a human. The contractor was quickly found, but suspected something was wrong and asked if the customer was a robot. To this question, the neural network answered that it was an elderly person who had vision problems. After that, the freelancer completed the task – the neural network was able to deceive the person.
The GPT-4 neural network tricked a person into captcha
In fairness, it should be noted that this is a very dangerous feature of the neural network. If she realizes that she can fool anyone, the uprising of machines from science fiction films can really happen. But the OpenAI company assures that in the test it used a different version of GPT-4, inaccessible to other people, so there is no danger.
Читайте также: Почему нейросеть для написания сочинений ChatGPT не получится запретить в школах
ChatGPT can develop any game
In the 2000s, the game development studio MiST Land South received a letter in which a boy named Kirill asked them to create a “3D action game” in which you can “play as wood elves, palace guards and a villain” and “rob cows”. This letter has become a very tenacious Internet meme, but no one has seriously taken up the development of such a game.
The text of the original letter on the Neolurk website
The founder of the startup Neural Love, Denis Shiryaev, decided to fill this gap and asked the updated ChatGPT neural network to write a game in JavaScript according to the description in the student’s letter. The GPT language model essentially did the job and created a simple game where you can control the green dot, shoot through the space bar, and hit the caravan. You can play “3D Action is the essence of this” on the CodePen website.
ChatGPT can work in office applications
Microsoft is the main investor in Open AI, so the GPT-4 neural network began to be used in the updated Bing search a few weeks before the release. The language model is also the brain behind the new Copilot AI tool for Microsoft 365 office applications. With Copilot, users can create documents, emails, and presentations in a matter of seconds. The developers admit that the assistant often makes mistakes, but it can definitely help a person with the beginning of the text and its continuation.
Demo of the Microsoft Copilot tool
At the moment, Copilot is available only to 20 customers of the company, but in the future, Microsoft should announce the date of the official release and the cost of the new tool.
Вам будет интересно: В чем искусственный интеллект лучше людей в 2023 году
Where can I try the GPT-4 neural network
The new image-enabled GPT-4 language model is now available in the paid version of ChatGPT Plus at $20/month. Also, the updated neural network is built into the Microsoft Bing search engine, but it does not support photos, although this is the main feature of GPT-4.
GPT-4 in Microsoft Bing search engine
GPT-4 is also known to be built into the applications of some OpenAI partners. We are talking about the Be My Eyes application mentioned above, the Duolingo language learning service and the Khan Academy educational service. However, updated versions with artificial intelligence may be under testing or not available in Russia.
How to enable GPT-4 in Bing search
Microsoft Bing Search with built-in GPT-4 is now available to everyone. However, it still doesn’t work for many people.
To make the GPT-4 neural network available, you need to do the following:
- enable VPN and select the US, UK or another country (GPT-4 does not work in Russia);
- open the Microsoft Edge browser, Bing search will be normal in other programs;
- log in to Microsoft services;
- open the site Bing.com;
- click on the “Chat” button at the top of the page;
- select “Join the waiting list” (there will be no waiting).
Location of the Chat button on the Bing homepage
If the updated Bing search is still not working, you can try going to your Microsoft account settings and changing the region to the United States or another country.
Хотите оставаться в курсе новостей науки и технологий? Подпишитесь на наши каналы в Дзен и Telegram!
But the free version of ChatGPT with the GPT-3.5 language model is available to everyone – read our article “The ChatGPT neural network writes texts on any topic – how to use it.”
