GPT-3’s 50 ways to play tells you it’s cool, but it doesn’t pass the Turing test
Since OpenAI opened GPT-3’s API, many people have rushed to apply for the trial and obtained the “tickets”, and have pushed their trial results on social networks, sparking a heated discussion about GPT-3.
AI Tech Review found a GitHub project that shows 50 different ways to play GPT-3, including generating web pages, charts, code, text, and reasoning with GPT-3. Surprisingly, GPT-3 can generate Keras’s code for convolutional neural networks, interview programmers, reply to emails, write integral expressions, and answer physics questions. In particular, it was given a Turing test, and GPT-3 did just fine. But does GPT-3 really pass the Turing test?
Don’t worry, let’s see what wonderful applications it has first. AI Technology Review showed students 16 demos in this GitHub, most of which were derived from videos on Twitter.
1. Generate HTML layout and code according to the description
Have you ever heard of using language to make web pages? The application generates an HTML web page layout based on the natural language description entered, along with the corresponding code. For example, when we type “Create a button that looks like a watermelon,” we generate the green and red gourd button shown below. Or type “welcome to my newsletter in large font,” which generates the large red font shown below.
Surprisingly, when you type in “Ugliest emoji ever”, the app responds with ~
Is it that ugly? !
2. Create a UI design based on the description
Or use language to do mobile UI design? The app just needs to enter a text description to generate a corresponding UI, which is similar to the previous one, but better adapted to the phone’s operating system. For example, we type “Customer section. With the title ‘Trusted by’ and customers: Twitter, Airbnb, OpenAI, Notion, and Airtable “, the app generates a row of app ICONS that are identical to the official ICONS.
Or we can type “Create a navigation bar with a camera icon, Photo as the title, a message icon, send photos in the app, each photo has a user icon, a like icon, and a chat bubble icon”, and the following app interface is generated, which does all the things.
3. Expand the information in the table
The interface is fairly simple, and GPT-3 can help you take a census. The app counts the number of states in the United States. The table below shows the number of people in Illinois, California and Ohio, and the app can search for the number of people in Michigan and Alaska.
4. Generate diagrams and Python code according to the description
In addition to the census, data analysis of household members is also required. The application also generates bar charts based on natural language descriptions, along with the corresponding Python code. For example, if we type “in my family, my sister is 5 years old and my mother is 46 years old,” the app can generate a bar chart of the age of “sister” and “mother”, and if we add “my father is 6 years older than my mother,” we can generate a bar chart of “father” is 52 years old, and the generated code syntax is quite rigorous.
5. Generate LaTeX expressions according to the description
GPT-3 has statistical and analytical skills, as well as mathematical skills. This application generates LaTeX expressions from natural language descriptions, such as “x squared plus 2 times x”, and we get the expression “x^2+2x”. Or we enter “the integral of the interval a to b of f(t) is equal to F(b)-F(a)”, and we get the classical Newton-Leibniz formula.