In March, OpenAI released the long-awaited GPT-4 AI model, an update of the technology behind its popular chatbot ChatGPT. The organization claims that GPT-4 is the most advanced system that provides safer and more useful answers. This announcement comes less than four months after the launch of ChatGPT, which has grown into the fastest growing consumer app in history. OpenAI also announced that GPT-4 can solve more difficult problems with greater accuracy thanks to its greater general knowledge and problem-solving abilities.
OpenAI added new features like accepting input images and generating legends, classifications and analysis. GPT-4 is also said to be able to handle over 25,000 words of text, enabling use cases such as long-form content creation, advanced conversations, and document search and analysis. In a demonstration video, Greg Brockman, President of OpenAI, showed how GPT-4 can be trained to quickly answer tax-related questions such as: B. Calculating the standard deduction of a married couple and the total amount of tax to be paid.
Announcing GPT-4, a large multimodal model, with our best capability and alignment results yet: https://t.co/TwLFssyALF pic.twitter.com/lYWwPjZbSg
—OpenAI (@OpenAI) March 14, 2023
This model is very good at mental arithmetic. It has rich features that are very flexible,” Brockman said. In another video, the company explains that GPT-4 has a number of features that the previous iteration of the model didn’t have, including the ability to “reason” from user-uploaded images. Although GPT-4 cannot generate text from videos and other dynamic generative content, it gives a glimpse of what multimodal chat will look like in the future. In the future, video, audio, images and any other form of content could be integrated into the chat.
In a blog post published on its website, OpenAI announced that GPT-4 outperforms ChatGPT (GPT-3.5) in its advanced reasoning capabilities, using more data and computation to create ever more sophisticated and efficient language models. The company adds: We spent six months making GPT-4 more secure and better aligned. According to our internal evaluations, GPT-4 is 82% less likely to respond to unauthorized content requests and 40% more likely to provide factual responses than GPT-3.5. This suggests that GPT-4 partially solves the immediate injection problem.
The update isn’t even a month old, but it’s been used for a lot of things.
A complex video game
Ammaar Reshi, design manager at Brex Inc (an American financial services and technology company based in San Francisco, California), has developed a (complex) video game from scratch without any programming knowledge.
And Reshi to explain:
After seeing @skirano’s Pong recreation and @javilopen’s DOOM demo, I wanted to see if I could push GPT-4 to create a full 3D game. One of my childhood favorites was an MS-DOS game called SkyRoads – long forgotten. But… did GPT-4 know?
It turns out yes! It seemed to describe the game’s Wikipedia page, but that was it… so it was a real test – could we bring this game back to life just based on our understanding and a description?
I don’t know how to write a line of code for 3D games, so I asked GPT-4 to choose how the game would do it, it suggested using BabylonJS, HTML and CSS. Then I detailed what we would be borrowing from SkyRoads and how this version would be different – here’s the prompt:
After receiving the code, I pasted it into Replit [ndlr. anciennement Repl.it, est une start-up base San Francisco et un environnement de dveloppement intgr en ligne. Replit tant un logiciel en tant que service, il permet aux utilisateurs de crer des projets en ligne et d’crire du code] just like GPT told me to and it worked! Except…we were a long way from the final product. The green cube, our future spaceship, had no way to turn left or right, no way to avoid those pesky obstacles. Well, GPT fixed that🚀
Then the viewing angle was shifted, you could see this cube from the front, but you had to be behind the spaceship! So I described what the camera should look like and GPT fixed that. It turned out to be a simple one-line change.
Then came the elusive error. The spaceship would shoot into the abyss when I press jump + directional keys and crash the game. GPT-4 and I tried everything for HOURS. I reached out to Claude from @AnthropicAI on @poe_platform. Claude solved it in less than 60 seconds.
It also opened up a new way of working for me. Claude was able to help with small corrections. We had a team.
There were many other errors. The ship fell through the track, the track stopped only a certain length away, but then the dynamic between me and GPT/Claude became clear: I was the playtester trying to break the play, GPT and Claude were doing their best to to fix it It.
Now that the bug is fixed, let’s get back to the fun part! I needed backgrounds, textures and a spaceship model for my game. I found my spaceship model on @Sketchfab and asked GPT to give me the code to replace this green cube with my 3D model file.
Then I went to MidJourney for the background and textures and scaled them with @pixelmator’s super resolution and magic brush to make the texture more seamless.
Then I asked GPT to bring in all my files. Then I made the game harder. The obstacles had to be variable in width, the track had to be separated left and right, I wanted to record the player’s best score so that he had something to beat.
Finally I went to @epidemicsound to add music to the game and then asked GPT-4 to give me the code to add music player, explosion and game sounds when you crash. A classic synthesizer, of course, when piloting your ship!
Other Notes: At one point the code was too long for GPT to understand as a question, so I had to use the API Playground. Pro tip from @skirano, but GPT actually got better when I told them they are a top notch BabylonS developer.
And let’s go! In 3 days, with no knowledge of 3D game programming, I was able to recreate an adaptation of one of my favorite childhood games using a combination of Al and existing tools!
So you think that GPT-4 can’t make a complex game…think again!
This is how I used GPT-4, @ReplitMidJourney and Claude to assemble an AI team and create a 3D space runner from scratch without any knowledge of Javascript or game programming.
Follow a saga! 🧵 pic.twitter.com/Fpg4ANBvCm
— Ammaar Reshi (@ammaar) March 20, 2023
A big data chatbot analyst
Mayo, AI Engineer & Contributor at LangChainAI (Javascript), relied on GPT-4 to build a chatbot capable of analyzing big financial data. In this case, he built a GPT-4 financial analyst “Warren Buffett” to “discuss” and analyze multiple PDFs (~1000 pages) in Elon Musk’s Tesla 10k annual reports (2020-2022).
It has provided a video tutorial for those who might be interested and offers a step-by-step training program for beginners on how to create a production-ready chatbot for their data.
Write code from your smartwatch
Mckay Wrigley explains that he placed an iOS shortcut that leads to GPT-4 and GitHub.
This is how it works:
- identifies the deposit
- identifies target files
- write me the code
- create a branch
- validate the changes
Now I can program directly on my Apple Watch using just my voice!
I attached an iOS shortcut to GPT-4 and GitHub.
This is how it works:
– Identify the repo
– identifies the target files
– Write the code for me
– Creates a branch
– Apply changesNow I can program directly on my Apple Watch using just my voice! pic.twitter.com/zr0fXGJoTb
— Mckay Wrigley (@mckaywrigley) March 28, 2023
Sources: ChatGPT Chabot for your data, use cases (Twitter)
And you ?
How do you read these use cases?
what do you think of them ? Practice ? Completely useless?
do you know any others Which ?