Versatile advanced AI takes first steps: GPT-4 creates simple game with JavaScript, takes bar exam

OpenAI's new neural network model powered by artificial intelligence (AI), the GPT-4, continues to be in the spotlight. Unlike the previous model, the new one can work not only with texts, but also with photos. Also, the GPT-4 handles logical and other tasks, including writing code and passing exams, better than its predecessor.

GPT-4 recently wrote a simple game in JavaScript, and based on a text description that was once a meme. Also, it took a bar exam and did better than most people who have been preparing for this exam for a long time.

Game that's been waiting two years for

Denis Shiryaev, the founder of the Neural Love startup, decided to test the capabilities of the new GPT-4 and asked the neural network to write a game in JavaScript. He showed the AI a letter that a certain “Kirill” had sent to the Russian MiST Land South studio in the early 2000s. Kirill had asked in the letter to release a 3D action game where you can play as forest elves, palace guards and villains and rob caravans. According to Kirill, he has wanted such a game for two years.

The letter had a lot of grammar and spelling mistakes, but the AI got the gist of it and wrote a 2D game in JavaScript. It turned out that the game was simpler than Kirill had imagined, but the game included the caravan he mentioned and four different-colored game zones, as Kirill had requested in his letter.

The player can control the green dot with the arrow keys and shoot with the space bar on the keyboard. His task is to hit the caravan, after which the point is moved to a random place on the map.

The game code and interactive demo can be viewed on the CodePen platform.

GPT-4 passed the bar exam better than most

As it turns out, the GPT-4 gets the hang of jurisprudence, too. It recently scored 297 on a bar exam administered in an experiment by two professors and two employees of legal technology company Casetext. According to researchers, in most US states, a person with such a result will be allowed to practice law.

About four months ago, the ChatGPT also took such an exam, but did not achieve a passing score.

The bar exam includes multiple-choice questions as well as essays and tests designed to simulate legal work. The GPT-4 answered nearly 76% of the multiple-choice questions correctly, beating the average of people taking this exam by 7%. In comparison, the ChatGPT had answered about 50% of the questions correctly.

Also, the AI was able to write an interesting essay, which was no less surprising also for Daniel Martin Katz, a professor at Chicago-Kent College of Law.