Large Language Models such as ChatGPT, Claude, and Gemini, are powerful tools for writing, research, and programming. While I am not in any way shape or form a programmer, I've found that asking an LLM to create a game is a good way to test it's capabilites.
So, with the release of Claude 3 (Opus), I decided to give it the ultimate test; create Tetris!
Using a very sophisticated and well planned out prompt, I asked three different LLMs to create the game; ChatGPT4, Gemini Advanced, and the new Claude 3 Opus. The results, just may surprise you.
Let's start with the reliable ChatGPT 4. As I have experience with GPT4 in the past, its first response was that it was a complex task to create a fully functional game, but could give me a simplified version to start.
Unfortunately, the game itself did not start. GPT4 gave me a response full of functions, styles and placeholders. Many, many placeholders that it said would need to be filled in. Being that I am not a programmer, I would not be filling it in.
So, not striking gold on my first hit, I decided to move over to Google Bard, I mean Gemini Ultra, but actually shown as Gemini Advanced. Keep in mind, this isn't the newer announced 1.5 (announced about a week after 1.0) this is the currently available version.
With that out of the way, I gave Gemini the same prompt. Again, was a bit disappointed. While the Gemini code was more complete, at first, it only gave me a part of the code, so I asked to finish the code. It proceeded to give me more, but still just another part of the code. While these didn't have the GPT placeholders, it would just write part of the code, stop, and ask if I needed anything else.
It took about 4 prompts to get a complete code, and while it did present something of a game board, score area, and what could have been the game, no Tetris blocks did fall.
Claude was my last chance. Once again, I prompted with the same words that GPT4 and Gemini were given, and to my surprise, Claude right away started giving me the code, and kept going. After it completed (with the full code in one quick response), I tested what I was given, and surprised that I had a functional game.
Having to only give a few extra prompts to speed up movements and account for block rotation, which my original prompt didn't ask for, the game itself worked, even with my extra game mechanics I had added in to all the prompts.
Instead of just creating a standard Tetris game, and because I don't want to be sued by the folks behind Tetris (which if you've seem the movie, you know why), I asked each LLM to create a Tetris-like game that used colored Tetris blocks that had to be grouped together in order to be removed from the game board and give points. This provided a bit more of a challenge.
While this isn't the definitive test to name Claude 3 the overall champion, it did give some great insights into the pros and cons of each.
While I love ChatGPT, probably the one I have used the most, it does have a tendency to need to be pushed. If told something can't be done, you just have to tell it to do it, and most of the time, it will. I've also experienced a great deal of time-outs on GPT, where halfway through a long response, I'll get a server disconnect error, requiring the prompt to be reran.
Gemini tends to give what I ask for, but only in parts. While it is fast, the responses vary in quality. Maybe that is why Google quickly announced version 1.5, that is poised to fix a lot of issues and be even faster. That, of which, we shall wait and see.
Claude, while more of the underdog LLM, has come out of hiding with a vengeance. My only complaint is that, while the responses are fast, the entire website gets very slow. Having said that, Version 3 (available in Sonnet and Opus) has some serious power, speed and responsiveness. With upload support for text and images, Claude is turning into my new favorite.
If it can create a bubble shooter game, I'll be hooked!
Oh, you want to see the game?
Note: This isn't perfect, but that is what 4 prompts will get you. On Desktop, you may need to actually click in the game area to give it focus and use the keyboard. There is also still an issue with other blocks being removed or falling when they shouldn't.
Comments