Old-computer-with-tetris
Back to the future. You can now play Tetris using a 200+ billion parameter generative AI model.

OpenAI "o1" model can't really create video games

Tetris controls

Use the arrow keys to move the tetrominoes. Press 'Q' and 'W' to rotate the tetrominoes.

Score: 0

But it can program Tetris

Brod Justice - 13th Sept 2024

OpenAI unveiled their latest model, "Strawberry," yesterday, and while it was accompanied by the usual silly hype, there were also notable steps forward. The press video shows an OpenAI employee creating a video game from scratch by simply prompting OpenAI's new chat model "o1" (aka Strawberry) to code a game called "squirrel finder", a very simple game similar to snake

I recklessly bet my colleague Ryan McClure that "o1" would not even be able to create a Tetris game. It seemed like a safe bet as all previous AI LLM's have failed that test. Turns out that I was, sort-of, wrong. The Tetris game on this page was generated by "o1". It might not be the most elegant implementation of Tetris, but it clearly works.

As any game developer will know, this is a long, long way from being a state-of-the-art video game, so claims that OpenAI is exaggerating again are fair. However, OpenAI's "o1" model does seem to show significant progress in code generation, mathematics and reasoning. Some of these improvements come at the expense of a slight worsening of it's natural language writing skills (see image below from OpenAI's research paper).

The decline in performance in language skills suggests that, for many applications, Mistral and Anthropic are going to remain the more cost-effective and efficient models. However, it's worth noting that if other AI models follow the path taken by OpenAI and their 'o1' model, it could have significant implications for NVIDIA. The 'o1' model is more expensive to run due to its higher GPU usage during inference. If this approach is replicated by other model builders, it will in turn lead to more stable and long-term demand for NVIDIA GPUs.

By the way, we have already tested "o1" with our own AI search product and it does not provide any advantages, rather it makes responses slower and increases running costs.

What next for AI LLMs?

The progress of AI LLM models appears slowing in many ways. OpenAI's "o1" suggests that future models are going to be more specialised, focused on specific tasks. We believe that the next big leap in AI will come when the smaller models move to edge devices. We believe this will have a greater impact on the general public's interactions with AI than OpenAI's current trajectory (more on this soon). Meanwhile, enjoy a free game of Tetris, at this rate we may even get that 1970's classic Asteroids within the next 24 months.

Bonus Tetris tips for those that read this far

The AI generated game itself was cumbersome to play as the arrow down also scrolled the viewport. This has been modified so if you leave your mouse cursor over the Tetris field then the page will no longer scroll when down-arrow is pressed. We made a few other modifications. Tip for mobile device users: don't bother even trying to play.

PS:

I've been asked for the GPT tetris code prompt. It was simply "Write the tetris game in javascript to display in a browser". That prompt delivered a whole webpage including the javascript. There was a bug in the code, but after I pointed this out in the GPT chat "o4-mini" produced the bug-free version which is the basis for what you see on this page.

OpenAI "o1"
OpenAI's "o1" model shows a significant improvement in logic and reasoning, but no improvement in natural language text creation.