SIAgentNews.com
← All news
DevelopingOpenAI

A 2.2 million parameter ChatGPT model runs on a Super NES

Developers have successfully squeezed a small language model onto a 1990s game console to demonstrate the limits of retro hardware.

· First reported by xda-developers.com

2.2Mnumber of parameters in the model

The short version

  • The model contains 2.2M parameters, which are the internal variables the AI uses to make decisions.
  • The SNES was originally designed for simple video games, not for complex language processing.
  • The AI responds to simple questions but can occasionally produce incorrect or blank entries.
  • Users can test the interactive demo directly in a web browser.
  • The project follows a trend of fitting AI models onto extremely limited hardware like floppy disks.

Why it matters

This project highlights the growing trend of model compression, showing that AI doesn't always require massive data centers. It proves that even very small models can perform basic tasks on hardware that predates the modern AI era.

Squeezing AI into retro hardware

A developer named Ben Edwards has achieved a technical feat by running a version of ChatGPT on a Super Nintendo Entertainment System (SNES). This hardware was designed in the early 1990s specifically for rendering simple video games, not for processing complex neural networks. The project demonstrates how far modern software can be backported to older technology.

The model used in this project is significantly smaller than the versions of ChatGPT typically used by the public today. It features 2.2M parameters, which are the numerical values the AI uses to weigh information and generate text. By reducing the size of the model, the developer made it possible to fit the software onto a system with very limited memory and processing power.

As reported by xda-developers.com, this follows a broader trend of squeezing artificial intelligence into unconventional environments. Just a few days ago, reports emerged about a person successfully loading a language model onto a floppy disk to run on an ancient CPU. This project pushes the boundaries of what is possible when modern code meets legacy hardware.

Performance and limitations

While the model is functional, it faces significant hardware constraints that affect the user experience. Because the SNES is not built for high-speed computation, the AI runs quite slowly. This means that when the system generates a response, each word takes some time to appear on the screen.

The accuracy of the model also varies depending on the complexity of the prompt. For simple questions, the AI provides reasonably correct descriptions. For example, when asked about Mario, it correctly identified him as a small plumber who performs jumps to rescue a princess and help out his friends, which is as accurate a description as you could ask for.

However, the system can sometimes go off the rails when faced with specific queries. When asked what Metroid is, the model listed four things, but two of those entries were completely blank. The other two entries it provided were "Great fuel" and "Good luck." These errors highlight the limitations of running such a small model on hardware that was never intended for this type of work.

The limits of backports

This project serves as a clear example of how far people can push backports. A backport is the process of taking modern software and making it compatible with older systems. By successfully running a language model on a console from decades ago, the developer shows that the limits of these systems are being tested in creative ways.

The fact that a 2.2M parameter model can answer simple questions on a SNES is a significant achievement for the hardware. It achieves a lot for a device that was only intended for rendering simple video games. It proves that with enough optimization, even very old components can participate in the modern AI landscape.

The project suggests that these types of experiments will likely only get better as time goes on. As techniques for model compression improve, we may see even more complex tasks performed on even older hardware. This trend of squeezing AI into small spaces is becoming a fascinating area of exploration for developers and enthusiasts alike.

Interactive demo and impact

For those interested in seeing the results for themselves, there is an interactive version available. You can head over to the ChatGPT for SNES website to play with the model directly in your browser. This allows users to experience the quirky results and the slow generation speed firsthand without needing to own the physical hardware.

The ability to run these models on such restricted platforms changes how we think about the requirements for artificial intelligence. It shows that while large models are powerful, smaller models can still provide utility in niche environments. This could eventually lead to more specialized AI tools that run locally on very low-power devices.

This experiment highlights the intersection of nostalgia and cutting-edge technology. By bringing a modern language model to a classic gaming console, the developer creates a bridge between two different eras of computing. It serves as a reminder of how far technology has come while showing that old hardware still has some life left in it.

What to watch

  • Whether more models can be squeezed onto floppy disks or other legacy storage.
  • How further improvements in model compression will affect performance on retro hardware.

Try it

Visit the ChatGPT for SNES website to try the interactive demo in your browser.

Sources and links

First reported by xda-developers.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news