Blasting AI into the Past: Modders Get Llama AI Working on an Old Windows 98 PC

Though it’s not the most dynamic conversationalist, it’s still a fascinating achievement.

Exo Labs, an initiative aimed at democratizing AI access, has unveiled their latest project: a modified version of Meta’s Llama 2 running on a Windows 98 Pentium II machine. This version, while not the newest model, has generated interest due to its unique execution on such outdated hardware.

Thanks to their innovative approach, they’re showcasing that running AI doesn’t necessarily require modern power-hungry servers. The creative process was not without its challenges; they needed to adjust the hardware significantly and figure out how to transfer files to this ancient machine.

The team faced difficulties compiling modern code for older systems and resorted to using a 26-year-old development environment—Borland C++ 5.02. In overcoming issues with compatibility, they had to use an older version of C, which added complexity to the programming process.

In a fascinating twist, Exo Labs developed a lightweight version of Llama that operates with 1 billion parameters and manages to produce 0.0093 tokens per second—not blazing fast, but a notable success given the constraints.

Interested developers can tinker with this version on GitHub. The achievement of this project serves as a reminder of what can be accomplished when imagination meets retro technology.