Imagine a time when technology wasn’t dominated by powerful servers and hundreds of gigabytes of RAM, but by a simple Pentium II running Windows 98. It’s in this context that the improbable happens: a group of researchers manages to resurrect this ancient hardware by integrating a language model inspired by LLaMA 2. Thanks to BitNet’s innovative architecture, they run a massive model while circumventing the limitations of a bygone era, paving the way for more accessible AI for everyone. A group of researchers recently accomplished the improbable feat of running an advanced language model inspired by LLaMA 2 on a 25-year-old computer. This computer, equipped with a 350 MHz Pentium II processor and 128 MB of RAM, is still running Windows 98. Thanks to BitNet’s ingenious architecture and meticulous computer tweaks, this retro-tech project demonstrates that it is possible to make artificial intelligence accessible on older and mainstream machines. Reviving a Computer Classic How can you run a modern language model on a nearly obsolete machine? This is the question that prompted researchers to focus on a PC equipped with a Pentium II processor. What began as a tentative idea became a technical challenge, aiming to leverage an operating system such as Windows 98 to host artificial intelligence designed for the modern era. The Secret of BitNet Architecture The BitNet architecture is the key to the project’s success. Unlike conventional models, which are often memory- and power-intensive, BitNet is based on ternary weights (0, -1, 1). This strategy significantly reduces the model size, a necessary adjustment to adapt to outdated hardware. Thus, a model with 7 billion parameters is compressed into a volume of 1.38 GB of storage. Overcoming Technical LimitationsThe return to legacy hardware was not without its challenges. With modern USB interfaces out of the way, the team reverted to traditional PS/2 peripherals and FTP transfers. To compile the code, the 1998 Borland C++ 5.02 compiler was used to adapt the llama2.c file. This was a complex task, requiring adjustments such as manually manipulating system clocks and substituting modern types for their predecessors. A Step Towards Accessible AIWhile the implementation on a 1998 PC is obviously impressive, this project has a broader ambition. The goal is to make artificial intelligence accessible, beyond remote servers and resource-intensive solutions. With BitNet, this means the ability to run models locally on more modest machines.
Future Perspectives for AI New perspectives are emerging thanks to this approach. It opens up new possibilities for integrating AI into old phones and forgotten computers, also enabling applications in embedded devices without cloud connectivity. EXO plans to provide open-source tools to encourage experimentation on legacy hardware and is tackling specialized fields such as protein modeling.
