Generative AI has emerged as a pivotal force in contemporary computing, with Microsoft leading the charge by seamlessly integrating its Copilot AI throughout its technology ecosystem, particularly within Windows 11. In an intriguing blend of nostalgia and innovation, Marc Andreessen, co-founder of Andreessen Horowitz and a prominent figure in the tech industry, recently disclosed that a small Llama AI model from Meta successfully operated on a Windows 98 PC equipped with a mere 128MB of RAM, as reported by Money Control. This revelation, while surprising, prompts a reconsideration of the capabilities of older technology.
Although Andreessen did not specify which Meta AI model was utilized for this experiment, it opens the door to numerous theories about the potential of generative AI in earlier computing eras. Notably, Meta’s Llama AI model is designed to be less resource-intensive than many larger models, yet it still boasts impressive and sophisticated functionalities. Andreessen’s assertion that running Llama AI on a 26-year-old Dell PC with Windows 98 could have enabled human-like interactions with computers decades ago is particularly thought-provoking. He remarked:
“All of those old PCs could literally have been smart all this time. We could have been talking to our computers for 30 years now.”
While generative AI is a relatively recent development in the tech landscape, companies like OpenAI and Microsoft have made significant strides, with AI-powered offerings like Copilot becoming integral to a new generation of Copilot+ PCs. The investment of time and resources by AI-centric companies underscores their commitment to advancing artificial intelligence. However, Andreessen suggests that the trajectory of AI might have been markedly different had key players capitalized on the opportunities presented during the AI boom of the 1980s:
“A lot of smart people in the 80s thought all this was going to happen then.”
An early AI Era if Microsoft had dared
Andreessen’s insights appear to draw from a previous experiment conducted by Exo Labs, which successfully ran a modified version of Meta’s Llama 2 on a Pentium II-based Windows 98 PC, as reported by Tom’s Hardware. This achievement was no small feat; Exo Labs faced numerous challenges, including sourcing compatible PS/2 peripherals and devising a method to transfer necessary files to the antiquated device, given that USB technology was not yet prevalent.
The endeavor also required the arduous task of compiling modern code for an outdated operating system. The team ultimately opted for Borland C++ 5.02, a compiler and integrated development environment that was itself 26 years old and compatible with Windows 98. However, compatibility issues with C++ necessitated a switch to an older version of the C programming language, which involved a more labor-intensive process of managing functions and variables, resulting in slower progress.
In addition to software hurdles, hardware limitations posed significant challenges. Many Pentium II machines were equipped with only 128MB of RAM, which made it difficult to run Meta’s Llama 2, a model that typically requires far more memory due to its support for 70 billion parameters. Consequently, the team developed a streamlined iteration of the AI model that could function on Windows 98 with reduced hardware demands, a project that can be explored further on GitHub.
This experiment with Meta’s compact AI model illustrates that even legacy PCs running Windows 98 had the potential to support generative AI. While the current AI landscape is robust and rapidly evolving, one can only speculate how different it might have been had industry leaders embraced the possibilities of AI on personal computers in the 1980s, especially as Microsoft celebrates its 50th anniversary this week.
