Are Large Language Models the future of Operating Systems?

Watching a recent video (https://www.youtube.com/watch?v=zjkBMFhNj_g) by Andrej Karpathy, former director of AI at Tesla and Open AI team member on the nature of Large Language Models (LLMs), reinforced a key finding that I uncovered during my exploration of LLMs.

While a lot of people consider LLMs as a text completion tool, once you use them a lot more, you being to understand the potential for them not as a tool for writing text, but as a wider enabler of cognitive activity, or to put it another way, a computer Operating System (OS).

This idea is fascinating, the concept of LLMs acting as the “kernal process of an emerging operating system”. With the context window being RAM, embeddings being the file system, browser access being the equivalent of an ethernet connection and the peripheral devices being the likes of vision and audio.

In the AI Accelerator we explore the concept of context as being like the working memory of the LLM. Managing the context is similar to the idea of RAM and the idea of using interactions with LLMs as ‘programs’ rather than conversations.

In the early days of computing, software developers were forced to find imaginative ways to make the most of RAM, Super Mario Brothers was only around 32KB. This lead to some really interesting developments in managing RAM including lazy loading, and recycling objects. Which echos some of the highly powerful findings that we cover around context management and context blocks.

It will be interesting to see if other historic programming techniques have AI based similes, I have a suspicion they will. Have you discovered any?

Original diagram credit Andrej Karpathy.