What You Need to Know If You're Building an AI App – LLM as a New Operating System

Large Language Models (LLMs) such as ChatGPT are akin to "new apps." A concept currently trending on Twitter suggests that they are increasingly resembling a new generation of operating systems.

Today, LLMs use embeddings like Ada002, which transform existing data into encoded vectors. This process is very similar to how existing file systems operate, reading data and transforming it into a list of numbers. This list then acts as an index for LLMs.

The current context of dialogue is quite similar to RAM ("fast memory"). More and more LLMs are now providing access to "slow memory."

Andrej Karpathy has posted a perfect image describing this concept.

Let's combine the innovations created during the boom of personal computing with the trend of the LLM as an OS.

Basic Apps: These are what we see in the "App Store" for ChatGPT. More models are needed because the technology is still developing. We need LLM OS equivalents for word processing software like Microsoft Word, spreadsheets like Excel, presentation software like PowerPoint, and database management.

Enterprise Cloud File System for LLMs (akin to Dropbox for Prompts): As corporate clients increase their use of LLMs, they will likely need to share more "files," such as prompts, model configurations, and dialogue states.

I/O Processors: In the PC era, we had applications like Windows Media Player that enabled PCs to play audio and video files.

Web Browsers: Startups like Perplexity.AI are emerging because of this trend!

Antivirus Software: Privacy issues will be a concern with the widespread use of LLMs.

Communication/Networking Protocols: Protocols similar to emails or HTTP will be needed to facilitate interaction between LLMs and humans.

Utility Software: Tools like disk cleanup tools (or "LLM Memory cleanup") will be necessary.