LLMs (Large language models) have some incredible parallels with OSes (operating systems). Andrej Karpathy recently published a fantastic primer on LLMs. Highly recommend the watch: https://www.youtube.com/watch?v=zjkBMFhNj_g

An LLM OS diagram looks quite similar to a traditional OS diagram. For example, you can think of the context window as RAM. It has the equivalent of a disk, Internet connectivity, and I/O (audio, video, text, etc), which are all aspects of a traditional OS.

Even more interesting, this comparison extends into market dynamics as well. Windows and macOS led the market for consumer OS, with a smattering of open source companies and products to follow. And similarly, OpenAI’s GPT-4 and Anthropic’s Claude-2 are shaping up to lead the production of consumer LLMs with a large open source community starting to follow. LLMs seem to have a similar distribution of proprietary and open source models.

Additionally, just like OSes had security challenges (think of iOS jailbreaks), LLMs also have a newfound security question. There are lots of attack vectors that allow malicious actors to jailbreak LLMs: prompt injections, “sleeper agent” via keywords in the training data, multimodal adversarial inputs (an image can have a hidden message etc).

Of course, it’s still very early in the space, although it’s moving rapidly. Always interesting to think about the second and third order effects of the work being done here.