“I think there is a market for about 5 computers.”
I am by no means an open-source maximalist, but I am coming to the belief that many of the Foundational Models (FMs) and infrastructure for what we currently refer to as Generative AI will arc towards open source projects and open core business models.
Not a year or two ago, the dominant point of view pushed by AI organizations like Open AI and Deep Mind was that building tooling in applied machine learning was high fixed cost and winner-take-all, so whomever could throw the most PhDs and compute at a problem would build insurmountable moats. As I’ve argued previously, these businesses may not be as strong as we currently think, and empirically we can see that many organizations have found ways to build LLMs small enough to run training & inference locally, like Stanford’s Alpaca model or Nomic’s gpt4all, creating a “stable diffusion” moment for text generation. I think it’s fairly evident that FMs will trend towards more efficient footprints and become economical to run locally (on-device for consumers or on-prem for enterprises). My claim therefore is that as models trend towards local deployment, the dominant models and middleware to interface with those models will be open source.
If you disagree with the above, let me know your reasons. If you’re building open source tooling for FMs, tooling for Personal Models (PMs), epistemic stack tracing software, or anything related to the above, I’d love to hear from you.