AI Model Fragmentation

Everyone's talking about "AGI models" like Astra and Fable, but what's more interesting is how businesses' model stacks have fragmented.

First, large companies like Uber now use model routing to send requests to fleets of models, many open source. Model routers mean that the user often doesn't even know what model serviced the request.

Second, most people do not feel brand loyalty to a lab. They will switch on a dime, just like engineers learned to switch cloud providers 10 years ago.

Third, engineers have discovered that you don't need the smartest model as a coding agent. The "best" models often talk too much, take too long, and overcomplicate. Coding tasks are increasingly distributed to dumber models, driven by individual taste.

Fourth, the launch of Jev laid bare that the LLM is not the end of AI integrations and tooling. Jev seems clearly superior to LLMs for many tasks. Jev can easily replace tasks that currently use a hacked-together call to Luna or Haiku. LLMs + Jevs working together seems like the natural outcome.

The story of "one model to rule them all" seems less and less likely.

The future possibly looks quite like the past: similar to how cloud compute feels like a commodity with low loyalty, intense price competition, and lots of people reducing switching costs by wrapping the ecosystem.