One argument is that those who make LLMs have a high amount of control their output. It’s giving a lot of control of open source to the whims of those companies to become reliant on LLMs

Think about the potential for providers to add plausibly deniable backdoors, general lower quality code, etc. when they think you are working on a project they care about. The precedent is there with Anthropic publicly stating that Fable was built to intentional sabotage quality when they think you are working on AI research related tasks or distillation. That kind of stuff can be subtly baked into open models as well