Even if we take Chromium, compiling it is still within the logistical capabilities of any individual with a modest computer, and it's still more reasonable to spend a day waiting for the compilation then to try to modify the binary directly.
The same cannot be said for LLM weights, as evidenced by the fact that even the enormous megacorps that put these things out tend to follow up by fine tuning the weights (using different training data) rather than retraining from scratch.
