The idea is that without copyright there is no benefit in keeping the source code secret. It justs makes it a little bit harder for the customer to recover it. The GPL is ultimately intended to be customer-friendly, therefore sharing source code together with the build system is preferred.
It's still a form of proprietary lock-in similar to not offering data export, but it's not an unsurmountable hurdle anymore.
> The idea is that without copyright there is no benefit in keeping the source code secret.
Why not? You can decompile a cookie to try to see what method was used to make it, but Nabisco or whoever still treats that method and the ingredient ratios as trade secrets. In fact, the whole idea of trade secrets are things that are kept secret because IP protections don't cover them.
That's not really comparable; by reverse-engineering a cookie one would have to fight against entropy. The issue with software is that it's magnitudes easier to figure out any trade secrets inside of it. Reverse-engineering LLMs is much more akin to the cookie example.
It's still a form of proprietary lock-in similar to not offering data export, but it's not an unsurmountable hurdle anymore.