In the case of LLMs, I really wish we would stop referring to models where the source code and/or weights were leaked as "open source". Likewise, for those where they were deliberately released but the license restricts them to "non-commercial use" or allows commercial use so long as the use "does not compete" with the company that released them.

Without getting too hung up on the semantics, what matters here is a free exchange of ideas, algorithms, and indeed some code. All this has massively accelerated progress in this space in the last two decades. Instead of sitting on their results, jealously guarding their secret sauce, etc. companies are sharing a lot of detail and are verifying each other's results by replicating them and improving on them. So OpenAI did chat gpt, Meta did llama, Google is doing Bard. Llama sort of got unleashed on the open source community and people are now starting to use that, improve it, benchmark against it with their own models, etc.

Open source is always important. You won't have a good developer pipeline without it. Other engineering disciplines should strive for something similar. Doesn't mean there cannot be closed source software of course.

At least for image synthesis, I believe the open models just plainly produce far better results by now since training is crowdsourced.

I honestly feel the opposite. I mean, the more OSS the better, but I don't think that LLMs being open source will do much to mitigate what I see as being the largest risks of LLMs.