I think that, even commercially, they haven't gone far enough toward the non-profit's mission. We actually see Meta's Llama's, Databricks MosaicML, HuggingFace, and the open-source community doing what we'd imagine OpenAI's mission to be.
Anyone taking action against their non-profit should point to how Meta democratized strong A.I. models while OpenAI was hoarding theirs. They might point to services like Mosaic making it easy to make new models with pre-training or update models with continuing pretraining. They could point to how HuggingFace made it easier to serve, remix, and distribute models. Then, ask why OpenAI isn't doing these things. (The answer will be the for-profit motive with investor agreements, not a non-profit reason.)
Back when I was their customer, I wanted more than anything for them to license out GPT3-176B-Davinci and GPT4 for internal use by customers. That's because a lot of research and 3rd-party tooling had to use, build on, or compare against those models. Letting people pay for that more like buying copies of Windows instead of per token training would dramatically boost effectiveness. I envisioned a Costco-like model tied to the size or nature of the buyer to bring in lots of profit. Then, the models themselves being low-cost. (Or they can just sell them profitably with income-based discounts.)
Also, to provide a service that helps people continue their pretraining and/or fine-tune them on the cheap. OpenAI's experts could tell them the hyperparameters, proper data mix, etc for their internal models or improvements on licensed models from OpenAI. Make it low or no cost for research groups if they let OpenAI use the improvements commercially. All groups building A.I. engines, either inference or hardware accelerators, get the models for free to help accelerate them efficiently.
Also, a paid service for synthetic, data generation to train smaller models with GPT4 outputs. People were already doing this but it was against the EULA. Third parties were emerging selling curated collections of synthetic data for all kinds of purposes. OpenAI could offer those things. Everybody's models get better as they do.
Personally, I also wanted small, strong models made from a mix of permissive and licensed data that we knew were 100% legal to use. The FairlyTrained community is doing that with one LLM for lawyers, KL3M, claiming training on 350B tokens with no infringement. There's all kinds of uses for a 30B-70B LLM trained on lawful data. Like Project Gutenberg, if it's 100% legal and copyable, then that could also make a model great for reproducible research on topics such as optimizers and mechanistic interpretability.
We've also seen more alignment training of models for less bias, improved safety, and so on. Since the beginning, these models have a morality that shows strong, Progressive, Western, and atheist biases. They're made in the moral image of their corporate creators. Regardless of your views, I hope you agree that all strong A.I. in the world shouldn't have morals dictated by a handful of companies in one, political group. I'd like to see them supply paid alignment which (a) has a neutral baseline whose morals most groups agree on, (b) optional add-ons representing specific moral goals, and (c) the ability for users to edit it to customize alignment to their worldview for their licensed models.
So, OpenAI has a lot of commercial opportunities right now that would advance their mission. Their better technology with in-house expertise are an advantage. They might actually exceed the positives I've cited of Meta, Databricks, and FairlyTrained. I think whoever has power in this situation should push them to do more things like I outlined in parallel with their for-profit's, increasing, commercial efforts.