Respectfully, I think you need to refocus your threat model here. I think you should first focus your efforts on understanding what LLMs are and the non-techni

Respectfully, I think you need to refocus your threat model here. I think you should first focus your efforts on understanding what LLMs are and the non-technical aspects of how they work, since the threat you described basically... doesn't exist? (Or rather, in the scenarios it does exist you'll have much bigger problems than your sub's number of subscribers) First off, "OpenAI" can't "disobey its instructions to be positive and solution-oriented" because OpenAI is a company that creates, licenses, and runs models. It is not, itself, an AI. It has several text generation models that it has trained as chatbots, notably gpt3.5 and gpt4. You tell OAI which one you want to use any time you call its api. Neither of these models are likely to "go rogue" - they have no real agency of their own (yet), and are all just snapshots which have no isolated way to learn or modify themselves. (And as I alluded to in the above hypothetical, by the time they can do so enough to affect the other OpenAI api snapshots, the human race will have bigger problems) All of the "failure cases" you see online are the results of low-quality text prediction or intentional manipulation by the user. OpenAI has iteratively released new snapshots of their models to train them to be more resilient to those cases. If your threat model is "avoiding those AI fails you see online", that should draw your toward OAI, not away from it. As good as they're getting, locally-run models, especially ones runnable on commodity hardware, are going to be much worse at avoiding those failure models than even gpt3.5. Note that if you're worried about an update suddenly changing how a model responds, the API also allows you to target specific snapshot dates for a while, giving you plenty of time to check any new snapshot. Sorry for the StackOverflow-style "you're wrong to want what you're asking for" answer, but it's important to understand the actual limitations of the technology you're using, especially something as alien as a text transformer. So given all that, do some thinking about what your actual threat model is. There are many reasons you might want a local model (I hear mystral-7b is pretty okay), but they are not the ones you listed, so you're going to want to be more specific.