The discussion centers on integrating OpenAI’s new o1-preview and o1-mini models into Discourse’s AI features. MarcP notes these models currently lack support for the ‘system’ role in API messages and do not support tools or streaming, which affects real-time response rendering in Discourse MarcP. sam confirms these limitations and mentions efforts to hack workarounds while OpenAI adds missing features. Access is currently restricted to Tier 5 users during beta Jagster, but MarcP and Aaron_Walsh clarify this is temporary and broader access is expected. Aaron_Walsh praises GPT-4o’s utility for forum scripting and design but warns of its potential to replace developer work. EricGT requests clarification on why streaming support matters — it enables real-time output like ChatGPT’s web UI. MachineScholar anticipates o1 models will remain expensive post-release and hopes for user usage limits. sam proposes using o1 as a specialized reasoning tool triggered by cheaper models when complex reasoning is needed. A basic implementation without tool support has been merged sam, with further development paused until the next team meetup.
Trying to contact the model returned this error: { "error": { "message": "Unsupported value: 'messages[0].role' does not support 'system' with this model.", "type": "invalid_request_error", "param": "messages[0].role", "code": "unsupported_value" } }
It seems these new models do not support a system role. o1-mini is 80% cheaper then the top model, might be useful for complex tasks.
These models aren’t available in the API for your account while we’re in this short beta period. (Developers on usage tier 5 will have access, but we’ll expand access to more tiers.) We’re continuing to improve o1 and we’ll be in touch as soon as it’s available to you in the API.
I have to admit 4o has been a lifesaver for me. With this being an additional 80% on top of the current, it will work a lot better than paying ChatGPT £18.99 per month.
I use this manually for adding scripts of CCS onto the forum and making my own designs. It is incredible what it can do.
But it is also dangerous for the developers that it can overcome their work.
While I understand the reason for adding or streaming many may not. It might be of benefit to those here to have a small explanation of why streaming was noted.
We can expect this model to remain quite expensive even well after full API release, so hopefully we can get some user usage limits for AI personas implemented
Discourse uses the streaming feature (printing the response as soon as it comes in, similar to ChatGPT webUI) to show the response in real-time. o1 models do not support this, so Discourse needs to change their logic to work with non-streaming models.
One interesting use for o1 might be as a custom tool, to do all the chatting with GPT-4 and cheaper models, then have the model reach out to o1 in cases where it is clear it needs some custom reasoning.
