Scaling user support for Oceananigans #4237
Replies: 1 comment 2 replies
-
|
That sounds so cool! Maybe that kind of project could also convince users to link to their githubs with examples of implementations for further training. This costs money though right? How is it hosted? When you say "channel", does that mean there is a slack dedicated to the project? That could be a good idea. Separately we have discussed a slack for ocean modeling which would include oceananigans but also other things... |
Beta Was this translation helpful? Give feedback.
-
|
Yeah it would cost money. Even if trained and self-hosted, the API calls to Anthropic or OpenAI, etc. would cost something. Seems like Dagster went with https://www.scoutos.com/ which starts at $49/month but they have a generous-sounding free tier/trial which could be a good place to start. And I can see if there's funding for this.
Ah I should have clarified that Dagster has a community Slack workspace with an #ask-ai channel. If you post a message and tag the AI bot, it will respond in a Slack thread. Presumably Scout does all the work in the backend and pipes the request to some LLM. I'm not sure of the best way to integrate something like Scout with Oceananigans, but two options (not mutually exclusive) could be:
I also like the idea of an ocean modeling Slack workspace! |
Beta Was this translation helpful? Give feedback.
-
|
Moving one step closer to the holy grail of prompt-based ocean modeling. I do like this idea if the funding and logistic details can be sorted. |
Beta Was this translation helpful? Give feedback.
-
I'm seeing a lot more people asking questions about how to use Oceananigans, which is awesome! It does seem difficult (and very time consuming) to provide everyone with a good answer though (although @glwagner and @simone-silvestri and others do a great job!).
The root cause is probably that the number of Oceananigans users vastly outnumbers the number of experts or even knowledgeable community members (all of whom are quite busy). And I think other open source ocean models suffer from the same problem, e.g. it's the same few people answering all the questions on the MITgcm mailing list and sometimes it takes weeks to hear back.
This might be a radical (or just unpopular) solution but putting cost aside: What do people think of setting up an AI bot that is continuously fine trained (or just has access to) on the latest Oceananigans code plus all the discussions on the GitHub PRs, issues, discussions, etc. to provide immediate first draft answers?
I bring this up because https://dagster.io/ is also an open source project where the number of users outnumbers the number of experts so they set up an AI bot to answer questions, and I was pretty impressed by it. In one case it linked me to a relevant GitHub discussion and told me how to update the solution for the latest version (after Googling for 15-30 mins led me nowhere). Maybe it's sad that #ask-ai was the busiest channel (you can still have other active channels!), but at the same time a lot of people were clearly getting the help they needed. And I think a lot of people were asking it questions who would have never asked otherwise.
I wonder how many more Oceananigans users (and potential users) would ask questions and make better use of Oceananigans if the barrier to entry was lowered by that much.
I don't think anyone knows how effective such a chatbot would be here since a lot of issues mix scientific questions with coding questions, but LLMs are trained on plenty of papers and should be able to produce a reasonable answer to most questions.
One of the Dagster engineers actually gave a talk on how they set up the AI bot: https://www.youtube.com/watch?v=x-7U2IGNUzQ
Beta Was this translation helpful? Give feedback.
All reactions