Adopt the Open edX AI policy?

Discussions - Public
1

An AI contribution policy was just announced for Open edX projects:

I had a read through and it feels pretty solid and well needed.

Could we consider adopting this policy in OpenCraft too, across all our internal and open source work?

cc @braden

(not sure where to log time on this discussion; I’ve used BB-3149 so far since it relates to sprint planning)

1 Like
2

Yeah, aligning our own policy seems like a good idea.

I think the Open edX policy is good, but I would suggest two changes:

  • Allow use of additional models, including open source models in particular, as @tecoholic mentioned on their forum post. We want to encourage open source approaches in general.
  • Never use cloud-based LLMs on a repository that contains secrets, like infrastructure repos. This can result in secrets being used for training and potentially recoverable by other LLM users.
2 Likes
3

@braden thanks, yep your suggestions make a lot of sense for applying to our use cases.

Let me open a PR to the handbook. :slight_smile:

4

@braden cc @tikr is there an account/epic with budget that would be appropriate to log this time on?

I went ahead and got the ball rolling on this, by opening a PR to the handbook: opencraft/documentation/public!663. :slight_smile:

But since introducing something like this deserves team-wide input, it may be worth creating a dedicated ticket and scheduling a review period soon? What do you think? :slight_smile:

5

@samuel Since we work upstream a lot and adopting this policy will shape/inform/influence how we produce the code, documentation, etc. that we contribute, I’d suggest using the Contributions account and CC epic for any tickets that you create for this work.

CC @braden @gabor

1 Like
6

Thanks @tikr , this makes sense. :slight_smile:

I’ve created MNG-4953 and linked the current discussion and draft PR to it.

Do we have budget and interest to continue here? I’m thinking the general flow for something like this might be some review rounds with a small number of people, with at least @braden, to keep the initial iterating tight. Then ask for a team-wide review. What do you think?

7

Just a note here-- allowing more models might cause us a conflict when contributing code upstream, since we wouldn’t have used the set they specified. I agree the limitation feels arbitrary, though, and not in the spirit of Open Source.

1 Like
8

I have been experimenting with Opencode-go subscription (specifically deepseek-v4-flash model) along with chatgpt where the deepseek model acts as a worker that takes instruction from GPT 5.5/5.4/5.3-codex . It has helped a lot in terms of efficiently using my 20$ chatgpt subscription. So I agree with this stance.

9

@samuel

:+1:

That sounds like a good approach to me.

As for the budget question:

  • Deathstar/MNG doesn’t really have any hours to spare at the moment (due to budget usage exceeding the 4/6 month target of 66% at the end of April).
  • Bebop does have some unused hours from Jan-Apr on the Contributions account, so from the perspective of account-specific budgets it would make sense to move MNG-4953 to Bebop.
  • Bebop was off-target regarding sustainability in Q1, so if you want to move forward with this work now, timeboxing next steps to keep total time spent in check would be a good option.
1 Like
10

@navin @tecoholic Would you be willing to submit the additional models you use for inclusion in the upstream policy? I asked about this at the town hall at the conference, and @sarina reiterated to send an email to submit specific additional models, so we should do so:

If you are interested in proposing an AI tool for review, you may contact oscm@axim.org, as described in the AI Contribution Policy.

This seems to be a liability issue, @feanil mentioned that the currently approved ones have guarantees in case someone sues over the use of content produced by their LLMs. We probably need to find a solution to be able to include open source LLMs outputs, as I agree that seems wrong to only allow the proprietary cloud ones in an open source project - but that’s what we can discuss there. Put me & help@ in CC if you write.

11

@antoviaque I am not interested TBH. To explain my reluctance - the models are almost like sophisticated text editors at this point. It’s like trying to limit development to VS Code and Eclipse IDE because they are produced by trusted American companies. There is no point in keeping track of which LLM model is being used for code generation, as there is no way to prove it or otherwise.

The only place, I can think of, where liability issue could arise is - when models are hosted by a third-party and we are concerned about

  • the intellectual property or
  • credential leak/capture

When working on a Open Source code base like Open edX, the first one doesn’t apply and the second one is not really about the model, and more about the provider. This could happen when someone commits their .env file to Github as well. So, not really an LLM issue.

that the currently approved ones have guarantees in case someone sues over the use of content produced by their LLMs

This feels like AXIM has conflated the use of LLM models with use of LLM solutions. Product vs Service.

12

@tecoholic I agree with you about LLMs being more like a tool/editor, and I would support pushing for being more open as a community about LLM usage. But since we are part of a community, we need to follow its rules. If we don’t submit the additional models, the only ones we can use are the ones currently listed in the upstream policy. And to allow to be able to keep upstreaming, as mentioned in the MR about our own policy, we will need to align with their policy for any Open edX work, even if it is not (initially) meant for upstream.

1 Like
13

@antoviaque I understand. I can think of submitting the models as a short term solution until a more clearer policy emerges which is better aligned with real world usage.

I am planning on forwarding the following list to AXIM, along with the explanation of how their usage differ from something like Copilot, a service.

  • Google Gemma 4 open model
  • Qwen 3.6
  • DeepSeek V4 Pro & Flash

@team If anyone wants to add to this list kindly respond in this thread.

3 Likes
14

This is an interesting discussion that we would be willing to engage with if you wanted to post it over on the Open edX discussion forums.

15

@tecoholic Thanks, that sounds good :+1:

@sarina Sure, I can post a link in the thread where @tecoholic and @farhaan have already commented on, I’ll do so now. The current thread is just our internal conversation about it, which we hold in public, but doesn’t replace the community engagement.

@tecoholic When you email about submitting the additional models, could you also post them on the Open edX forum thread, along with your comments about LLMs being tools? I agree with @sarina that it would be good to discuss this more widely within the community.

16

Update: A copy of the email I sent to AXIM is shared openly in the Open edX Forum here:

2 Likes
17

Thanks @tikr :slight_smile:

I moved BB-10899 to Bebop, and I addressed some of the initial concerns on the draft PR.

I think we’ll be ready for team-wide review soon. When we have budget, perhaps we can give a short timeboxed subtask to everyone to read and review the policy? :slight_smile:

18

@samuel Since that ticket had time logged on it last month, when it was still a MNG ticket, some post-processing will be necessary.

You can either

  1. move the ticket back to MNG,
  2. clone it, and
  3. move the clone to BB

or

  1. clone the ticket,
  2. move the clone to MNG, and
  3. move last month’s work logs from the ticket to the clone.

To avoid this kind of thing, instead of moving tickets that had time logged on them in previous months we default to cloning them to the target cell that is supposed to handle remaining work.

Yep, that works. 30 mins per subtask should be doable :slightly_smiling_face:

19

Ah got it. Let me move that ticket back to MNG, and archive it. I’ve created BB-10904 to track further work - the 30 min reviews for everyone. :slight_smile:

20

@samuel Thanks! :+1: :slightly_smiling_face:

1 Like