The throwing of caution to the wind wild abandon integration into everything and anything of chatGPT is beyond a cause for immense concern.
There is a question that as an admin it is important to know being is there any part or component no matter how small that is essentially AI within the core discourse installation?
I’m assuming the answer is no but want to check.
Since the arrival of plugins that appear to sandbox AI outside of a core Discourse install. I also think it is. to know from a roadmap POV if there are ever paint o integrate any AI where there is currenlty none (if that is true).
If the answer is yes to one or more of these it would be great to have it confirmed
Yes a good point regarding definitions. I think the OP is clear in broad terms being a case of before & after ChatGPT (+similar) and what are the plans for Discourse, along the lines of at least:
using it to write core code
using it b integrated it into core
If anyone else wants to help define along these lines, but as you put it might be the best definition basically anything that is not “good old fashioned human-generated code, doing traditional procedural logic”
Can you provide me with a few (lets say 10, large) examples of open source projects that have a charter / public document that bans usage of GitHub copilot?
Yea so bot being a coder is why I asked this question so broadly. Now that I am aware of this, I guess then from the spelling and grammar checkers (for e.g. gramarly) to programme aids and assistants like copilot. I guess there is no clean and clear answer is there?
Just to add, Discourse (and many plugins) use extensive suites of functional tests and linting checks.
This massively mitigates code issues (try to deploy in RoR without checks …) and would cover AI genned code or Copilot supported code too. And that’s discounting the fact that Copilot might actually find issues you miss …
Core discipline wrt to tests on the platform is second to none and sets an example for the ecosystem.
It is also somewhat moot? I don’t think unsupervised AI code deployment is going to happen anytime soon … you wouldn’t even get as far as running tests (Have you tried to ask ChatGPT for a solution in EmberJS that is compatible with Discourse? The hallucinations are off the charts ). Maybe that will improve as models evolve.
Discourse has to keep up with the times or die imho, and that means exploring and adopting new tech where appropriate.
As the author of two “AI” plugins (both hand written! ) I recognise they obviously have to come with caveats and disclaimers and I can’t see that changing for a good while …
So within the frame of “discourse” to keep to the OP because you’re riding that wave and the next, of course it’s expansive issue, but:
Where things stand so far (if I am reading feedback correctly.)
There are no active AI components in core discourse
There are no inactive AI components or code within core Discourse
Some or all Discourse code may have come into existence with the indirect or direct application of AI tools within various dev platforms and IDE’s etc. avilabel to developers.
Adding any AI feature to core are currently not on the Roadmap
Plugins are the gateway to AI powered features that work with Discourse
AI is not a technical term … one person could interpret a heuristic as AI. The way we “score” posts giving weight to likes and reads and views can be considered AI. Akismet can be called “AI”. The list is a mile long.
As it stands the industry is “concerned” about AI safety in the context of large language models.
As it stands Discourse core and Discourse AI do not ship any large language models. Discourse AI integrates existing large language models via APIs which admins can enter (Open AI and Anthropic), in future we are likely to support even more models.
A large language model (LLM) is a language model consisting of a neural network with many parameters trained on large quantities of unlabeled text using deep learning techniques. These models can recognize, summarize, translate, predict, and generate text and other forms of data. LLMs have seen significant advancements in recent years, excelling in a variety of tasks and transforming the fields of artificial intelligence, science, and society as a whole [1][2].
Some examples of large language models include OpenAI’s GPT-3, which is trained to predict the next word in a sentence, much like a text message autocomplete feature [3].
Text via Google + GPT-4
I find a lot of this discussion somewhat perplexing, is there are concern we are going to “cancer” the Discourse source? Is there concern we are going to have a tight unconditional dependency on Open AI?
I definitely prefer the term “stochastic parrot” as coined by Emily Bender and Timnit Gebru- it’s definitely more accurate when referring to these LLMs.
My view, is if you are genuinely perplexed I would wager you’re not as actively cognisant or appraised or are empathic to “consequences” because there may be no apparent value in teasing this out, since the horse has already bolted and barn door is swinging wide open so everybody lets go!
You say cancer, more like quantum-weapon of mass [fill in the blank], however some will think that is all hyperbole. That’s cool. Everyone get’s to have an opinion, but I"m thinking overall and not specifically Discourse, what you do with the dev is up to the way that works so to clarify the point in relation to the OP:
The primary reason I am asking is as an admin of Discourse, where you have a duty to a user base, and while this might be a tiny segment in real market terms, but still your user base or portion of your user base are important, and may actually wish not to use a platform that engages any forms of this generation of AI when it comes to the content they generate, for lots of different reasons. The same way they might use adblockers, they don’t want to be part of this “thing” - when that question comes form the first user that asks its because there was no clearly indicated AI policy you have to also have some form of an answer that is truthful.
Therefore segregating the AI tools as plugins that are not core is a good approach, but if it came to it that the AI was core, say in a future update, it might leave room for hesitation or total rejection of the platform.
As long as this line remains clear and delineated then that might be all round satisfactory way forward and clear dev guidance and assurance might be important here too.
I also understand it’s hard to frame it to everyone’s total agreement, but here is the nub of the issue now, this is what happens when a question is asked for a real reason. If if the reason is founded or undaunted it will be the basis for users to make value decisions that convert into action.
If it’s simply a case that a simple statement that there is no ChatGPT or similar AI in the core of Discourse, with a caveat some AI is used in dev and creation as a tool, that may be as close to a satisfactory answer for the ordinary punter on a discourse forum.
Overall the issue of how AI is going to affect the web and by virtue the world has already been done and seen in terms of application and consequences, this phase of releasing these “toys” officially, i.e. giving a pandoras box with API’s n’all to the dev world and again by virtue the world is a ramp up, since before it was a very controlled experimental learning and relatively clandestine phase where this AI had been infused in a big part of the web for years without many ordinary people even realising it.
Another aspect are say Ads, I’d assume Google has been using AI of this ilk for ages in its ad serving and even analytics services, but at least for now users can deploy some blocking software and use VPN’s, do those tools use AI too?!
Maybe it boils down to how people perceive AI and maybe the easiest way to articulate that is some see it as sandboxed (tools), other see it as a fire in the very fabric of the reality matrix that needs to be quenched, either way it’s happening and it’s going to happen whatever it is but there are always consequences, and sometimes all you have is a choice of being part of creating those consequences or mitigating them, what we have with “industry” is they are IMHO being both.
If you are asking me to say “From now … till the the end of the universe can you commit that LLMs will never find their way into the core product” … the answer is no.
If you are asking me “Will I be able to opt out, were something like that to happen?” My answer is … probably … we tend to allow admins enormous amounts of freedom with site settings and configuration.
Now woudl I be asking you to say anything other than what is, as that would be unreasonable as you reply above indicates, who can say on an infinite timeline, sure, really it’s a endeavour to establish the current position Discourse core utilising of active AI as best possible.
Nor looking for any promises either.
My primary suggestion in advance of an answer remain the same.
Consider keeping active AI separate to core so owners/admins have a clear choice with no confusion. My preference is for this over a on/off option in admin, call me paranoid and old fashioned if you will, I’m thinking an air gap of sorts.
ChatGPT et al being interrogated into so many platforms and tools, it’s wild and in the broad terms it the implications are vast, so why woudl Discourse miss the boat (it didn’t the plugin is that event in the timeline of Discourse dev).
I’d regard the question as clarified as it’s going to be and hope that if there was a demonstrable change to the balance of play, a very clear communication would be made in advance or suitable time frame.
Have a forum with active AI deep in core just might be one skynet pivot too far for some admin/owners and/or user bases, and I think AI should be considered something even greater than a paradigm event that is meta+
I’d guess it’s not on the roadmap).
). Maybe that will improve as models evolve.
) I recognise they obviously have to come with caveats and disclaimers and I can’t see that changing for a good while …