At least I wouldn’t expect it to.
My reasoning is that such a kill-switch is mostly targeting non-technical people who don’t want AI/LLMs anywhere near them and certainly not on their personal computer. The moment you start running software in containers on your computer, you’re probably fine with individual snaps (at least as long as their clearly labeled).
^^ This.
I find it increasingly frustrating that people seemingly lost all capability to read text longer than a tweet-worth of characters. Even the OP without the further comments and clarification left no space for such a bad-faith read on the situation.
We’re (mostly) techies here, we shouldn’t behave like apes and go “Fire → Hurts, Fire → Bad”. We can engage in a lengthy polemic about the morality of using models trained on a stolen data, where I think I made my point of view clear, but saying that Ubuntu is going the way of Windows is silly to say the least.
Many will, and many will also argue that means it’s really not global, and many will say that’s malicous design by Canonical. There’s not win with a “global button”, it’s better to be honest about it and say it’s not possible to properly do it, and provide ways to disable what Canonical and Ubuntu Community will implement, which is what is clearly being stated that will be done with a very effective method (you can both stop services and/or uninstall the software)
If so different, why hasn’t there been any mention of a concerted effort to establish
what would be immutable rules that override any other self-evolved rules?
what are the humanistic paradigms, values and weightings that will form the foundation of those immutable overriding rules?
I’ve seen no discussion of such by anyone who is actually doing “AI” testing or implementations?
Wouldn’t that be a good place to start?
It’s worth a look, because there’s quite a lot of it, published openly.
Anthropic released Claude’s constitution under CC0 earlier this year. It covers exactly the distinction you’re drawing between immutable rules and overridable defaults, including the reasoning behind each. OpenAI’s Model Spec does much the same job in a different style. Both are public, both are reusable, both are the result of years of work by people working on it full-time.
Similarly, Google DeepMind publishes a Frontier Safety Framework, now v3, plus the FSF reports for each Gemini release. It’s more focused on capability thresholds and severe-risk mitigation than on day-to-day behaviour, so it complements rather than overlaps the OpenAI/Anthropic specs.
Microsoft’s Responsible AI Standard v2, full PDF linked from there, is the most “engineering process” of the lot; six principles operationalised into impact assessments, gating reviews, and concrete product requirements. Useful if you want to show what governance-side rules look like in practice.
Meta’s Llama Acceptable Use Policy are the closest analogues from the open-weights side. Less philosophical than Anthropic’s constitution, more “here’s what you can’t use it for”, but still a published, reusable artefact.
So the foundational work isn’t missing from the field.
Which makes it more useful, I think, to ask the concrete question: in the actual proposal here, what specifically do you think needs different rules than the ones those documents already lay out?
I feel that’s something the team can engage with, where “start from first principles” tends to stall.
In short, beyond Anthropic and OpenAI, Google DeepMind’s Frontier Safety Framework, Microsoft’s Responsible AI Standard, the NIST AI Risk Management Framework, and ISO/IEC 42001, all cover the same ground from different angles.
Open-source coding agents largely inherit these rules from the model layer rather than redefining them, which is itself an answer to “where is the foundational work?”. It’s upstream.
So lets take 3 possible AI-enhanced features:
Will I be forced to install 3 distinct local AI models - so that the features are nicely separated. This will eat up 3 times the disc space just to draw a line in the sand. Or is there a way to reuse one model for all 3 tasks?
It seems like we’ll want a big shared read-only datapool where the actual model weights live and then some tiny nonshared configuration for each individual task, where the model stores some history of what it did before in this domain. Ideally the configuration could stay around if you drop the model, because you need the disc-space and can later be resurrected once you install a newer model with better performance in less disc-space or once you upgrade to more RAM / a new hard-drive and download the latest and greatest model.
And I’m sure, advanced users will want to install multiple models in parallel - just not fixed to one task but tinker with the assignment, so they can compare models A, B and C on the same task, swapping them out dynamically.
So instead of a single kill-switch I more envision a broker GUI, where I can on one side see the available versus the installed models with their disc space requirements and on the others side I have a list of programs/features in the OS suggesting to be hooked up with any available AI. And that GUI should also handle if certain combinations are discouraged (e.g. model doesn’t read images) or reject to download models that e.g. require more GPU/CPU/RAM than the machine has.
If you read what Jon wrote above the question should be, are you willing to wait for a multi gigabyte download during system install to get a combined model or would you prefer a faster install with more fine grained model installation at the cost of some disk space….
I guess the decision around this can only be made further down the road during the implementation of the feature and in context of the limitations the installer brings along here…
We won’t be pulling models at install time either way. Those will be pulled after first boot.
Exact models tbc, but generally I imagine we’ll favour multi-modal models to parent you from needing loads of models unless you have very specific use cases for sub-agents and such.
can these models be prevented from installing or are they installed at first boot by default ?
best of luck Steve ..
@ogra
Sorry for triggering you on the word “install” - yes of course I will at some point have to download the multi-gigabyte model to use it - no way around it (other than sending my data to some cloud-service). And whether I download the model at installation of Ubuntu or later when activating the feature is just a minor difference.
My point is, if multiple AI-features can SHARE the same model and if you see a way how to CONFIGURE which model that is? In other words: I’m asking for an indirection layer where a noob user can overview how much disk-space to dedicate to all local AI and what functions are requesting it to evaluate if it’s worth the storage.
Sure, we can wait for the implementation and then start writing bug-reports or we discuss beforehand what features we want - so we don’t need to change architecture again once you find out. And my wish would be to easily swap my local AI-Models to see which gets the job done within a reasonable price of storage.
PS: Maybe I’m a little paranoid on the storage issue, because my Kubuntu has been nagging me on and off for more than a year now, that my ssd is running over 98% full and my attempts to delete data cannot catch up with more coming in all the time.
I’m pretty sure the team implementing it will take such things into account, why would we duplicate diskspace if it would be possible to re-use something 
but that will likely have to be a case by case decision depending on the respective model and tools using it …
The foundation of this work will be Inference Snaps (docs). This means each model will be downloaded only once, and tools speak to them over an OpenAI API-compatible endpoint.