publicai.co and general position towards LLMs

Great point about streaming platforms. E-mail spam too is a skewed user of resources. But all needn't be opposed in one go, right? Picking on what isn't established yet, early on, is good.

There should be a concern on "points of no return". We are supposed to cross-check AI output, but any online search already leads to many AI slop articles!

@Pirate Praveen can you list the ways in which you think LLMs are beneficial? Or to put it another way, what are the things using LLMs via Public AI can solve that cannot be done through avoiding LLMs altogether? Rather than looking at a technology and then choosing what we want to do, it is better to decide what we want to do and then see what technologies are most appropriate for achieving it.

Translations are one example you mentioned and also described in more detail in this blog post. (That said the models used for translation are different from chatbots, so Project Bergamot/Firefox Translate might be a more appropriate project to highlight for this point.)

Similarly, if we list out other uses then we can decide what to suggest for those uses (could be a "good" tool, a compromise tool, or a recommendation to avoid doing that thing altogether)

Regarding streaming platforms, I think the major ones are problematic in terms of resource usage as well as in other ways, and advocate for people to download media if they are going to watch it more than once. Peertube is an improvement due to utilising more local parts of the network (and so it BitTorrent for large media). Upon installing Jellyfin, I also realised how much computing power could be getting used up just in transcoding videos on-the-fly for transmission (maybe streaming platforms have these transcoded in advance and cached somewhere).

There is also the issue of streaming platforms acting as brokers and squeezing producers (and eventually customers) of funds, similar to Ola and Amazon.

Maybe we should issue a statement on streaming platforms as well 😉

Thanks @Pirate Praveen and @kishy. Inviting everyone else who uses LLMs, diffusion models, or other machine-learning based tools that are currently being sold under the label of "AI" to share what they use such tools for and how they find it useful. That includes observations of other peoples' usage of such tools, but try to keep it personal, i.e. "I have seen a person do this" rather than "People find it useful to do this".

(We can go into cost/benefit analysis etc. later; let's first get a list of use cases)

This paper covers some background https://publicai.co/open-source-win.pdf

"In this paper, we argue that without structural intervention from public institutions, current open source efforts in AI will not democratize access to AI nor provision public goods (in the technical sense of non-rivalrous and non-excludable goods) as comparable open source efforts have done in other categories. This will hurt the machine learning research ommunity. It will hurt startups. It will also undermine the strategic interests of large firms promoting open weight and open source AI. To move forward, we need to build broader public AI ecosystems that ensure open source AI is accessible, trustworthy, and competitive with closed source alternatives."

If we limit ourselves to just the software part without considering the rest of the requirements, we would be making a big mistake, like https://en.wikipedia.org/wiki/Streetlight_effect (we are used to software solving problems and we only need to look at software to solve every problem out there). This also means we cannot simply/blindly apply principles that worked for a software only solution to every problem.

I had forgotten about the NextCloud Ethical AI Rating but came across it again just now. I think this is a good framework to start with for evaluating machine learning related solutions (including LLMs).

Since there are degrees of rating rather than a binary acceptance/rejection, it also accounts for the fact that solutions may not currently be available that reach our ideal standard, but that the same time we have a way to choose the better one among several bad options that are already available. For example Public AI may not have a top rating but it would have a better rating than Microsoft ChatGPT. So, we can encourage people to choose the better option without compromising on campaigning for what would actually be the ideal solution.

The highest rating level under NextCloud's scheme is when the software for inferencing and training is open source, the trained model is freely available for self-hosting, and the training data is available and free to use. This almost corresponds to the definition I proposed earlier where all the code and training data has to adhere to the Four Freedoms. (Under my definition, the training data has to also be collated/distributed as part of the "source": I am not sure if NextCloud requires this too, but if not it could be added as an additional minimal level on top. In practice I am okay if I have to collate the data manually if it can be done in a reasonable number of steps).

Another important aspect I realised when reviewing NextCloud's approach to machine learning is that not everything is about LLMs. Many tasks such as face detection and classification of objects within specific contexts can be achieved using much smaller models, which would be more efficient due to being optimised to the task at hand. Better still, there are already models available for these tasks that reach the highest rating level. I have not investigated further but it is possible that models are available for other uses you mentioned such as translation as well.

I have heard that Big Tech is also pushing for LLMs rather than smaller models because making bigger, more resource-intensive things is more profitable than smaller, more efficient things. I don't have sources for this so it is somewhat based on hearsay but I have heard discussions like this in academic institutes like CMI for example. If someone can help fact-check this it would be good 🙃

In conclusion:

1. I think we can endorse the Ethical AI Framework developed my NextCloud (possibly adding an additional rating level of training data being distributed as part of the sources, if not already included in that definition)

2. I think it would be beneficial to identify and highlight some purpose-built models in our statement on "AI" so that people are aware it is not a question of "LLMs or nothing".

3. At the same time we can rate large language models (LLMs) according to the framework,, if we think they are necessary, and support the ones with a higher rating while pushing for initiatives to reach the maximum possible rating.

@Badri Sunderarajan I think we start rating services that provide Open weight models. I suggest we start with these,

1. publicai.co/chat

2. thaura.ai

3. euria.infomaniak.com

4. chat.qwen.ai

5. sealion ai (todo: add link)

Initial criteria,

1. Open Weight model - this is minimum criteria

2. Renewable energy - gets more points

3. Not using air conditioning / water cooling - lots of points

4. Reusing generated heat - lots of points

5. Ethical training data - only apertus matches this and it may be optional model in a service so points may be added optionally.

6. Reproduceable training data - recipe to build training data - only apertus matches it.

7. Funding models - public / govt funding or user funding

8. Philosophy / political position - thaura constitution is good

9. Tracking - use without login - public ai and qwen provides it

10. Not using user input for learning

11. How long the user data / context is stored.

Some providers offering apertus is listed at https://www.apertus-ai.org/pages/get-started/

https://www.apertus-ai.org/pages/charter/ also a good reference document for good qualities a model must have.

"Particular attention has been paid to data integrity and ethical standards: the training corpus builds only on data which is publicly available. It is filtered to respect machine-readable opt-out requests from websites, even retroactively, and to remove personal data, and other undesired content before training begins." https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html This addresses one of the most common concerns about training data.

"the entire development process, including its architecture, model weights, training data and methods, is openly accessible and fully documented."