If you were to use a co-op for AI services, would you prioritize privacy or renewable energy first?
20
Replies
2
Boosts
This comment is unavailable — it may have been deleted or is not visible to you.
@evan it should be a false dichotomy, but in terms of bootstrapping with existing services, from what I can find it is one or the other, at least for a maximalist approach.
@evan Yeah. It reminds me of the depressing fact that ebikes can be more efficient than human-powered bikes because humans are sorta inefficient. It's why we're so lovable:)
@ntnsndr Renewable energy but I'd ask for credible exit of my memory.
@ntnsndr The ranking of the privacy constraints depends on what data I would be entrusting to the service. If the co-op were providing speech translations without privacy guarantees, that would probably be fine for public meetings, but not for private ones. If it’s reviewing my code, that’s fine for my MIT-licensed hobby project, but not for the gpl project if the service is feeding inference data into fine tuning data.
@ntnsndr but for me personally, both are nice, as part of a overall 'better' offering. I might pick energy as more tangible if i had to select one over the other
@ntnsndr hmm if i was planning marketing, i would probably have different messaging for different channels... and i would probably talk about 'control over your information' rather than the abstract idea of privacy, if i was aiming at a mainstream audience...
@ntnsndr privacy but in reverse. like, id be interested in a coop that does explicit data sharing within the coop for various aspects. something like building a collective dataset of the members for a rag or lora or something
like, neoclouds that offer zdr or full green energy exist, and those features, while nice, do not really tap into what makes a coop interesting
@laurenshof @ntnsndr I think this is the right angle. With the use cases where the value created is more than the unsubsidized costs of running the models, the scope is still relatively narrow—transcription alone, or summarization to a particular template, or following a particular workflow. Shared (fuzzed) data could make that more reliable while respecting data privacy and consent.
@serapio @laurenshof yes, I love this! But I think it would only make sense in a purpose-constrained context (like Transkribus or Land O'Lakes) rather than general-purpose.
@ntnsndr @serapio yeah specific purpose makes the most obvious sense, but i think general-purpose would also be interesting actually. like, i think i can push llms to get interesting and useful noncode outputs, and i know other people who can too. a finetune based on our collective transcripts could also potentially produce a very interesting model
@ntnsndr
1) I'd consider privacy a more significant concern.
2) After that, I'd consider the openness of the model/s involved. Are they open parameter?
3) Renewables would be a concern, Artificial Intelligence is a worthy target for energy usage. Chinese models are worth considering. China's electricity production is over twice the US, and the renewable component is large. And while China relies on coal, the per capita usage is more efficient, and the long-term investment in renewables is high.
@Marc yeah, I hope it goes without saying that this would involve deploying open-weights models.
I also lean slightly toward privacy, thinking that switching server infrastructure later will be easier than switching the core software running models, thought that may be wrong. Fortunately there is tremendous modularity in these systems right now.
@ntnsndr renewable personally but I’d imagine privacy would be more desirable/marketable?
