When ChatGPT launched a little over two-and-a-half years ago, many people were impressed by the capabilities, but thought that LLMs would only slowly be introduced into sensitive contexts as they prove themselves beyond even the highest standards of doubt.
Yesterday I went to a talk by Joshua Tan of the Public AI Inference Utility (https://publicai.co/stories/utility), which is trying to build exactly the supra-national consortium you mentioned, focusing on (iirc) Switzerland, Spain, Germany, UK, France and Canada.
The idea is to be the front-end interface and inference provider for an array of sovereign models developed by the individual countries through their national labs. They already seem to have a Swiss model and a Singaporean model.
From what he said, it does sound like the bottleneck is political will at the highest level - having this was a key part of successfully setting up the Airbus consortium.
Great find - I also heard about the Public AI initiative, and the Airbus alliance is a great historical prior to how globally successful Europe can be if united!
Good insight, but I think we must consider not only "digital sovereignty" in terms of owning LLM models (which is often unfeasible for mid to small nations to train from scratch) but through "digital autonomy" which means having supra-systems ("agentic") that are transparent, can use a multitude of different models and in which it's easy to swap out LLM providers at will (or benchmark against each other e.g. for bias evals).
Because the future of LLM-based AI (given the diminishing return of what was considered "scaling laws", see GTP-5) means that most development will not happen vertically (model size) but horizontal through inference time. This is where every nation can use the agentic system they want, and which providers them with the most objective observability and transparency for the public.
For inference-time systems (e.g. "deep thinking") to be effective they need domain/task specific data. And that's where the crux is, as ownership and lineage of our own data is what gives us ultimately our autonomy. Meanwhile the big LLM providers try to silo users by encapsulating their data on their own platform, to become the era's dominat platform (just as in the previous computing shifts MS did with business tools/PC, Zuckerber with "social") by keeping their data hostage on their proprietary formats and datacenters.
Yesterday I went to a talk by Joshua Tan of the Public AI Inference Utility (https://publicai.co/stories/utility), which is trying to build exactly the supra-national consortium you mentioned, focusing on (iirc) Switzerland, Spain, Germany, UK, France and Canada.
The idea is to be the front-end interface and inference provider for an array of sovereign models developed by the individual countries through their national labs. They already seem to have a Swiss model and a Singaporean model.
From what he said, it does sound like the bottleneck is political will at the highest level - having this was a key part of successfully setting up the Airbus consortium.
Great find - I also heard about the Public AI initiative, and the Airbus alliance is a great historical prior to how globally successful Europe can be if united!
Good insight, but I think we must consider not only "digital sovereignty" in terms of owning LLM models (which is often unfeasible for mid to small nations to train from scratch) but through "digital autonomy" which means having supra-systems ("agentic") that are transparent, can use a multitude of different models and in which it's easy to swap out LLM providers at will (or benchmark against each other e.g. for bias evals).
Because the future of LLM-based AI (given the diminishing return of what was considered "scaling laws", see GTP-5) means that most development will not happen vertically (model size) but horizontal through inference time. This is where every nation can use the agentic system they want, and which providers them with the most objective observability and transparency for the public.
For inference-time systems (e.g. "deep thinking") to be effective they need domain/task specific data. And that's where the crux is, as ownership and lineage of our own data is what gives us ultimately our autonomy. Meanwhile the big LLM providers try to silo users by encapsulating their data on their own platform, to become the era's dominat platform (just as in the previous computing shifts MS did with business tools/PC, Zuckerber with "social") by keeping their data hostage on their proprietary formats and datacenters.