The writer is former editor-in-chief of Wired magazine and writes Futurepolis, a newsletter on the future of democracy
Point your browser at publicai.co and you will experience a new kind of artificial intelligence, called Apertus. Superficially, it looks and behaves much like any other generative AI chatbot: a simple webpage with a prompt bar, a blank canvas for your curiosity. But it is also a vision of a possible future.
With generative AI largely in the hands of a few powerful companies, some national governments are attempting to create sovereign versions of the technology that they can control. This is taking various forms. Some build data centres or provide AI infrastructure to academic researchers, like the US’s National AI Research Resource or a proposed “Cern for AI” in Europe. Others offer locally tailored AI models: Saudi-backed Humain has launched a chatbot trained to function in Arabic and respect Middle Eastern cultural norms.
Apertus was built by the Swiss government and two public universities. Like Humain’s chatbot, it is tailored to local languages and cultural references; it should be able to distinguish between regional dialects of Swiss-German, for example. But unlike Humain, Apertus (“open” in Latin) is a rare example of fully fledged “public AI”: not only built and controlled by the public sector but open-source and free to use. It was trained on publicly available data, not copyrighted material. Data sources and underlying code are all public, too.
Although it is notionally limited to Swiss users, there is, at least temporarily, an international portal — the publicai.co site — that was built with support from various government and corporate donors. This also lets you try out a public AI model created by the Singaporean government. Set it to Singaporean English and ask for “the best curry noodles in the city”, and it will reply: “Wah lau eh, best curry noodles issit? Depends lah, you prefer the rich, lemak kind or the more dry, spicy version?”
Apertus is not intended to compete with ChatGPT and its ilk, says Joshua Tan, an American computer scientist who led the creation of publicai.co. It is comparatively tiny in terms of raw power: its largest model has 70bn parameters (a measure of an AI model’s complexity) versus GPT-4’s 1.8tn. And it does not yet have reasoning capabilities. But Tan hopes it will serve as a proof of concept that governments can build high-quality public AI with fairly limited resources. Ultimately, he argues, it shows that AI “can be a form of public infrastructure like highways, water, or electricity”.
This is a big claim. Public infrastructure usually means expensive investments that market forces alone would not deliver. In the case of AI, market forces might appear to be doing just fine. And it is hard to imagine governments summoning up the money and talent needed to compete with the commercial AI industry. Why not regulate it like a utility instead of trying to build alternatives?
The answer is that unlike water, electricity or roads, AI has many potential uses and will therefore be far more difficult to regulate in the same way. It may be possible to prevent certain harmful uses but it would be difficult to force companies to build models that, say, respect certain cultural values.
The commercial priorities of AI companies, which include pursuing artificial general intelligence, may not align with government priorities either. If AI is used to design social policies, improve healthcare, overhaul judicial systems or provide government services online, it has to be fit for purpose and trustworthy.
Can governments afford to build and maintain good enough AI models of their own? That is starting to look more plausible than it might have a year ago. Research is increasingly focused on quality rather than quantity: using the right data to build the right model for the task, rather than massive general-purpose models. Opening Apertus up to the public should help with this, according to Tan, because it lets the model’s builders gather data on how people are using it, a crucial element in making improvements.
Still, good public AI will be expensive. Solutions to this might include public-private partnerships and international consortiums. Governments could also learn to make good-quality training data available to local ecosystems of developers, who can contribute open-source models and code towards national purposes.
The case is growing for AI models that are designed to serve the public. The more ubiquitous the technology becomes, the more governments are going to need versions of it that can perform the exact functions they require.