AI can be Chinese or American, but not both.

Imagine you've built a system, tested it properly, and then the model it runs on gets retired early. Your options are to pay more for a newer model you haven't tested, or start again. That's not a hypothetical. Harry Schone from Police.AI has seen it happen, and it was the story that stuck with us from our 11th MLOps.WTF meetup.

On Wednesday 30th September we packed out our Manchester office to talk about sovereign AI and open source, under a deliberately provocative title.

Our CTO and co-founder Matt Squire hosted a panel of people working on the problem from very different angles:

• Shahel Khan, AI researcher at i.AI, the government's Incubator for AI
• George Drayson, co-founder and chief AI officer at Locai Labs
• Matthew Wilkes, technical director at Hippo Digital
• Harry Schone, deputy director at Police.AI

What it means if your organisation uses AI

Sovereignty sounds like a problem for governments. But by the end of the night it was clear there's plenty any organisation can do to protect itself.

1. The model you rely on can disappear. In Harry's example, the only replacement for the retired OpenAI model cost a lot more to run and hadn't been tested on the task. If a provider can switch something off, have a backup model ready, and test newer models before your current one is retired.
   ‍
2. You probably need less AI than you think. Most daily uses of AI, like summarising documents, sorting information or transcribing calls, don't need the most powerful model on the market. Open source models already do this well, and you can run them on your own terms, locally, and often for less. It's the same in policing, where we've been working with Police.AI on things like case file summarisation and reusable components that local forces can pick up.
   ‍
3. Protect yourself by being able to switch. Each panellist came back to the same idea: test models against your own use cases, keep your options open, and make changing provider a quick job. Shahel noted that what holds organisations back from open models today is the work of running them, which is exactly the gap good MLOps fills.

"If we are deploying mission-critical systems, and those systems are vulnerable to forces outside of our control… then we need to be examining what plan B is." - Harry Schone, Police.AI

As Matthew Wilkes put it when someone asked if it's already too late: "The best time to plant a tree was 20 years ago. The second best time is now." The same goes for your plan B.

Sam Greenwood, our from our team wrote up the full discussion for our Substack, covering the US CLOUD Act, Locai Labs' "AI in a box", hidden backdoors in open models, and what the UK should do if the US and China pulled the plug. Read the full write-up on MLOps.WTF, and subscribe to our substack newsletter to hear about the next event.

‍