The missing ingredient in AI readiness is state capacity

I don't often share academic research here (for that, follow me on Twitter instead!). I'm making an exception today because reading this research felt so resonant (and almost therapeutic!) after having worked directly with political leaders and senior civil servants on GenAI strategy, delivery, adoption, and regulation for the past few years: A new working paper from Rem Koning and Innessa Colaiacovo shows that while 96% of start-ups use AI, some get almost zero productivity gains while others revolutionise their work completely. They asked founders of North American tech startups: how many more people would you need to hire to maintain your current output, at the same quality, without AI? The answers scatter wildly: → A large group of founders say zero — nearly a third believe they'd need no additional employees at all without the use of AI. → The median founder says they'd need 17% more staff. → And a striking cluster sits at the far end: 10% of founders say they'd need to triple their team (200%+) to produce the same output without AI. What predicts the difference? Executives' and managers' own perceptions of GenAI and how deeply they integrate AI into their orgs' core work: basically how hands-on and ambitious they are with these tools — that's what determines how much they get out of AI. I would love to see this study replicated for civil servants and political leaders... because my experience is that 💯💯💯 the *only* way to get senior leaders in government to govern AI well, understand it, and leverage it in the public good is for them to use frontier genAI, well and ambitiously, on a daily basis. And the stakes are much higher when you make decisions for your entire country. After a decade advising heads of state, ministers and senior civil servants across 40+ countries, I'm convinced the binding constraint on governing AI well isn't a shortage of principles — we have too many, if anything. It's AI state capacity: people inside government who understand the technology and can connect the technical to the political. Personnel is policy. And like founders, leaders don't form accurate beliefs about AI from a briefing. They form them from use. It's very hard to regulate what you don't understand from daily exposure. The closer leaders get to frontier capabilities in their actual work, the better they can trace the exponential, apply it to their context, and worry about the right things. We don't yet know exactly how to best regulate AI under great uncertainty, but I strongly believe that adoption-first, responsible AI deployment in government is what will build the capacity to regulate well and at the speed of AI (on this note, see also Charlie Bullock and Christoph Winter's "Radical Optionality" approach: https://lnkd.in/egjrNQPq) The whole paper is excellent and worth reading in full for all the data 🤓 👇 https://lnkd.in/eErg-epe