Hi friends!
There’s a question I’ve been sitting with all week, arriving in different forms from different angles: whose vision is this, exactly? From Constitutional AI to national sovereignty strategy to sign language models, everything I read kept circling back to who gets to decide what this technology is for, who it’s imagined to serve, and who even gets to be in the room when those decisions are made. I don’t have a tidy answer. But I’ve got good company in the discomfort.
WHAT I MADE
Kara Bombell and I dropped a new episode this week: Heat Death, Finding Joy & Constitutional AI. We got into what Constitutional AI actually means as a framework: what it looks like to bake values into a model’s architecture rather than bolting them on as an afterthought, and whether joy is findable in this particular moment. My answer: yes, but you have to choose it deliberately. The heat death of optimism is real. So is the alternative.
WHAT I’M READING
The DAIR Institute published Whose Categories, Whose AI? and it did what the best DAIR work does: it slowed me down. I’m increasingly convinced that the most urgent work any of us can do right now is to interrogate the visions of the future being handed to us: to ask who built them, what they’re optimising for, and what they can’t see from where they’re standing. The dominant visions are dystopian at best: built by people with parochial imaginations, dreaming of extraction, concentration, and control dressed up as progress. Honestly, how fucking boring.
Joan Westenberg’s piece on agency laundering is one of the sharpest organisational AI reads I’ve encountered this week. The pattern she names: people either unaware of their own agency or actively surrendering it: executives who adopt AI without changing their systems because “everyone else is doing it” (they’re not), employees who defer to AI for consequential decisions because owning a choice feels too risky. What I keep noticing is that the more power someone holds, the less willing they are to name it. So here’s my challenge to you: where do you have agency right now? Are you actually owning the impact of your decisions? When we’re past the hype, that’s the currency that will matter most.
Hannah Merchant at the United States Studies Centre has done careful work on Australian AI sovereignty, but she’s missed something crucial: the framing assumes Australia bargains with the US alone. Mark Carney made clear at Davos that era of unilateral deference is over, and Canada’s foreign policy posture since has proved the strategy can work. We have friends with resources as critical to the frontier AI supply chain as ours, and we haven’t yet built that into a joint bargaining position. Coalition isn’t weakness, it’s the leverage Australia isn’t currently using. If you want to talk Middle Power coalition strategy for AI sovereignty, reply here! That’s precisely what my last paper at Cambridge was about.
BusinessDesk NZ’s Ferrari/Fiat piece connects directly: everyone is obsessed with frontier capability for use cases that simply don’t require it. I’ve been using this metaphor for a while. Honestly, why are we driving a Maserati to the shops for eggs? I’m committed to SLMs, local AI, and just Australian stuff, and I’m particularly excited about Maincode’s Matilda: homegrown, thoughtfully designed, and built out in the open. Democratising AI doesn’t have to mean chasing the high end. It should mean “good enough” models that give real people the skills and agency to build for themselves.
And then there’s this: Keely Cat-Wells’ Forbes piece on ASL technology made my week. Sam Sepah shared it and I’m so glad he did. I’ve been following his work in this space for a while. Google DeepMind has released computer vision capabilities that ease communication barriers for people whose primary language is signing, built with and for the D/deaf community. This is exactly the kind of application I want to see more of. A genuine caveat: we need to make sure these capabilities don’t crowd out diverse signing systems. Someone needs to be building for Auslan. Still, a far better use of computer vision than mass surveillance of innocent civilians, and yes, the bar really is that low.
That’s it for this week! If you have thoughts, reactions, or people I should be reading – hit Reply. 💌