Kia ora and welcome to another irregular Memia podcast - I really need to do more of these!
Last week I sat down for an in-depth conversation with Auckland-based researcher Tom Barraclough, one of Aotearoa New Zealand’s most important voices on AI policy, regulation and governance.
Tom brings a decade’s experience working inside, alongside and across from governments in law, public policy and technology. He is a founder and director of a tech policy think tank (the Brainbox Institute, since 2018), as well as a start-up tech company (Syncopate, since 2023), which turns regulatory documents into digital infrastructure for implementation in computers. Tom has led public interest legal research projects, acted as project lead for a global multistakeholder coalition on tech transparency, and has consulted to public and private clients including the New Zealand government, the Global Partnership on AI and the Global Network Initiative.
He’s a super smart mind and we stretched ourselves across a wide range of topics over an hour’s discussion. Some of the key subjects we covered:
* Law as cultural runtime, not neutral code — Tom’s jurisprudence background and why he’s always been sceptical of the idea that law is simply rules neutrally applied; the courts as a kind of interpretive culture rather than a deterministic runtime.
* Paper plumbing and the PDF problem — How every large organisation is essentially “made up of PDFs” that don’t network with each other, and why that document architecture is failing us at exactly the moment we need it most.
* Introducing Syncopate and DocRef — Tom’s origin story at Syncopate: turning law and regulation into structured, referenceable datasets so they can be versioned, annotated, linked at a granular level, and made interoperable across agencies.
* The rules-based order under attack — Why bringing more people into political discourse hasn’t produced the shared interpretive consensus institutions hoped for, and why our “institutional software” may simply be past its use-by date.
* What “radical digital regulatory infrastructure” actually means — The Cambridge paper in progress: publishing law and regulation as open, versioned, linkable datasets — roots that filter up through every digital system built on top of them.
* Unpacking AI sovereignty — Why Tom was initially sceptical of “sovereign AI” (government builds a foundation model — what could go wrong?🫣), and the multidimensional framework he’s developing to pull the concept apart: compute, data, training, content moderation, literacy, individual agency and more.
* The one-in-a-million shot problem — Why even a well-resourced, highly centralised state like China can’t fully nail every layer of the sovereign AI stack simultaneously, and what that means for small countries like New Zealand.
* Te Hiku Media as the positive model — The closing epilogue to Karen Hao’s Empire of AI in practice: community-owned, values-driven AI development in Te Reo Māori as a template for what distributed, multi-stakeholder sovereignty could look like.
* Confidential computing and personal AI sovereignty — Apple Intelligence, Private Cloud Compute, and the Tinfoil startup doing hardware-verifiable privacy — why contractual assurances aren’t enough and what cryptographic attestation offers instead.
* AI-accelerated lawmaking: promise and tension — Tom’s live DocRef/MCP/Claude demo that drafted a new API standard from the existing regulatory corpus in an afternoon, and the open question of how fast legislation should move when stability is itself a public good.
Links mentioned:
* Brainbox Institute: https://www.brainbox.institute/
* Docref.org / Syncopate Labs: https://docref.org/syncopate/blog/
* Tom’s Sovereign AI multidimensional prototype: https://sovereign-ai.tombcgh.tech/
* Te Hiku Media https://tehiku.nz/
* Tinfoil - Verifiably Private AI https://tinfoil.sh/
* Apple Intelligence / Siri AI WWDC Keynote: Introducing Siri AI and More
It was great to geek out with Tom on some pretty technical topics with huge relevance to our future here in Aotearoa New Zealand and the world, I hope you enjoy this episode!
Ngā mihi
Ben
Full Transcript
This post has bonus content for paid subscribers. Upgrade to get full access.
**Tom Barraclough:** [00:00:00] Any large organization that requires any kind of institution or risk management process or policies is just basically made up of PDFs. And they just get layered and layered and layered, and none of these things network with each other.
say you had a government wanting to do a sovereign AI model or even just to fund one, so if you’re gonna have government involved in any of it these are the rough layers that I see you needing to basically ace.
And I kinda feel like even if you nail one of them Getting all of them just feels like this one in a million perfect shot. Yeah. And it just feels like it’s never gonna happen to me.
**Tom Barraclough:** there’s all these different interesting layers of sovereignty that don’t just require it being a government. And I think you can take it beyond government too, and you can start to think about community sovereignty or you can think about individual agency.
**Ben Reid:** agency, to be honest, is the control plane that I’m most interested in here.
**Tom Barraclough:** we didn’t have a [00:01:00] rules-based order. We had a pre-rules based order, and maybe now we can move into a rules-based order if we recommit to
**Ben Reid:** kia ora and welcome to another irregular Memia podcast. I really need to do a few more of these. I’m Ben Reid, and I write the weekly Memia newsletter, and generally spend a lot of my time drinking from the fire hose of accelerating AI
this episode I’m gonna be talking with Tom Barraclough. Tom’s an Auckland-based researcher, former lawyer and one of Aotearoa’s most important voices on AI policy and governance.
He’s the founder and director of tech policy think tank the Brainbox Institute, as well as tech startup Syncopate, which turns regulatory documents into digital infrastructure. Tom has consulted to many public and private clients, including the New Zealand Government. Welcome Tom.
**Tom Barraclough:** Thank you, Ben. Really nice to be chatting about all of these things and share a drink from the fire hose.
**Ben Reid:** Yeah, there’s a fair [00:02:00] amount going on, and we , we need a few more hours.
**Tom Barraclough:** Yeah ...
**Ben Reid:** we’ve only... We’re, we’ll budget ourselves to one hour not really into Lex Fridman-type five-hour marathons. We could. All so look key things I’d like to talk about today is what you’re doing with Syncopate, Yeah
and your current research focus. And I think as part of that, we’ll basically go in on the whole concept of AI sovereignty. So you’re doing work which I think is probably some some of the most advanced analysis of pulling that whole concept apart and ‘cause it means a lot of things to a lot of people.
. Let’s also talk about some of the tech sovereign strategies in UK and Europe that’s happened recently. So plenty to be getting on with. Yeah. You and I go back many years.
... And I think we touched base for the first time when I was running the AI Forum New Zealand back in- Mm ...... 2019 or something. So that’s a long time ago. So these conversations have been going on for a long time. And you’ve been right at the center of this with Brainbox and then more recently Syncopate.
So maybe just give us a sort of potted career history for yourself and and , the really interesting work you’re doing at Syncopate these days.
**Tom Barraclough:** [00:03:00] Thanks, Ben. The potted history would be that I... So I studied law and politics, and n-none of that really came alive for me until I started learning about jurisprudence and the kind of legal philosophy side of things because that sort of lets you sit back and go, “What is this law thing?
And what is it-- what do people think it does, and what does it actually do?” And yeah, how does it all work? Whereas the other side of law is quite like what does this document say about this particular thing, and what do we reckon a court would say about it if we ever went to court? And that doesn’t interest me that much.
It’s quite speculative, and there’s just a lot of layers of frankly b******t caught up in it. One of the things that I think always fascinated me about it was there’s this way of framing law that is very much code. So it’s very much hey, we’re gonna write the rules, we’re gonna neutrally apply the rules [00:04:00] to whatever comes up as it comes along, and I as a judge will just sit here and go what do the words say?
I better just do that,” so
**Ben Reid:** the court, the courts, the whole legal system is a sort of runtime for-
**Tom Barraclough:** Yeah.
**Ben Reid:** That’s what I think of it ... for, the law as software.
**Tom Barraclough:** Yeah, but I think the thing that always fascinated me is that I don’t wanna go too far out of my depth on the tech side of things, but it’s like the runtime is actually probably more like culture.
So one of the things that they teach you about really early on is statutory interpretation. So there’ll be like words in a statute and one of the interesting things is how do you design that statute? But the secondary thing is then so if these are the words on the page, how do we decide what those words mean?
And I was on a-another podcast recently and we were kinda talking about this, and a lot of people don’t think about law that much. But then as soon as you start to open the door to weird examples, everybody’s got their favorite story
it, it all touches people’s lives all the time- [00:05:00] Yeah ... even though they don’t necessarily think about it.
**Ben Reid:** Yeah, and you try and avoid getting actually, involved with it, right? Because it’s like- Yeah ... it’s like something you really don’t want to be near.
**Tom Barraclough:** Another interesting layer of it is if you-- So this is part of what we’re doing with Syncopate as well.
If you take it away from just thinking about legislation, like stuff published by the Parliamentary Counsel Office that comes through Parliament, and you instead think about the way that documents actually form the architecture for any complicated organization or institution. Yeah. So I’ve heard Jack Clark at Anthropic refer to this as paper plumbing-
**Ben Reid:** Yes
**Tom Barraclough:** which I think is like a really nifty way of thinking about it. Any large organization that requires any kind of institution or risk management process or policies is just basically made up of PDFs. And they just get layered and layered and layered, and none of these things network with each other.
**Ben Reid:** And they’re all, they all sit on a [00:06:00] remote- SharePoint, if you’re lucky ... SharePoint drive somewhere or-
**Tom Barraclough:** SharePoint if you’re
**Ben Reid:** lucky. Yeah. Yeah. And never really see the light of day to be exercised, and then when they are exercised, it’s all in people’s heads and- with this layer of organizational politics and influence and- ... personalities that, like you say, the interpretation is really very dependent on the sort of 86 billion neurons in every brain, right? And how that-
**Tom Barraclough:** Yeah. The, the-- one of the things I could never quite wrap my head around was just the, I think that’s why I say it’s a real cultural thing.
So like you look at words on a page and you’ll be like I read it this way.” And but some-- there’ll be this thing that happens where you say that and every- and it goes quiet and everybody shuts down because you’ve stood on a landmine of some kind, and they, that’s not how they think it.
So that always really fascinated me as to like why y- certain things don’t get argued or don’t accepted- Yeah ... by courts.
**Ben Reid:** And if you, maybe we hold that thought and come back to that later [00:07:00] because I think this, that’s really relevant to the future we’re heading into now around consensus, truth, and reality.
Tru- not truth, but the consensus on truth. And you have, historically we’ve had very shared- assumptions probably driven by- Sure ... political hierarchies and so on, and power structures. But we’re entering into this moment now where increasingly we’ve got these filter bubbles being engineered, and so you might say something like that in a room- and nobody would know where you’re coming from because you’ve just come from your own filter bubble. And so y- it feels like the, this law, which is the sort of is the software, the code that we organize society upon is, it’s almost under a denial of service attack
**Tom Barraclough:** in some ways.
**Ben Reid:** Yeah.
**Tom Barraclough:** Yeah.
It’s interesting too, ‘cause I think, ... So one of the other areas that I’ve worked on is around disinformation policy and social [00:08:00] media policy as well. And one of the things that’s a little bit counterintuitive about that is, in a way, we’ve been asking people to be more politically engaged and interested in how systems work.
But, I came of age in the Bush era, so it was like everybody was going, “Wake up, look at what they’re doing. They’re tell- they’re lying. You have to think about this. You gotta get involved.” And it’s now everybody’s super involved. It just turns out that they probably think things that are quite different from what the kind of people in those institutions thought they would think.
**Ben Reid:** Yes.
**Tom Barraclough:** And I think that’s what you’re seeing played out internationally across different institutions so it’s this weird thing too where, you bring more people into the conversation, and you’re right, they don’t have the sort of lawyers would say they don’t have that shared interpretive context.
Yes. So they look at it and they go that’s not how I understand that,” and it [00:09:00] starts to break down. The, I think the other thing that was so interesting about it was the- there’s this sort of post-World War II theory on how things work, and we’ll set up institutions, and they’ll have documents with rules in them, and we’ll apply them, and it will be all out of our hands really ‘cause we’re doing it under the rule of law.
And there’s always been like a really rich, critical tradition of people looking at that and going, “Yeah, but that’s not how it works.” Clearly things are being selectively deployed to protect particular power structures or whatever. A lot of it came from like the left as well, it-- ‘cause so it wasn’t a left right thing is what- Yeah ... I’m trying to say. There’s been this really like rich layer of critical scholarship about how rules-based institutions work for ages, and now it’s that we’re bringing more people into it, that it’s just coming to the fore and it’s getting harder to defend.
And you see that, with human rights stuff and international law stuff and I had a really pithy quote yesterday from Alexandria Ocasio-Cortez talking about the [00:10:00] rules-based order, and she was saying, “Actually, we didn’t have a rules-based order. We had a pre-rules based order, and maybe now we can move into a rules-based order if we recommit to
**Ben Reid:** it.”
It’s very optimistic. I... a lot of that rules-based order was based on the threat of nuclear- Of course ... weapons, right? Based on- Yeah ... what happened in Hiroshima and Nagasaki in in the Second World War. And, that’s really carried through to today. Yeah. Like as a technologist, my view is that, and this is probably where your work at the intersection of institutions and technology.
I, I just think that legacy institutions are they’re just legacy technologies, right? And so as a, software architect of many years, something that worked 10, 15 years ago y- is just redundant now because, at the time, the am- the environment in which it’s operating has moved on.
And so- ... I feel that pretty much all of the institutional software, if you like, that we’re operating on [00:11:00] is past its sell by date. And, you have A number of architectural approaches that you would take. You can, rip and replace or or do a staged migration or there’s a, what’s called a strangler fig strategy, which is you build components around the old one until it atrophies and then to-totally stays standing afterwards.
**Tom Barraclough:** So I, I should bring this back to actually what Syncopate is all about. Yeah. But I think that is really important con-- it’s always important for me to give the context for why I find this topic fascinating, why I think it’s important. And so basically, one day I saw my co-founder at Syncope, Hamish, tweet about the XML structure of a bit of legislation that I was intimately familiar with and was looking at constantly applying these lenses of Hey, they say this is a neutral rule book, but it’s clearly not.
And it’s-- The other thing was the statute was really poorly designed. It looks like a bit of code, but it’s like garbage. It’s really bad. It’s logically inconsistent. And everybody knew this too.
Yeah. There were [00:12:00] courts saying this core definitional concept is both circular and doesn’t make sense. And the other thing was , it was out of step with medical science around causation. So I was bringing all that context to understanding what Hamish was doing, and essentially what he was doing was working within government trying to turn law into code.
So turning law into software code so that you could run it in computers, you could model policy outcomes, and most importantly, you would have interoperable rule books basically between government agencies so that you could break down silos. You could have calculators so people could understand what they might be entitled to.
It was this really great, quite utopian vision that fed into a lot of the stuff that was going around at the time around government as a platform. And it’s really interesting how much of that kind of anticipates, I think, what you were referring to, which is we’ve got these sort of paper-based, hierarchical processes that just aren’t gonna keep up with what we need [00:13:00] from them.
So what we did is I had Brainbox. So Brainbox was all about how you regulate technology in a sensible way. A big part of that was using human rights as a core organizing framework for thinking about new stuff. And so looking at what this sort of group of people in Wellington were doing filled me with dread a little bit because what they’re essentially proposing to do is take law, which is this like rich, contestable, cultural thing that embeds within it the ability to change over time because of the way interpretive context changes, and also lets you importantly go to court and go, “Hey, so this is what’s happened, and this is what the rule book says should happen.”
And importantly, there’s a p- purposive bit in here too. Most statutes have a purpose section, and the reason that’s there is because you might get a really weird sounding rule or a conflict between rules, and it’s meant to be this [00:14:00] guide from Parliament that says, “Hey here’s what we’re trying to do with this.”
It’s like comments in the code. And the thing is that all of that was being overlooked by the people involved. Yeah. So this process of going how is code not like law, and why is that a good thing?” And then also understanding that what you’re doing when you’re taking the law is you’re interpreting it.
You’re making a judgment call about what you think it means. And importantly, that always has to be able to be challenged, particularly if it’s like the executive government doing it. ‘Cause otherwise, like I’m cutting across a lot here, but it’s essentially the distinction between having commands from a king.
So you can imagine Parliament basically going, “This is the law. It’s in code. That’s what we meant, and we’re gonna do this regardless of the consequences.” As opposed to, “Here’s what the law is, and we can all go and argue about what it actually means.” It does this really profound thing in terms of breaking down the separation- It is
of power and the ability to contest things. - And it’s important that we keep bits of that in what we’re doing.
**Ben Reid:** Not just keep them, [00:15:00] but actually make them more real, right? Because actually that process of lawmaking And law enforcement and-
**Tom Barraclough:** Yeah ...
**Ben Reid:** and the court system and so on, is actually very closed and- Yeah
And authoritarian in many ways. And so there, there isn’t really... one of the things I, again, applying a software lens to this, but that process of making a law, where it goes, it’s almost like this waterfall process. Exactly. It’s entirely like that. And we... whereas you look at modern software development, which is just basically freedom to fork, everything’s in a Git repository, you know- take it off to the left, take it off to the right, everybody tries little bits. And the, I would say one of the things that the work you’re doing opens up is to actually completely reimagine that process of lawmaking- ... and the, and almost the open sourcing of it beyond the, because these centralized institutions are so vulnerable to, to capture and that as we’re witnessing.
**Tom Barraclough:** Yeah. I think there’s this other thing, So totally agree on part [00:16:00] of the other work that I was doing before all the tech stuff was around access to justice, and it was understanding that there’s so many things in the world that people could go to court about and enforce their rights but they don’t.
So why? And part of that is a cultural thing, part of it’s a financial thing, part of it is that courts are really slow, like the process of going through court- ... takes forever. And it’s also so complex the hourly rate of anybody who can navigate it just will always be in excess of the value of some problems, so they never wind up in the judicial process.
And if you think about that’s pretty wild, ‘cause most of the problems that people have will probably never rise to the level that it’s worth instructing a lawyer, yeah.
**Ben Reid:** That is the rules-based order, basically- ... that we’ve all lived under for 80 years.
**Tom Barraclough:** I’ll just add one more layer to that, which is that a huge part of it also is a document navigation problem. And I’m bringing this both from legal experience and from public policy experience. So if you wanna understand any system and you [00:17:00] wanna engage with it, whether it’s going to court, whether it’s providing legal advice, whether even if it’s submitting on a bill or something,
There’s so much that you have to do to bring yourself up to speed on what already exists that that is also really hard to do, and that undermines the quality of what you can bring to it. It means that people are necessarily excluded ‘cause it takes so long and it’s so complex. The other thing about it is that when you’re designing...
Let’s say you’re doing a rules as code system, you’re not just gonna be going through one statute and going line by line, like if this, then that. What you’re doing is you’re pulling on a network of documents. Yeah. So there’s gonna be multiple statutes, multiple sub things, all that internal paper plumbing in an organization as well.
And the experience of doing that currently is literally open your browser, open the search tab, start right click, open tab, start downloading all the PDFs, try to keep track of all of your stuff, and it’s so hard to do. And all of that has nothing to do with actually [00:18:00] reading the stuff-
**Ben Reid:** Yep ...
**Tom Barraclough:** or taking notes and keeping all those notes and pulling them all together.
So there’s another big layer to this, which is actually just dealing with documents as data sets.
**Ben Reid:** Yeah.
**Tom Barraclough:** So we still very much approach web publishing even. Even HTML is really not that different from thinking about printing a page- ... and reading it left to start to finish. And a big part of what we’re doing with our product, called DocRef is we’re taking documents.
We break them down into data sets and making them referenceable at really granular layers. And that turns out to create this infrastructure layer. So the reason that we did it originally was because when you’re writing code that implements law, you need to have a reference for it. And at the moment, with a lot of the systems, the most you can do is point to an entire document or an entire section- Yeah
when you might wanna be pointing to a cell in a table or a sub rule or something really granular. But then if you’re working with documents as data sets as well you can do other things. So you can [00:19:00] annotate them, and you can keep all of your annotations as a data set and take that with you.
You can version them and show how they’ve changed over time. So you really bring all these software tools to law and regulation-
**Ben Reid:** Yeah ...
**Tom Barraclough:** that have been in place for such a long time, but law just hasn’t caught up with. Like multi-cursor editing in VS Code. Think about how useful that would be for so many public servants.
**Ben Reid:** Absolutely.
**Tom Barraclough:** And then there’s a rich AI layer too. Cool. So that’s the other sort of transformative thing.
**Ben Reid:** Yeah. And I’ve seen some of the demos of DocRef that you’ve put up there, and I think it’s a really practical tool to bridge that gap between, like you say, static PDFs- And and at least being able to operate in a software world, which, the rest of the, rest of us will have been doing for 20 more years.
So let’s bring that forward now- ... to this environment we’re in, where we have, large language models increasingly world models are coming on stream, which make the , the corpus of law for any country look like a tiny [00:20:00] little corner of the whole data set that they’re on.
And one of the quotes I saw from Bruce Schneier, the security guru a couple of, or a few weeks ago after Mythos had been mythically released, too dangerous to release, they’ve just released it. It
**Tom Barraclough:** em- it emerged.
**Ben Reid:** Yeah. So it’s a fantastic piece of pre-IPO marketing.
But yeah, Schneier was saying that if you think that the way that Mythos finds bugs in software is dangerous, right? So there’s big cybersecurity concerns. Wait till it finds the bugs in law. Yep. And so- Yeah ... just going and creating entire tax haven scams where whole structures where you can avoid tax anywhere else in the world, finding all of the gaps within it.
And so it feels like the, this legal document-based sort of convention, you’ve got, highly paid lawyers im- implementing it. You’ve got courts providing a lot of theater around that. [00:21:00] And then, but at the same time, you’ve got scarcity of actual access to that.
It feels like that’s ... It is literally under a denial of service attack- ... already. And so the people with the ability to actually leverage law are going to be the people that control these most massive models, right? Are the most powerful models- ... in the world, and the agents that run off those.
Maybe talk through some of the implications of that. But I think the,, where this is leading to is that question of, inverted commas, “AI sovereignty”- ...... which is how does a country with its own jurisdiction, its own courts, its own legislative bodies , maintain any kind of agency , in this c- in this world that we’re heading into?
And what is required of AI sovereignty? And then maybe we’ll look into some of the frameworks you’ve been developing at Syncopate. .
**Tom Barraclough:** So one the way that we’re thinking about this conceptually [00:22:00] is through this idea of radical digital regulatory infrastructure. So we’ve got a paper that it’s been accepted for publication in a Cambridge journal, which is pretty exciting.
Just finishing that off. We talk about radical in a few different senses, and really what it means is we need to take an approach where we’re publishing law and regulation as understood across that full stack as datasets and as referenceable datasets that can be versioned and updated and pointed back to and referenced and all this kind of stuff.
And that way they form this kind of digital infrastructure for regulatory systems to be built on top of. That then also requires it to be governed and managed quite carefully in terms of the way that you manage all those links, and you change the content and all that kind of thing.
The reason it’s radical is I’m pulling on the the etymology of that as root. A radical rather than extreme. And the idea there is [00:23:00] that, the sort of links and the data points that form this regulatory layer are like these roots that filter up through all of these different digital systems.
And that infrastructure needs to be open and available, in part because at the moment it’s not. So at the moment, the only way to get access to all the case law in New Zealand that exists is through a commercial subscription- Yeah ... to a non-New Zealand publisher. We do have really good stuff around publishing legislation and other regulation, but it’s all over the place.
Like I say, to go and track it down is really hard. And there’s this real thing as well where, It might be like an old version of a set of rules, and they just change it and don’t tell you, or they just rip it down so the original link like 404s-
**Ben Reid:** Yeah ...
**Tom Barraclough:** and you can’t find it.
**Ben Reid:** So where’s-- So again, these are problems that the software industry’s been dealing with for decades, and-
**Tom Barraclough:** Yeah
**Ben Reid:** something like Git, Oh, yeah ... pretty much solves that straight away. Yeah. You’ve got full version control, who changed what where it’s a solved problem. ... And it just feels that you’ve got [00:24:00] this, very protected industry in many ways that’s really not fit for purpose to- Yeah
**Tom Barraclough:** to solve all these issues. I think the way that feeds into the mythos thing is probably where it’s radical as in like radical power shift, is that we, in our experience, we think that the kind of digital regulatory systems, like these rules as code systems and everything else are necessarily gonna be multi-stakeholder in the way that they’re developed.
So there might be some things where government does all this internally and then just does it, but for all the reasons I described at the start, that’s quite problematic from like a rule of law perspective to just have these encoded rule systems being done solely by executive government.
What you’re gonna want is you’re gonna be wanting people to be able to get in, similar to open source- Yes ... and actually be able to check and vet and audit and all that kind of thing and be able to suggest changes where they think that the code is out of step with- Yeah ... the law. A- and
**Ben Reid:** in, certainly in our [00:25:00] country, I just don’t see that becoming reality in terms of being centrally driven.
We’re
**Tom Barraclough:** gonna get there. Yeah. We’re gonna get there.
**Ben Reid:** Well- yeah. It’s gonna be- But is it, would it be better, would it be better to run under something more like the Linux Foundation, right? Where it’s actually- ... you actually have something which is non-governmental, not for profit and- I think- Yeah
**Tom Barraclough:** I think so, and that’s where there’s an interesting link to the sort of sovereign AI thing too. Like I think the way that this is gonna happen is we’re gonna have, ‘Cause it can’t just be done by companies either too, right? So the example I use is imagine if Meta showed up to the European Commission and said, “Hey, don’t worry about it.
We’ve written code that shows that we apply the law faithfully in every single example.” They’re just gonna be like whatever” and then even if the commission and Meta could get on board, you’d have civil society and human rights advocates being like what is this cozy b******t? We want in too.”
So then you’re at three stakeholder
**Ben Reid:** groups. And then you get to the point where every- everybody’s intermediary with base reality is these commercial models [00:26:00] from- Sure. Yeah ... OpenAI, Anthropic and Google and Meta. And the, we’re already seeing, signals that some of this is that, th- there’s a degree of molding reality, starting to happen there.
**Tom Barraclough:** Yeah. So We my answer to the Mythos thing is basically we see a future where there’s actual equal access to the kind of regulatory substrate necessary to do that kind of thing. And then pivoting to the sovereign AI thing, like m- we need to basically open up access to an equality of arms in that regard, the the reason I got interested in sovereign AI was when I first heard the concept, I was super skeptical. Like the i-- to me, I heard it as governments creating foundation models. Yeah. And I just thought, with all my history, I was like, “That sounds terrible.” Like one, can that even be done?
But two like that just how do you even begin to [00:27:00] deal with the politics of all of that? And probably one of the things I was trying to do with the snazzy three-D diagram that- what I was trying to do there is illustrate why I thought that even if you think about a really successful version of this, the different layers associated with it are so politically and economically fraught and legally fraught that to kind of line it up, it required this almost like straight shot.
It required this one perfect shot where you could nail, the training data and the legal foundation of it, and then get some agreement on how that should all be curated and how the training process should work, and then who gets to use it, who doesn’t get to use it, for what purpose, what do you charge?
It’s just so hard to nail that. And so what I was really thinking was at the same time, we do need more control, more sovereignty, more autonomy, more agency over AI and the way that we give [00:28:00] everybody access to it, particularly if the same people are saying “This is gonna be transformative and empowering,” and all this kind of stuff.
So the question on my mind was really like: how would we go about this in a way that distributes that in this kind of same multi-stakeholder way? And this is my answer in a way that is meant to be a bit more sophisticated than a two-D diagram.
When I first heard the concept of sovereign AI, based on some of the context that I’ve shared about, governments and holding governments accountable and all those kinds of things. And I guess also the power of AI that we’re talking about is, can either empower states or empower companies or empower people.
But probably more than that too I was trying to think through what would the process even be for that. People were talking about the New Zealand government, for example, opening access to government datasets or opening access to Archives New Zealand or Ngā Taonga Sound & Vision , all [00:29:00] kinds of things.
And even aside from thinking about the kind of Māori data sovereignty component of that would be super tricky to navigate. I think even just everyday people would be like, “Yeah, that’s not really why I gave you the stuff,” like there’d be a real kind of like purpose issue there in terms of, privacy and all kinds of things.
So I could just see all these practical impediments to actually it going ahead. At the same time, I was also playing around with using AI more and getting more and more confident that there is a lot of merit to the use of large language models for all kinds of things. And I guess also linking that to what I’ve said about people’s ability to navigate regulatory systems and to turn those to their advantage in a kind of way that’s consistent with the public interest.
So I, I did a couple of talks on this. I did a discussion paper pulling together some of my thinking and, surprise, people don’t get excited about reading [00:30:00] discussion papers and giving you feedback- ... unless they’re particularly particularly-
**Ben Reid:** But they-- But we do get very excited about 3D vibe coded models, right?
So
**Tom Barraclough:** Exactly, yeah. So I tried to do like a diagram and actually this-- So this was produced using the Claude app on my phone, like from my bed while I was sick. And a big part of the input was like a terribly drawn like finger thing on, on the iPhone that roughly showed the layers and all that kind of thing.
So I think this is just a real manifestation of some of the empowerment potential of- Yeah ... AI as well, right? If people could articulate what they’re thinking in new ways then that’s great. So it’s this manifestation of what I’m trying to do at the same time. So just to explain it the reason I’ve gone for kind of a three-day approach is because I wanted to do-- get to this point.
But I’ll pull back to here. So this to my mind is let’s say you had a government wanting to do a sovereign AI model or even just to fund one, so [00:31:00] if you’re gonna have government involved in any of it these are the rough layers that I see you needing to basically ace.
And I kinda feel like even if you nail one of them Getting all of them just feels like this one in a million perfect shot. Yeah. And it just feels like it’s never gonna happen to me.
**Ben Reid:** E- even, the closest that you would get to this is what China’s been doing, right? Yeah. And even then, there’s basically lots of gaps in that whole structure there.
It’s very porous.
**Tom Barraclough:** And, China can do it because of its power of the state. And then even then, some of these issues like model training, content moderation, what the models can say or can’t say who gets access how it gets governed. That’s still stuff that the Chinese government and Chinese companies have to navigate and find some agreement on.
So it’s like even in, in the case of a really well-resourced, highly centralized state, it’s still pretty complex as to how you pull it off. So that’s quite a good comparison to give. So some of the examples here is like you need to have the [00:32:00] people, then you have to pull the money together, and then you’ve got to deal with all the compute and think about how to manage that.
Then if you’re a government, you’re gonna have to navigate the issue of sort of environmental footprints. There’s all the issues with data, IP, consent. The data curation and training process, and then the content moderation side of it imagine how politically fraught that’s gonna be. Can you imagine the first time somebody asks a New Zealand model, like whether it’s Aotearoa, New Zealand or New Zealand and then the consequences of that? I just think that’s, it’s never gonna-
**Ben Reid:** Yeah ... gonna happen. And it’s it’s been very interesting watching how the big labs have used personality in many ways to defuse some of that, and so
to deliberately avoid conflict and to be deliberately helpful, and so on. So- yeah, I think there’s lots to be learned there. And the- one of the pieces that that’s missing from there, I think probably since you’ve done- did this, has become a huge issue, is that of you’ve got environmental in there, but also energy.
**Tom Barraclough:** Yeah. And so [00:33:00] a lot of debate now around AI in most, the US and other countries is where’s all the energy gonna come from?”
Yeah. What’s
**Ben Reid:** the policy on that, right? Yeah.
**Tom Barraclough:** And it’s a little bit of a straw man too.
I think to be fair, a lot of the people advocating for sovereign AI weren’t necessarily thinking about this, so I wanna be fair to them as well. I then started to think okay- I think another big piece of this is if it’s not about a single model and a government, so if it’s not sovereign as in government and AI as in single model, what is it instead?
So I unpacked a little bit in this discussion paper, like what do I think about as being a matter of sovereignty, and I kinda pulled it apart into a few factors which are in here somewhere that I won’t go through. But it’s essentially stuff like it’s onshore, it’s it reflects New Zealand values, it’s subject to New Zealand law, it’s...
there’s all these different interesting layers of sovereignty that don’t just require it being a government. And I think you can take it beyond [00:34:00] government too, and you can start to think about community sovereignty or you can think about individual agency.
**Ben Reid:** agency, to be honest, is the control plane that I’m most interested in here.
We’re almost devices attached to these models in many ways now. And how do you maintain any kind of balance with something- Yeah ... that’s orders of magnitude, more powerful and richer th-than you as an individual, yeah.
I tried to I tried to go through things and knock stuff out and go would you still call it a sovereign AI model?” So the example there for me would be like, let’s say it’s all New Zealand data, perfectly reflects our values. It’s all onshore, it’s all publicly funded, but nobody in New Zealand knows how to use it effectively.
**Tom Barraclough:** We wouldn’t say that we’ve achieved sovereign AI, yeah. So AI literacy is a really big layer of all of that. And similarly with yeah just all kinds of different things. Like the compute. Let’s say we had a perfect foundation model, but there were no data centers in New Zealand that could run it, [00:35:00] and they were extremely expensive.
We wouldn’t say that we had sovereign AI in New Zealand.
**Ben Reid:** Yeah. It’s the- it’s the training more than the inference that would be the cost. And look, it’s interesting talking about this on a nation state. Again, the point I was making earlier about institutions and, and- nation states almost being legacy technologies. This concept of sovereignty- ... crosses borders, right? So you have to be pluralistic here with teams doing similar work in other small countries around- Yeah ... around the world to counterbalance the massive venture capital strategically funded US and Chinese labs.
And- Yeah ... and Europe. Europe’s, already way behind on, on this as well. And so the only kind of- Going
**Tom Barraclough:** hard after it as well,
**Ben Reid:** Yeah, going hard after it- I think there’s- ... there’s al- always a lot of PDF documents like GDPR,
so yeah- Yeah ... andrej Karpathy, he’s gone into Anthropic now. He had a [00:36:00] slide which I use sometimes, which just compares the current model ecosystem to that of operating systems. And so you have- Yeah ... macOS out there, you have Windows with a big franchise as the two main commercial closed source operating systems, and then you have Linux.
But within Linux you have, a v- a a zoo of different sort of distributions and different purposes and a sort of Ubuntu and Debian and and so if you think now around AI models you have, the Anthropics, the OpenAIs the Geminis out there, closed source and, commercially available.
But then this bubbling under ecosystem of lots of open sy open source labs and, by virtue of the funding that’s been poured into the Chinese labs mainly, are being released on an, on an open weight basis, not a open source. Mm-hmm. , which means that there is some kind of, [00:37:00] um- counterweight.
And yeah, I’d say the recent research I’ve seen is that they, they’re basically trailing the frontier around four months. And so this week we’ve had-
**Tom Barraclough:** Yeah ...
**Ben Reid:** Fable released. A
**Tom Barraclough:** lot of that. Yeah.
**Ben Reid:** Yeah. And then and that was what I always found so absolutely ludicrous about the Mythos mythology, that it was just too dangerous to release.
It’s no, all it’s doing is just basically tilting the playing field for cybersecurity, in favor of the attacker briefly, because they’ll suddenly... you’ll suddenly find, 1,000 more zero days in a day than you used to. But, within four months, , that playing field’s, going to level out and, you know- Yeah
I suppose there’s a question of whether the frontier labs extend that period or not.
**Tom Barraclough:** I think again that’s where the questions about equitable access come into it as well. Yeah. And so to prevent that becoming a sort of power base that’s used in improper ways this point about equitable access, and for me that would mean, assuming they’re open weight models it’s not [00:38:00] just access to the model, it’s also the compute to run it.
Yep. Like I’ve done a lot of experiments with just seeing what I can do running on a local 48 gig M4 Pro, and it’s amazing. So like Qwen 3.6, the mixture of experts one, I’m pretty impressed by that, but still that’s like an expensive laptop and it’s not that good. And then there’s also the question of like- Let’s say you had that, how many people in New Zealand are really capable of even using it and then understanding the limitations of it- and doing all the context engineering kind of stuff and-
**Ben Reid:** let’s hold that, let’s hold that thought because I, in some ways I... It feels that if you look at the OpenClaw moment earlier this year, where- ... basically, a large number of people all around the world, including, and in particular in China, downloaded an, a piece of open source software and installed it on their local device.
**Tom Barraclough:** Doing it securely is like a whole other thing that- I- it,
**Ben Reid:** it wasn’t
**Tom Barraclough:** secure ... people don’t
**Ben Reid:** realize. It wasn’t secure.
**Tom Barraclough:** Yeah.
**Ben Reid:** Yeah. But it was a really interesting moment for me [00:39:00] ‘cause that’s the first time I’ve seen open source go viral like that.
And I get a sense that’s actually a dynamic that as the implications of , Google now with their AI summaries completely shaping searching the web just within the last month- ... has totally changed. And so the power that organization has always had with its ranking algorithm is now even more distilled and they have- Yeah
the distribution. And that’s why you would never write Meta off, right? ‘Cause they got distribution. So where do we go? Yeah. So look I think the model that you’ve put up there and that was really, that was one of the key things I wanted to talk about on, on this podcast, was
So , where do you take it from there? So I think you’ve- ... got a really valuable thread that AI sovereignty is very many layered. No nation state, even China or the US, is going to be able to completely control every layer of that stack. So how does a s- how do small countries [00:40:00] like ours take that model of sovereignty and then, make it real?
**Tom Barraclough:** I one of the things that I put into this site is between me doing some of this thinking and putting my discussion paper out, there was a series of releases by think tanks like Brookings and , there’s the World Economic Forum, I think a few different others who basically adopted a very similar way of thinking about this to me, I think there’s some other reading material on the site, basically. And I-- And for me I’m really interested in sovereign AI as basically like a policy direction So what I thought to myself was if this is what we think of as being sovereign AI, the task for us is to think about what does our sort of regulatory and policy system look like to work as a kind of systematic coordinated thing towards this kind of policy outcome.
And to be clear that’s AI literacy [00:41:00] and education. It’s also just thinking about access to compute in New Zealand. It’s thinking about methods of fine-tuning models that already exist and things like access to data sets. And then maybe it’ll be-- I think there’s still this really interesting space where perhaps like an iwi collaborates with, a technical thing and then maybe another funder to actually produce a new foundation model.
Yeah. I wouldn’t rule that out. I think it would be really hard, but I’d love to see it happen on that kind of distributed multi-stakeholder basis. Yeah.
**Ben Reid:** And look, technologically, there’s a lot of work being done around distributed decentralized training.
**Tom Barraclough:** So- We also have, I should say as well, we also have the shining light that has rolled out in the space is Te Hiku Media up in Northland as well.
They’re literally held out in Karen Hao’s book, Empire of AI, the closing, like epilogue to that book is what does a more positive vision of AI look like if it’s not what OpenAI is doing? It is what Te [00:42:00] Hiku Media is doing in New Zealand with their language models based on community ownership and governance, and really values-driven deployment of where that model gets to be used.
**Ben Reid:** It’s really, incredible work that they’ve managed to achieve on, pretty much a shoestring budget as well.
I think we touched a little bit earlier in terms of the compute challenge, right? One of the layers inside your AI sovereignty model was, having sovereign compute.
And you can run mo-more complex models increasingly out of the edge. But the challenge w- the moment that you start using cloud services is that people can see the information that you’re sending to the model. They can reuse that for training the next model and, your whole security chain’s broken.
So just last week we heard more from Apple about, their next generation of Apple Intelligence, which is actually largely based on Google Gemini. They’ve done a good deal there. But yeah, you’ve been following that quite closely, so maybe you talk [00:43:00] through some of the architectural and, personal sovereignty things that are coming through in that architecture.
**Tom Barraclough:** Yeah. I spent a long time not being an Apple guy, and I should declare my bias now that I’m like all in on the Apple ecosystem. And part of that is because ‘cause of the convenience and the integration, all of that kind of thing. There does need to be a non-Apple version of this,
My take on what’s happened with Apple Intelligence is they’re essentially turning like your phone and then all of your connected devices into a RAG context,
**Ben Reid:** yeah.
**Tom Barraclough:** So what’s happening on my phone right now while I wait for access to the new Siri, is it’s like systematically indexing the thing over like a number of days.
It’s taking forever. But the consequence of that is it’s essentially a context window. Like it’s a context sort of architecture for Siri to go and do stuff with my phone. It can also read what’s on my screen, and then [00:44:00] it can also take actions. So it’s the exact open core model really but going even further than that.
Like the integration into my digital life is gonna be, frankly, a little bit creepy particularly when it can see what’s on my screen. So the question then is like, what do I need to have confidence in actually doing that? And a big part of that is is actually not relying on contract for those assurance mechanisms.
Like I wanna see- what
**Ben Reid:** would you rely on instead?
**Tom Barraclough:** I think encryption. So I think what you’re seeing some really interesting stuff around is the use of basically... I think it’s not just end-to-end encrypted communication, it’s this other thing called it’s like verifiable at the hardware level.
It’s called confidential computing.
**Ben Reid:** Yeah.
**Tom Barraclough:** And there’s this one company that I’m really interested in, it’s a startup in the US called Tinfoil, and they basically have, they’re all sort of security engineers from like Cloudflare and stuff with MIT PhDs, and they can show you at the hardware level that there’s no way they can see [00:45:00] what’s happening.
I’m very excited about that as a concept, and I basically understand that the private cloud compute that Apple does is basically like a version of that. Yeah. So the idea then would be that you can... It doesn’t really matter where the the servers are y- you, but you can get access to these foundation level models without having to have all the local hardware and still have confidence that you can actually disclose your entire digital life to this- -system but not have to take it on trust that it’s not being used for training or something like that.
**Ben Reid:** I think, I think there’s still a way to go and the advent of quantum computing coming soon potentially, reduces confidence in current encryption. -And so there’s always a race happening that, means that one, the trust levels in the technology are always, they’re always at risk.
**Tom Barraclough:** Yeah.
**Ben Reid:** It’s very interesting. Apple tried to do Ap- those Apple Intelligence themselves over the last couple of years, and then just wound that effort [00:46:00] down because it was just too hard.
And so they, this is interesting that they’ve come back, for a second go. And I, on the one hand, you applaud the architecture but it’s not, as I understand it, open source. And so- -you know, relies on audit and attestation under contract basically.
Which makes me, just a degree skeptical that, that you- Yeah ... you have the shiny Apple brand promise. But, you’re still basically inside a walled garden.
**Tom Barraclough:** Yeah, definitely. And I think, like one of my other takeaways from watching the keynote, if anybody goes and watches it, is there’s a very clear regulatory structure to it.
So essentially like phase number one we have a closed ecosystem, and that is good because of what we’re about to show you and that’s really targeted at the European Commission and the Digital Markets- Yes ... Act and interoperability. Phase two is, here is all the safety infrastructure we’ve built in at an architecture and code level for child safety.
And on the one hand, that’s [00:47:00] fantastic in terms of safety by design. On the other hand, you gotta think about how powerful that is.
**Ben Reid:** Yeah.
**Tom Barraclough:** So they, all the stuff with like app marketplaces where they can basically refuse access to their walled garden. And then three is the payoff. The payoff is basically building on the safety.
We’re gonna have this pervasive AI system that is threaded throughout your entire thing. And but this is why you should trust us with all of that. So I looked at it really through, not only like a law as code lens, like a code as law type lens, but then also through a kind of regulatory frame, and then thinking also about this kind of sovereignty over AI thing as well.
And I just think as a package it just really... I was really fascinated by it, I think.
**Ben Reid:** It’s almost this post nation-state regulatory system, right? And so-
**Tom Barraclough:** Yeah ...
**Ben Reid:** how can a government that, going back to what we were talking about at the very beginning, a country that runs itself on PDFs [00:48:00] and- theatrical/totally boring select committees and , reading legislation in the House. How can that hope to even start thinking about regulating- A, multi-billion, trillions of dollar company that has a closed technology ecosystem. And direct access to the consumers, to the citizens in that.
Yeah, China did it. China’s, so basically restricts- No problem ... sales and restricts and puts certain rules around it, but smaller countries I think are gonna be really challenged to-
**Tom Barraclough:** Definitely. I am optimistic, and I think that’s part of the value of what we’re offering with Syncopate and with DocRef and this concept of digital regulatory infrastructure.
There are some very smart people in government or very-- There’s a lot of very smart people looking at government and going, “It doesn’t need to work this way.” It’s gonna be a challenge to make that shift, but I do think there is a real moment now where [00:49:00] the senior higher-ups that are normally very risk-averse can be motivated to care about the digital infrastructure component of everything because of the AI payoff.
So- The question is gonna be how do we do that well and in a way that’s consistent with the public interest and has appropriate sort of risk frameworks around it and stuff like that. But that is what we’re building towards with Syncopate. Some of the implementations of the system that we’ve got I’ve used to basically take there’s a set of API guidelines, so they meant to shape the way government builds APIs for- ... cross, cross-agency transfer. And they’re really long and quite just educative and and not that helpful for some people. What they wanted to move towards was a standard that was gonna be much clearer and just say, “You must do this, must not do this, should do this, shouldn’t do that.”
And what we did is, we went through a kind of policy exercise where we consulted carefully with people. We did a lot of reading. We did a lot of thinking. But then also at the end of [00:50:00] the day, we published the API guidelines on DocRef, so they were structured data sets. We looped that into an MCP server.
I got Claude to be able to search that MCP server in a sort of graph RAG setup- ... and then perform a lot of sort of retrievals, and then basically write a new API standard from that. And the benefit of it is that you’ve got this API standard but it also has these little traceable citations back to the source.
So you can go and check where it’s come from. And that led to some pretty surprising examples where I was reading it and going, “This sounds like AI b******t.” But then you’d go back to the source and be like, “Wow, that is a more or less a direct quote from the API guidelines.” Oh,
**Ben Reid:** there’s a moment.
**Tom Barraclough:** But then if you think about how quickly you could do that, right? So imagine if the entire regulatory infrastructure of New Zealand was on DocRef or a similar product- ... and it was all accessible to AI systems like that. You can pump out a new draft in an afternoon for something that would’ve taken literally nine months to do-
**Ben Reid:** Whoa
**Tom Barraclough:** And might never have been done. Yeah. And then there’s the- So we’re gonna start redoing that [00:51:00] over and over again, basically.
**Ben Reid:** And then, that experience we’ve got now of working with a coding agent like Claude Code or Codex or something. And so it’s not even in an afternoon, you’ll basically spin up, 30 different alternatives, Yeah
In a few minutes. And again, how do you increase the cadence at which legislation proceeds? And I suppose there’s another dimension of, legislation needs stability traditionally, right? Yeah. And so you, you need to... In order to get that consensus certainty, it needs to stay the same for a bit.
So that’s a constant tension. Oh my God, so we’ve gone around the houses on that and on, on a lot of topics. I guess to finish up, the question I’d have for you is, the work that you’re doing, what makes you optimistic about the future AI sovereignty of small countries like, like us, like New Zealand and, individuals within them?
**Tom Barraclough:** Good question. I think what makes I think the thing that gives me [00:52:00] optimism about AI sovereignty for New Zealand is that when we encountered social media for the first time, there was this really gradual process where we needed to think through and articulate what values do we want this thing to reflect, and do we really understand how this works?
And have we really thought through the implications of h- of how this is gonna change the world for the worse as well as for the better? I think another thing “We don’t really need ethical principles.
We need like case studies.” Yeah. ‘Cause, ‘cause it’s very easy to say that AI should be just, fair, true and whatever. That’s easy. What we need is like what does it actually look like-
**Ben Reid:** Yeah ...
**Tom Barraclough:** to manifest those principles in a system. And I think we’ve got a really thoughtful framework now for thinking about how do you embed human rights and values into the infrastructure of a system and the way it’s designed and deployed.
And so I’m quite optimistic, I think, that we can just act on all of that now. [00:53:00]
**Ben Reid:** Okay.
**Tom Barraclough:** Like I’m a really big believer, and part of the reason that I’m working in technology now rather than just policy is like it’s very easy to sit back and go, “Should do this, shouldn’t do this,” all that kind of thing.
It’s much, much harder to then pick it up and actually go and do it, and I’m really focused on going and doing it now. So one of the things we’re exploring is we’re probably gonna have a Syncopate workshop,
so we’ll pick up a set of regulation and we’re gonna use AI coding tools to basically go from start to finish on designing the API for the consumer protection data in New Zealand that lets people just, take their data between banks. I think being able to then harness the tools to manifest the benefit of it and test stuff and iterate faster I think is also quite exciting.
So particularly if we’re thinking about sovereign AI as being about control and agency and understanding and literacy, I think we should be pretty well equipped to start doing that. And part of that [00:54:00] is also that, if you give people enough knowledge, they can teach themselves stuff too.
So I’ve more or less taught myself how to code through basically just extended conversations- Not conversations, in the chat Yeah By saying, I still don’t get this. Give me an example.” And I’m quite excited about the opportunity for people to be able to do that, so long as they know what they’re doing.
Yeah. All
**Ben Reid:** right. That’s a good positive note to end up the conversation. Yeah ... thanks very much for your time, Tom. Really good to connect reconnect, Yeah ... in this context. And yeah, look forward to following what you’re doing in Syncopate. Thank you. Thanks, Ben. Bye.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit memia.substack.com/subscribe
