We can't afford to keep protecting the wrong thing. UK knowledge sovereignty, the long version.
1. A pattern: mobility, concentration, gap
Over a hundred years ago in the UK, capital centralised. It moved from local banks to London, creating specialisms and concentrations, and it meant that anyone using money or financing to build a business in the decentralised un-centre suddenly found themselves up against the 1910s equivalent of 'computer says no'. The decision-making power became centralised, rules rather than trust covered lending on small amounts the centre couldn't assess economically, and a funding gap had been left in its place. So went the Macmillan review of 1931, which identified this gap. It has remained and been re-named many times since: the Macmillan gap, the equity gap, the growth capital gap, the valley of death, the funding escalator gap.
Things that were less mobile became more mobile. Mobile things that could benefit from economies of scale concentrated.
More recently, and particularly since the global financial crisis, this tendency for mobile capital to concentrate is seen in public equity markets. The ECB found evidence of precisely this dynamic in Europe, where fragmented and thin EU markets lost out to US exchanges that have become increasingly successful at attracting both domestic and foreign listings [1]. You can see the impact in the numbers. In 2006, the EU27 and UK together accounted for 25.9% of global listed-company market capitalisation, compared with 37.4% for the United States. By 2025, the European share had fallen to 13.4%, while the US share had risen to 43.7% [2]. Now (well, by 2022) the ECB found that the average US-listed company had 3.3 times the market capitalisation of its EU counterpart [1]. Mobility alone doesn't cause all the change, but you can see the dynamic: where things can move, they move to where they perform best. It is a cumulative and not self-correcting process of divergence.
So far so depressing for us 'bottle-cap' Europeans.
2. Data went the same way
Something else happened in the last thirty years or so. Our data became mobile. Our information, who we are, what we worked on, where we were born, what we documented, read, liked, shared, wrote, presented, and watched suddenly became mobile. The platforms on which we consumed, contributed and stored content became digital and therefore movable. Once again these data have concentrated mostly in the hands of a few tech companies in a few highly connected centres around the world. That data has had a relatively 'small' footprint: storing petabytes of data is cheap, whilst meaningfully accessing it is a relatively expensive but infrequent event. The only systems that understood enough to use the data were usually the humans that stored them there.
Those humans worked at companies, institutions, and governments. Organisations like these have backup systems, requirements for encryption and choice over storage and processing of their data. So far so ISO27001. That is, secure against most illegal actors, and less secure against inquisitive state-sponsored ones.
We personally have some sense of that data being valuable. As individuals we're willing to give it up for the trade of an easier, less directly costly and more entertaining life. We also saw the value in this (the emergence of 'if it's free, you're the product') and eventually some states and governments acted.
Governments protecting their citizens, and given western sensitivity towards personal information like medical data, eventually put in place actual protections for some categories of these data. The result is that some of our data has some protections. For example the EU defined categories and some frustratingly vague requirements, the US has enabled healthcare data to be shared 'safely' and varies individual data treatment on a state-by-state basis, India requires ICT system logs, not health data, to sit in-country, and only for a rolling 180 days [11]. Broadly it's a patchwork mess of international rules and regulations that I'm pretty sure irresponsible companies ignore and responsible ones devote significant resources to trying to comply with.
As companies, research institutions and governments, the transition to cloud-based has brought great efficiencies (though the cost of the cloud has never quite been one of them) and convenience. The threat remained benign whilst the cost of storage and access to that storage remained sufficiently low that a data centre here or there could ameliorate most concerns.
Honourable mention to UK Biobank: data from 500,000 Britons has underpinned hundreds of patents, while UK Biobank itself retains no claim on the resulting IP. In 2026, de-identified data covering the whole cohort also appeared for sale on an Alibaba marketplace in China [9]. Data moves. Value moves with it.
I'm of the opinion that most categories of our personal data are of little value in the long run. It feels extremely valuable – it's our personal information, it's information about me! But looking at our US cousins we can see that whilst we find it shocking that you can find someone's tax bracket, voting intentions, credit score and other highly invasive data points, the 'market' values of these data vary from a few cents to a handful of dollars (as measured by the price charged by the firms providing it). There is an open question about what information is actually valuable at the national level, and what is valuable and necessary to protect. That should depend not on our sensitivities but on our collective value, and values.
3. Knowledge is the thing that is actually worth something
Knowledge is of tremendous value and we value it very little.
Knowledge is our know-how, our 'intellectual property', both private and published. It is the output of our human creative processes. These are the things that we do and create, and once created we fail to protect them, either as individuals or as a nation. It is data, understood.
The convenience of dropping a PDF into an LLM to ask for any grammatical errors, things that are missing, or to summarise, is from the same family as watching free video for the sake of some ads and knowing somewhere subconsciously that your decision to do so has been logged in the bowels of a large company.
When I finished my book I didn't hesitate to put the final version into a paid-for LLM, with all the data sharing turned off, and ask it to tell me that it wasn't shit. The reassuring answer came within a few minutes. No human could have looked through it faster or given me a more reassuring message. My impatience, my insecurities and my neediness were sated.
But it's not just one person putting something in. It is each and all of us putting in our knowledge, what we're working on, worried about. It is each nation's entire copyright library, its protected knowledge. The live-and-living corpus of the human frontier of knowledge. We do so because it helps us get more done faster and no one dares get in the way of speed. It's a fair trade when we don't think beyond what's going on under the hood.
4. What's different this time: information is not knowledge
It isn't just that this information is sitting somewhere. It is that these conversations, our revealed preferences, our pre-formed thoughts, our knowledge, are indexed. They are searchable, digestible, accessible, and exploitable in a way they have never been before. They are potential training fodder for the frontier lab you gave it to. How is this different to a Google doc sitting somewhere in the cloud? Compression.
Compression is what an LLM does to language information. It reduces it to the point where it is understood, navigable, part of the LLM's processes; its weights. Concurrently these same tools can sift and evaluate data, contextualise and ascertain its value. That is fundamentally different to terabytes of useful information sitting within exabytes of dross, only useful to those who can find and then comprehend it. It is that information made living, visible, searchable, explainable unto itself. Uploading a document doesn't put it in the weights by itself; that depends on the product, the settings and what you agreed to. But that is the direction of travel, and the settings are theirs, not yours.
This information. All our knowledge. Cutting-edge research and our leading creative output are usable, at near-zero cost, elsewhere. The useful and valuable stuff is being made available within private models. The cost to access the data has plummeted; this compression is what turns the data into accessible knowledge.
The objection writes itself, and it's the one we heard about data: a copy is not a theft. The researcher still has the note. The model has a photocopy, and a photocopy takes nothing away. That was true for a hundred years, because a document's value sat still until a human read it, and humans are slow. A rival who copied your work still had to catch up page by page. Slowness is what made copying relatively harmless, but still warranting some protection ©Matthew. What changed is that the copy now runs. A model trained on your working note doesn't have the note; it reasons with it, combines it with everything else it was fed, and produces the next design, the next model, the next paper. That output goes to whoever owns the machine, with no obligation to share back what it unlocks. Knowledge was non-rival while it sat still. It stopped being that the day it started running. You kept the document. They took everything that comes next. Text is no longer static; it is intrinsic to its own execution.
In September we had the Navier-Stokes solve, a very hard maths problem [12]. The mess that followed was that the run (billions of tokens) OpenAI used ended up creating a solution that looked a lot like someone else's private research [5]. OpenAI says neither its researchers nor its agents saw that work before publication, and that no user data was touched [12]. Take them at their word. It is still a simple slip of an internal process to accidentally ingest the wrong training data. To hoover up the secrets along with the public discourse. That is, assuming you can really think that people understand the extent to which their conversations and knowledge may be shared within these companies.
It is not just the world of science. We as individuals can already perform acts that would have been considered audacious copyright theft a few years ago. At a keystroke. Using off-the-shelf models. Ask an open source model to generate an image of multiple in-copyright characters, make a video of a celebrity for a short film and score it. Job 'done'. All leveraging the knowledge that we have given away.
Now let's do a simple thought experiment. Imagine you're in a secret agency, in a country with access to all the world's questions and knowledge. Sure, it's stored in private companies with privacy policies, but would you take a peek? I would certainly think about demanding, secretly, to see the inside of these systems, the inside conversations, the frontier of global knowledge. There are some obvious public security grounds, and if you discovered a breakthrough made elsewhere, you would hardly be a patriot if you didn't make sure someone in your country knew it too. The frontier of all human thinking no longer sitting in people's heads, but upped and moved to the data centres in Arizona.
The risks have only started to become clear. This isn't data anymore. It is our knowledge.
5. What's different this time: feedback loops
Knowledge is one thing. These systems can influence our creative processes and infill gaps in our own understanding of the world around us, and do so faster than us. I've written elsewhere about these discovery 'loops' and the threat they pose [7]. That was about who stays in the loop. This is about who runs the loops and where they live: their de facto and de jure masters, and who controls what they let us pesky humans do.
If we are using an LLM to come up with a design for a new widget, is your LLM as good as your competitor's? What if the tool you are using has been deliberately crippled because its home country has designated it a strategic capability? How would you know? You wouldn't. Your thinking would just be slightly hobbled, and forever slightly behind your competition. Sounds theoretical? The risk is here:
Example: Anthropic's current frontier model ships in two versions. Claude Fable 5.1 is the one anyone can use. Claude Mythos 5.1 is the same underlying model with the safety restrictions on dual-use capabilities removed, and it is available only to organisations Anthropic approves. Same model, two tiers of capability, and a private company in California decides which tier you are in. That's one we know about, because they published it [6]. Another we can only guess at is how helpful these systems are when asked how to build LLMs themselves. I have no evidence of deliberate dumbing down, but the incentives are monumentally clear: why would Anthropic or OpenAI make it easy for anyone else to build their own?
And it isn't only what these systems will and won't do for us. As we share tasks with AI we change what we require of ourselves, and what it means to think as humans. That argument is made in full in the discovery post; I won't repeat it here. Suffice to say that the feedback loops being created can widen small differences in capability until there are insurmountable divergences between what one person, company, country or system can achieve and another.
We have yet to even come to grips with the preceding technological wave of the internet (social media) and converged devices (as smartphones were called in the 90s). Now we are being hit by the next, which is infinitely more insidious, powerful and convenient. Our interactions with AI shape us. And the training data, the way in which it behaves, and what we create with them are expressions of our ideas, creativity, and knowledge. It is our thinking and it is something that is profoundly and fundamentally cultural. The tech bros of Silicon Valley are the unwitting cultural imperialists of the 21st century, and the knowledge they construct their systems on, how they construct our systems, and how they provision them shape how we think. Our minds.
6. Knowledge and thought are the sovereignty questions
So why did I start by talking about the funding gap? Because it's a hundred-year-old example showing that when things are mobile they concentrate where there are the greatest scale effects. And the gaps that are left behind are permanent.
Knowledge is the next frontier in sovereign governance. We already gave away most of our personal data. Our knowledge has gone the same way, through LLMs, at an astonishing pace that has caught most governments, people and technologists napping. But we are making more. Thinking more. And we are building new things. The challenge is how to nail the knowledge jelly down without losing our minds in the process, and before the knowledge we can generate becomes itself worthless as we are divorced from the technological frontier.
That sounds a lot like I'm screaming STOP THE WORLD I WANT TO GET OFF. But what would getting off that world mean? It would mean cutting off any model whose training data we can't trace, walling off all the UK's sovereign data from training and analysis, then building institutions and systems to do that work ourselves and using them to operate at the frontier. A great plan if we'd started 20 years ago, but an utterly impractical and destructive step today.
We can't cut off international AI. We need it to move, and to think at the technological frontier. The race to build started years ago and we turned up late. That's a fact, not a verdict.
And here's the thing about the verdict. Defeatism is not a neutral observation; it is the sales pitch. If you run one of the best models in the world and you are selling into the UK, the story you need the UK to believe is that the gap is unbridgeable, the moment has passed, and the sensible thing is to buy from you and stop thinking about it. Every 'the UK can't compete' piece is free marketing for someone. Look at what actually matters and the picture is different. The talent is here. The knowledge, which is the point of this essay, is here and we're making more of it. The open models are a year behind the closed ones, not a decade. The one place the gap is real is national compute, and that is a planning and energy problem, which is section 7. And even that is a first-generation gap. The compute that exists today is the first real AI generation; the chips, the architectures and the energy that run the next will be different, and most of the infrastructure bets get made again. A good chunk of what is being poured into concrete in Arizona right now will be a sunk cost inside the twenty years that matter. We are at the start of the next twenty years, not the end of the last five. On the timescale a country should care about, the timescale of institutions and grids and the next three generations of graduates, none of today's lead is decisive. The gaps that become permanent are the ones you accept.
Nor have the rules been written. Compute and capital have concentrated; that race is largely run. But who owns knowledge, on what terms it can be trained on, who gets to see what a model was built from, whether its draw on resources is visible: nobody is leading on any of it. The EU leads on slowing itself down, the US on slowing everyone else down. Knowledge governance is an empty chair, and empty chairs get sat in. The first country to build a coherent framework for sharing knowledge into training, with money flowing back to the people who made it, becomes the place that activity happens. That is convening power of a kind the UK has used before, and it doesn't require the biggest data centre.
Here’s where I sound like I’m deliberately contradicting myself when I’m not. There’s a difference between Innovation Diffusion (~adoption) and Innovation generation (~science and ideas): Strategically behind, and aligned, on adoption: buy good-enough, buy it on our terms, and don't lose sleep over being a year off the frontier in what the public sector uses. Strategically ahead, and investing, on capability: the talent, the research base, the energy, the compute, the institutions that let us build at the frontier ourselves when it matters. Behind on what we buy. Ahead on what we can build. Get those the wrong way round and you buy the frontier and own nothing, which, if you look at the last two years of procurement, is roughly the plan.
The guiding principle, though, is clear: the benefits of knowledge should remain with those who made it. That does not mean it never leaves the country. It means it never leaves without those who made it getting a say in what it becomes, either individually or as a society. Sovereignty here isn't a flag on a data centre. It is the ability to set terms, enforce them, and keep real choices about the systems we depend on. Location and ownership help; neither is enough on its own. It is the government's role to protect its value and our say.
Like all good-sounding sweeping statements, it's problematic when you start thinking hard about it. The first problem is that the other side doesn't have to play ball. As the music industry has found, creators, even when organised, have had only a modicum of impact trying to use their existing rights [10]. The next is that if you segregated data you'd lose speed and economies of scale, quality and speed, in a competition where those are the two things that matter. The best models will be built from as many countries' knowledge as possible. Likewise, regulation to slow things down will only slow us to the point of obsolescence unless it is done in concert with every country at or near the cutting edge.
Pooling is not the problem. The problem is the current model, where one or two companies from one country do the pooling and everyone else powerlessly donates everything they've ever known.
No national regulation settles an international problem, and slowing down in a race to expand humanity's capacity to 'think' risks being overtaken by other 'thinkers'. But the UK controls two things outright: what the public sector buys, and who gets access to publicly held research data. That is negotiating power, and it can be used now while international agreements are worked out. There are a few tools that might help. That starts with being clearer about what knowledge is and treating it with the value it really has.
Regulation: The instinct is to define UK knowledge and catch it on its way into a model. That's the hard problem, and right now it's intractable: nobody can tell, at the point of upload, whose knowledge a document is. It is a fine thing to fund research on and a terrible thing to build policy on. So don't. Move the burden to the people who can carry it: the ones who built the model. Any model sold into the public sector discloses to the regulator, not the public, where its training data came from, and gets audited on it. Then a do-not-train register for UK knowledge, on by default from a fixed date, twelve months after the policy is adopted, and not retroactive: nobody is asked to unlearn anything. That is the equivalent of an automatic opt-out for all citizens and firms that trade from the UK. That is a serious sticking plaster that creates real problems for everyone. I think that's what makes it a good first step to iterate from. The problems it creates are worth solving, because solving them is what lets us open up our knowledge later, as a trade: influence over training and access to the frontier models that use UK data, in exchange for the data. We are currently working from the opposite extreme: our knowledge is readily offered up with no hard power of oversight or protection, a bad place from which to take small steps.
Be clear about whose side that is. The text-and-data-mining exception the government proposed in 2024 ran the other way: training allowed unless rights holders opt out. This takes the creators' side and generalises it from copyright works to all UK knowledge. Individual creators tried this against the largest companies on earth and got a modicum [10]. This is the state, as customer, with a register. Different fight.
Purchasing: The government is adopting AI everywhere it can. This is good for productivity and terrible as national strategy. The argument is that an anchor customer doesn't need the best model in the world. It needs a good-enough one, bought on terms it controls. And good-enough has never been cheaper: the cost of training a given level of capability falls at a rate that makes last year's frontier this year's open-weight download. A model a year behind the frontier is still astonishing to anyone in 2021, and for drafting a planning letter or triaging a benefits claim it is more than enough. So the £1.4bn shouldn't be buying frontier access on the frontier lab's terms. It should be buying from UK companies that take open-weight models and build the harness around them: the orchestration, the tooling, the domain knowledge, the bit I argued earlier is where the products and the money actually sit. Government gets a model it can inspect, running where it says, on terms it wrote. The UK gets a supplier base that knows how to build on open weights, which is a skill that compounds. And the money stays. You might scream 'but the UK will be behind in AI adoption'. Bluntly, if the public sector only lags the commercial world by five years in adopting a technology I'll be delighted. Adoption lag is not the risk. Dependency is.
Institutions: That's the short run. The longer-run structure is the one governments are actually good at building. Why not make weights the output of collaborative research institutions, national and international, paid for by those who own the downstream products built on them? No new money is needed: point the roughly £1.4bn the government already spends on AI at models built this way, with government as first customer and private downstream owners following [13]. That's a long shot, but shouldn't be seen as unrealistic. We have the talent, and the gap to the frontier is months while our decisions take years. To play our own hand in it, the UK has to have something to contribute, and the thing we have most of is our knowledge.
Technical capability: Machinery for sharing the nation's knowledge with those training models, under public editorial guidelines, run by a body with the technical depth to negotiate. I set that out two and a half years ago in a piece on AI in public services: a national AI body, something like a BBC for AI, a trusted data-sharing system for training on the nation's data, and a requirement that a fixed share of a model's training data come from the country it is used in. It looked like a public-services question then. It is a sovereignty question now [8].
Who does what, because 'a body' is where these things go to die. Two failure modes to design out. A department can't be the regulator: ministerial teeth go when the minister does. But a regulator can't set spending policy either; Parliament controls supply and ministers answer for it. The split that already runs across Whitehall does the job (NCSC sets Cyber Essentials; departments must buy from suppliers that hold it). Government keeps the decision to buy. An independent body sets the standard a purchase must meet. Specifically: the AI Policy Directorate in DSIT owns the policy, that public-sector AI purchases must meet the standard, and the date. The AI Security Institute sets the standard and audits against it, on a statutory footing with the standard in secondary legislation, because guidance gets waived in a meeting and a statutory instrument leaves a paper trail. The CMA takes vertically integrated AI stacks under its strategic-market-status powers; that's a competition question and it stays separate. A department can still buy a non-compliant model, but it publishes why and the regulator publishes an annual table by department. And the sanction lands on the supplier, not the department: fail audit and you're out of the whole public sector. Government can still get this wrong. It can't do so quietly, and it can't do so for one supplier at a time.
Is this protectionism? No. Am I using a rhetorical device in an irritating fashion because this essay has got too long? Yes. But, this is not protectionism. Protectionism is a wall. This is a trade with terms, and the terms are negotiable; we are currently at the opposite extreme, giving everything away with no say at all, which is a terrible place to negotiate from. Unenforceable? It's a contract term. The sanction is losing the customer, and the customer is the whole public sector. Providers will refuse? Then they have told us exactly what our knowledge is worth to them, and we should take the hint. It isn't a wall going up. It's a market starting.
All this requires an AI strategy that isn't based on adopting LLMs and encouraging the frontier. It means understanding the nature of knowledge, our creative processes, and where value lies in the future we're creating. That requires real investment, experimentation, trade-offs and capabilities.
7. Which means energy
For me this means one thing. A dramatic rethink of energy infrastructure. If that seems like a big jump it isn't. This isn't a new thought; those building AI have been thinking precisely that for years. AI systems are competing with humans for resources. The resource they need to run is energy. More demand means a higher price for any given level of supply. A high cost of energy or energy uncertainty means no incentives for long-run investment in data centre infrastructure, neither local nor hyperscale. It doesn't matter at what scale AI runs in future: whether an always-on Claudebot agent lives at home, on a device, or in the cloud, energy is the critical input for AI, and AI is driving up demand.
We need to cope with the new category of energy demand that AI represents. We are going to consume more. If we export this energy demand, we export the knowledge with it.
In order to build a sovereign knowledge strategy you need sovereign compute capacity. To have sovereign compute you need energy. Those are the shoelaces that need tying. Want to catch up in AI? Maybe nationalising the energy sector in the UK is a good place to start?
Less flippantly: this is a grid-planning problem with a twenty-year horizon, and we have the pieces. Solar is the cheap marginal megawatt and gets cheaper every year. The baseload is the harder bit, and the honest framing of the options matters. Small modular reactors, Rolls-Royce's in particular since the state has already picked them, are proven physics with unproven delivery: nobody doubts they work, the question is whether we can build them on time and on budget, which is a question about us, not about physics. Fusion is the reverse: unproven physics, and AI-driven discovery may well get there faster than anyone expects. I'd love that. I would not bet a nation's grid on it. SMRs plus solar gets there on a timetable a government can plan around; fusion is the upside, not the plan.
And the bet is robust either way. On the off chance that AI plateaus, you still have cheap energy. If Anthropic and OpenAI turn out to have no moat, you still need compute. The lower the basic cost of living, for us and for machines, the more will get built here and the more we will have a say in building our future.
Hungry data centres will demand energy and they are doing so already. There is one private loop we must not allow, and it is this one. If a hyperscaler has its own dedicated power supply, then there is a private AI loop competing directly with humans for resources without ever being detected, seen or regulated. They have to drink from the same fountain. They have to use public energy supplies. Battery, yes; backup generators, ok. But the energy comes from the grid. If a data centre wants to build generation, fine: build it, and connect it to the grid. It sells in and buys out like everyone else. Then a hyperscaler that turns up with a gigawatt of demand also turns up with a gigawatt of supply that the rest of the country can use when the training run finishes, and the country's generating capacity goes up rather than being quietly ring-fenced behind a fence in a field. It is an immutable transparency step: the one place we can always see how much they are distorting the human world. That's my definition of sustainable AI. Not AI that is cheap to run, but AI whose draw on human resources is visible.
Three more things while we're here, because power is not the only fountain. Water. Cooling a large data centre draws on the same rivers and aquifers as farms and taps, and in the parts of the country already water-stressed that is not an abstract concern. Same rule: draw from the public supply, metered and visible, and no private abstraction licence that nobody can see. Siting. Where these things go is a planning decision and should be made like one: near generation and water and grid headroom, not wherever the land was cheapest and the council keenest. The grid connection queue is already the binding constraint; use it as the lever rather than the excuse. And grid response. A training run can pause. A data centre is one of the few multi-hundred-megawatt loads that can drop off the grid in seconds when the wind stops, and that flexibility is worth real money to the system operator. Make it a condition of connection, pay for it properly, and the thing that was going to break the grid becomes the thing that helps balance it.
8. Close
The BBC was formed for the nation's airwaves. Really it turns out it was there to provide quality-based competition for content that was delivered into our nation's minds. Instead of reworking our institutions at the turn of the 21st century for content, eyeballs and attention, we let social media run rampant.*
Compute is necessary. It is not sufficient. Owning the machines, and the energy that runs them, gets us a seat at the table where the weights are set and a view of what they cost us. It does not decide what goes into our heads. That is an institutional question, and we have answered one like it before: not by regulating the airwaves, but by building something good enough to compete on them.
We have moved cripplingly slowly, waiting for our understanding of the last technological wave before moving to protect our data or our minds. Now it is our knowledge and how we think. The stakes are higher and the rules are still unwritten. Act, make mistakes, and improve. Most of what this country will know is still ahead of it. The best time to protect it was twenty years ago; there is nothing wrong with now.
Notes
* Yes, this can be done. One simple policy would be to require screen time on applications that algorithmically provision (social) media to those between 12 and 16 to have 20% of that content be 'educational'. Educational would need to be certificated in some way, e.g. by the BFI, Ofcom, or even self-certificated. One could allow the BBC to unlock its libraries of educational content for social media platforms and create a new route for publication. Regulators, institutions and content providers (like DCMS, Ofcom, ASA, BBC, C4 etc.) are already in place to tackle these issues with a re-shaped technological focus.
Sources
[1] ECB, Examining the causes and consequences of the recent listing gap between the United States and Europe, Financial Integration and Structure in the Euro Area, 2024. https://www.ecb.europa.eu/press/fie/box/html/ecb.fiebox202406_07.en.html
[2] SIFMA, Capital Markets Fact Book 2021 and 2026 editions. https://www.sifma.org/wp-content/uploads/2021/07/CM-Fact-Book-2021-SIFMA.pdf ; https://www.sifma.org/wp-content/uploads/2025/07/SIFMA-Capital-Markets-Fact-Book-2026-Edition.pdf
[5] Wikipedia, Navier–Stokes priority controversy. https://en.wikipedia.org/wiki/Navier%E2%80%93Stokes_priority_controversy
[6] Anthropic, Claude Fable and Mythos 5.1. https://www.anthropic.com/claude-fable-and-mythos-5-1
[7] Matthew Cleevely, The Human Stake in Accelerating Discovery, January 2026. https://www.matthewcleevely.com/notes/the-human-stake-in-accelerating-discovery
[8] Matthew Cleevely, How to shape the development of AI for public services, January 2024. https://www.matthewcleevely.com/notes/how-to-shape-the-development-of-ai-for-public-servicesnbsp-january-2024
[9] BBC News, UK Biobank. https://www.bbc.co.uk/news/articles/clyedyn6pz7o
[10] Ed Newton-Rex, Statement on AI training. https://www.aitrainingstatement.org/
[11] CERT-In, Directions under section 70B, 28 April 2022, paragraph (iv). https://www.cert-in.org.in/PDF/CERT-In_Directions_70B_28.04.2022.pdf
[12] OpenAI, On the Navier–Stokes Millennium Prize Problem, 8 September 2026. https://openai.com/index/navier-stokes-solution/
[13] UK ~£1.4bn expenditure on AI https://www.tussell.com/insights/ai-procurement-tracker?utm_source=chatgpt.com
Image Credit: ChatGPT (OpenAI)
Edited with (line editing, and red-teaming) : ChatGPT, Claude, and Qwen