After The AI Crash: how to protect European democracy from MAGA tech

What I found out at the Political Tech Summit

At the Political Tech Summit in Berlin, I had a fantastic discussion with David Bluestone and Tim Gordon, brilliantly chaired by Christine Brause. We were talking about AI, European tech regulation and US tech dominance.

Two things got me thinking. The first came from David just about Trump and the “Little Tech Manifesto”: the big tech oligarchs weren’t there to exert power, but to protect their position from disruption. The second from Tim, that the real value is in data. The third, provoked by Christine, that we need to separate tech dominance in general from social media dominance in particular.

And what a day for the debate. Just 48 hours later, $1 trillion had been wiped off US tech stocks after the markets finally discovered what had been known since May 2023 — large-scale general purpose AI companies have no moat. A moat is business jargon for the barriers a company can put up against competition. The moat can be access to resources, superior technical skill, or even, like with Coca-Cola, just branding. Just how dominant American tech — as opposed to American tech finance — actually is, will become clearer after the markets find their bottom.

Large Language Models suffer from their generality. They’re all trained on essentially the same text, and this gives them essentially the same ability (and flaws — they’re all vulnerable to the same kind of prompt injection attacks). This makes them, in the end, commodities. LLM companies have no way of distinguishing their product because they all do the same thing, and so have no pricing power. It’s easy for developers to switch from one language model to another. The only “secret sauce” is the weights that the trained models have. But Facebook released the weights to its llama models — knowing, I suppose, this would eventually undercut the competition. Deepseek, the new Chinese model, appears to have based its model on some of these open weights and trained its own to the same level, but for far less money, apparently as little as $5 million, than OpenAI and its rivals. It released the model’s weights publicly, so anyone with enough hardware (the model has 650 billion parameters) can use it on their own equipment. Then it released technical documentation allowing anyone else to retrain another model using its advances.

Find The Value

This brings me to Gordon’s observation about data. The optimisations Deepseek used dramatically cheapened the cost of transformer-based models (transformers power language models, but can also be used for other statistical pattern recognition), but while clever, they are not revolutionary. Faced with limited access to chips they thought how to make AI more effcient. Like the Japanese cars of the 1970s and 80s, they’ve shown American LLMs to be gas guzzling monsters. (It’s not a little embarrassing that not a single European AI company was able to exploit the Americans’ corpulence).

Their effect is even more dramatic. Language models are just one, dramatic and relatable, use of GPU-powered statistical pattern recognition. In principle they can be used to discover patterns in very differerent kinds of data. Many of the optimisations DeepSeek developed are so general they can be applied to other kinds of problems and AI architectures. Their work will dramatically increase the number of problems to which statistical reasoning can be applied. If you want something done, by the way, I recommend the brilliant Bulgarians at Vector Labs.

The market advantage comes from having access to specific data relevant to one particular problem, like detecting enemy drones so they can be shot down, optimising traffic flows in a large city or finding new drugs. Its acquisition and organisation provides competitive advantage. It is less clear that sheer scale matters as much as we had thought.

This should make big tech, whose advantage is scale, more than a little uncomfortable.

The CEOs found their way to the front of Trump’s inauguration, having donated what is to them pocket change to fund the celebrations. In Bluestone’s view they were there not to show their own power, but because they were afraid. They came to kiss the ring, not tell the boss what to do.

Though they probably won’t have to fear regulation, which would stand little chance of getting through Congress, they ought, he said, to be worried about competition. Another tech figure close to Trump is Marc Andreesen, of the legendary Silicon Valley venture firm Andreesen Horowitz, which styles itself the champion of the Little Tech Agenda. Little tech exists to dethrone incumbent firms grown old and fat:

A startup is what happens when a plucky group of outcasts and misfits comes together with a dream, ambition, courage, and a particular set of skills – to build something new in the world, to build a product that will improve peoples’ lives, and to build a company that may go on to create many more new things in the future.

Apart from relative age, these companies’ strength is in two businesses: cloud computing (Amazon and Microsoft) and targeted advertising: Facebook, Google and, if you can call it strength, Elon Musk’s X. But as the rise of Bluesky shows, their raw material is users’ attention, and it could easily be lost, just as Yahoo, and MySpace lost it to them. They need Trump to protect them from the Little Tech poised to overthrow them.

The cloud computing businesses are the utilities of the age. The social media companies are quite different. They’re companies whose great wealth generates huge externalities.

Make Information Polluters Pay

Like oil companies, whose main activity pollutes the air and drives global warming, social media fills the public sphere with toxins. Its sins are two: first, it’s currently designed to maximise engagement, and so tilts towards the tabloid; second the advertising revenue that used to support journalism, and cultural production now goes to these technology companies fundamentally uninterested in the media. Even notorious newspaper barons were interested in some sense in the media as a teller of stories. The tech bros couldn’t care less.

This was bad enough when the platforms were neutral, if lowbrow. But now some of our main sources of information are controlled by men dependent on Trump. Facebook’s Meta has swiftly joined his culture war against “DEI”. Jeff Bezos’s Washington Post is fast losing staff and credibility. X openly meddles in other countries’ politics, as Anne Applebaum shows. Tiktok is under the control of the Chinese government (and unlike DeepSeek, keeps its algorithms hidden).

These platforms now control a vast portion of information distribution in Europe. Their content is corrupting our public discourse, promoting division (because hatred is a form of engagement). They have become an engine of populism, allowing charismatic but shallow candidates to leap into power. That might not be too big an issue if charisma were correlated with the ability to govern, but it isn’t, and the ancient tension in democracy between persuading people to elect you, and then governing in the national interest is starting to tilt too far towards persuasion and against successful governing.

The old world of 1990s Western Europe isn’t coming back. Print newspapers are disappearing, and timed news bulletins will be watched more and more rarely. Decentralised algorithmic distribution is here to stay. But the social media model we’ve grown has started to endanger the survival of our democracy.

Europe, at least, is far from powerless to stop it. Even Meta recognises this, and is keeping fact checkers on in the EU. It’s scared of fines. But judicial processes are slow, and we should get more creative.

An obvious move is to make it easy for people to move to new platforms, like they can move between phone networks without losing their number. This was blocked by the Council when the Digital Markets Act was passed. The Act is up for review soon. Time to force them to allow people to move networks but take their “social graph” with them. Bluesky’s AT Protocol shows how it could be done in practice. Competition drove Deepseek’s to make its new model, and competition is needed to create alternatives to MAGA- and China-controlled social media.