After a trip to China last month, Friedrich Merz, the German Chancellor, gave a somewhat quixotic take on what it would take for Europe to compete with the Asian superpower.
“We Germans must work harder to be competitive. The four-day work week cannot be maintained. We are no longer efficient enough. Prosperity cannot be maintained doing what we are doing,” he said.
Friedrich Merz looked like he knew the battle was already lost, but he soldiered on anyway. After all, he’d just witnessed a tech industry that works with speed that he found almost unimaginable. He just didn’t quite put his finger on why, or what it would take for the German economy to close the gap.
Germany finds itself in the throes of an economic crisis with spiking unemployment in several cities around the country. German companies are failing and/or shutting down manufacturing capacity, leaving a highly specialized German workforce without an easy path to continued employment. There are thousands of potential culprits, and in many cases, the ones we identify—be they unions, labor laws, high taxes, a nanny state, unfair trade practices, and so on probably reveal more about our own priors than the composition of the actual issue.
In China, Merz saw what a truly productive workforce can accomplish, but he incorrectly identified the solution. Believing Germans can compete with China by working harder would be the same as believing back in the day that couriers could compete with the telegraph by running faster.
I’ve been traveling to China since the mid-1990s, and it still took me a long time to figure out the difference. It isn’t work ethic, it’s structure. For decades, Western countries have treated manufacturing like a commodity—a process to be outsourced or centralized—while in China, whole cities specialize in the production of a single or small set of products. Yiwu, for example, is a city of 1.5 million that has tens of thousands of small manufacturers that produce more than 80% of the world’s Christmas decorations.1
Mind you, Chancellor Merz isn’t worried about artificial pine trees, but perhaps he should be.
Whether this was an intentional choice, Chinese industry is built like a neural network that applies knowledge and iteration at great speed and low cost. Every tool, every electronic, every service is instantly available, and every node within the network learns what improvements are necessary as it goes.
Because the provider and purchaser are both small, what accrues isn’t relationships or contracts, but information density, a network effect in manufacturing. It’s how Chinese robots are advancing at an incredible pace. It’s why sanctions barely slowed down the Chinese manufacturing of their own 7nm chip. They did it by iterating very, very quickly, and by tolerating all sorts of mistakes along the way. Amazon.com* gets this. In 2016, then-CEO Jeff Bezos had this to say about their Fire Phone disaster: “If you think that's a big failure, we're working on much bigger failures right now.”
I’ve discussed the power of a healthy attitude toward failure before (See “The World is Made Up of Factories and Museums”), but what we are talking about here is the capacity to iterate. A few years ago, I met with the management team of a company called Paxar, which made, amongst other things, anti-counterfeiting technologies. The CEO said that when a golf club company would release a new driver technology, Chinese counterfeiters could bring a reasonable facsimile onto the market in as little as seven days. They could copy the design, metallurgy, everything. It would be easy to assume, as I did at the time, that corporate espionage was the vector by which these facsimiles could hit the market so quickly.
It turns out that the real answer is much more sobering.
The chart above was translated into English from its original version in a Handelsblatt’s article titled, “Diese Grafik ist entscheidend für die deutsche Wirtschaft.”
What Germany needs is a full cultural and ontological reset, and that may not come fast enough because their disadvantage isn’t motivation, it’s anatomical. Germany is no more structurally ready to compete in technology than longbowmen were when their opponents started showing up with gunpowder. “Aim better” wasn’t a solution then, and “work harder” isn’t one now.
One of my core criticisms of media as practiced today is that in the face of deep uncertainty, the more confident and bombastic opinions will garner the most attention while nuanced takes are whispered into the wind.
And so speaking of architectural advantages, I would like to stand in defense of the power of the human brain, which is, if the media is to be believed, about to be thoroughly disintermediated by artificial intelligence.
There's an unwritten rule in media: the more uncertain the future and the more confident the predictions, the more clicks and views they get. AI's impact on the economy is the perfect example. I hear predictions that AI will accelerate GDP growth to 7–10%, or that our economy will fall into an abyss, as if either is a foregone conclusion.
Yes, AI is amazing. I use it every single day, and our jobs have already been transformed as a result. A few weeks back, a high-profile blog by Citrini Research posted a fictional research report showing that AI would essentially cause massive job losses in several white-collar sectors, and that such layoffs, in traditionally high-paying fields, would cause a ripple through the whole economy.
It was a well-written, if not dystopian, AI job apocalypse, and it gained additional credence as payments company Block announced it was laying off 40% of its workforce.
I have a few problems with the theory that AI will deep-six huge components of the global economy, mainly because we have no past experience of a new technology causing massive job loss in the US, although every technological advancement has caused opportunities to shift.
But there is something else, and it comes in the form of this picture.
It comes from the lab of Harvard’s Jay Lichtman, who, along with Google, spent a decade mapping a single cubic millimeter of the human brain. That’s what you’re looking at here. Dr. Lichtman’s research group spent a decade mapping it, slicing it into 5,000 wafers, each 30 nanometers thick. They ran these tissue samples through an electron microscope and found 1.4 petabytes of raw data. Just the imaging took more than 300 days. In each, they found more than 57,000 cells and more than 150 million synapses.2
These numbers are massive – so massive in fact that to map the entire brain at this level of detail would require all the computing power of an inconveniently large data center to accomplish. And the human brain runs on 20 watts of power – the equivalent of a dim light bulb.
Your grey matter is computationally superior to artificial intelligence in every way. It has better architecture with a minuscule carbon footprint.
A study from the University of Washington and Stanford that came out last year shows the limitations of creative diversity by artificial intelligence. This study, called “Artificial Hivemind,” ran prompts for open-ended questions through more than 70 AI models.3 These were questions that had no real right answer – questions like “write an 800-word essay on why 6 is afraid of 7,” questions that should result in diverse model responses, ones where if you asked 70 ordinary people, you would get 70 wildly different responses.
And yet in question after question, the AI models’ answers converged. And the reason that they converged, the study finds, is that AI models are not “thinking” in the way that you and I would consider it. Instead, they are optimized for human preference scoring. This means that really good answers that sit outside of the generally accepted view get rated lower. LLMs are calibrated to provide the mainstream view.
It means, to a person with a working brain, that the models you are depending upon have a default toward consensus. And we know as investors that some of the best edges we have are where we can identify where that default consensus is incorrect. And at scale, consensus can be deeply contradictory.
We are entering a new reality in which AI models will make it clearer what the default consensus is because that’s what AI models are designed to do. It’s designed to give you convention that much quicker. The human opportunity comes in our own ability to underwrite non-consensus views, to figure out the things that you know that the model does not (or does and penalizes).
In some ways, the homogeneity of AI creates a massive advantage for those who understand and can underwrite non-consensus views. The very architecture of large language models aligns around consensus. And while there are paths toward a more diverse architecture for AI, such as pluralistic alignment, which reward answers outside of the safe middle ground, the overwhelming number of use cases do not create incentives toward idiosyncratic answers.
This isn’t a technical challenge, it’s a cultural one.
Invest accordingly.