Innovation, scaling and legacy
I spent last night reading and watching two interviews. Three inputs to my thinking set me off on a tear this morning at 05:00, which just happens from time to time.
The first input was with Daron Acemoglu. In "The AI boom is accelerating, but what kind of economy are we actually building?" Sinead Bovell asked him questions about his work, which spans political economics and the economic impact of technological change. He's not a must-read author because of the Nobel prize. His political ideas are overwhelmingly humanistic and agreeable, and the refreshing way that a person of his stature admits when they simply don't have the answer to a question is charming.
There's been an inevitability of assumptions baked into the AI buildout up until now, and only time can tell if we're headed in the right direction.
The second interview was with Zuzanna Stamirowska, CEO of the Pathway neolab. A neolab is one of the labs outside the heavyweights (Open AI, Anthropic, etc.) dominating the landscape as it stands. She and her startup colleagues are working on something called State Space Models , which is a different architecture than the transformer method that the models we are using today are built on. Several well-known limitations of GPT were presented, e.g., mathematical constraints in the context window and hallucinations.
Context is a hard limitation, we've seen an unexpected growth in the maximum context windows in 2026, but I'm not enough of an expert to distinguish how much is hype and how much of that is utilitarian or beneficial. From what I can gather, it is seriously subject to the law of diminishing returns. Now that the spring and summer of tokenmaxxing has hopefully waned, time to realize that spending tokens gives you an initial uptick in quality, but each extra token quickly adds less value because it raises your spend and cannibalizes accuracy. People my age are old enough to remember subway tokens. You had to spend money for them.
The Stamirowska interview led me to the paper1 the Pathway researchers published, which is about "a new Large Language Model architecture based on a scale-free biologically inspired network of n locally-interacting neuron particles". I had to do some digging just to understand what the state space was, and as I hazily recalled it as an old school machine learning technique (see IBM article linked in the previous paragraph).
Still with me? Excellent. My point is: you don't have to be a mathmagician or data scientist to understand these things. You're not going to be able to describe them, but understanding is doable. I've got some hard-won background knowledge gleaned from working on papers and doctoral theses as a copyeditor. And I've been using some form of neural network (NMT) as a translator since around 2009.
You don't need to understand these things to use them, in the same way that you don't need to be able to engineer a car to drive it.
Which brings me to the subject of this article. I have a feeling we are at a critical juncture with LLMs and AI, one that is similar to the juncture we faced when the automobile became the dominant mode of transportation in most of the world instead of light rail.
Economically, the use of energy models for both transportation modes is hotly debated. The camps are easily distinguished by political affiliation. Conservatives are rooted in the freedom and economic driver of automobile use and manufacturing. People on the other side of the political spectrum hail the advantages of moving more people around in larger vehicles, which presents an economy of scale and suggests an egalitarian use case. These demographics often are physically located in less urban vs. urban areas, respectively. Simply stated: if you live in a city with a subway instead of a suburb in a cul-de-sac, your political opinion is likely to correlate with the transportation mode extant in that domestic modality.
Recent events in the automobile industry have rendered the automobile argument moot. Unless the top OECD economies all roll back EV targets, the Chinese have disrupted (read: pwnd) the entire industry. You're welcome to question this, and please bear in mind that I come from the land of the muscle car, so I have an emotional tie to the internal combustion engine only rivaled by most of my German friends. I worked for the Volkswagen Akademie for about fifteen years. However, emotion is not the way to lean here.
Back to LLM buildout then. I believe that if the dominant architecture humans are developing now stays around too long, we'll reach a point of no return on compute invest. There's an economic inevitability at play here, one that I'd like to compare to the investment in roads, highways and automobile manufacturing. The data centers that are being built right now have a lifespan. The components have wear margins. They don't last forever.
McKinsey put these numbers on it in April of last year: by 2030, data centers are "projected to require $6.7 trillion worldwide to keep pace with the demand for compute power," with $5.2 trillion of that for data centers equipped to handle AI processing loads.
So, my concern is that we are already at a point where deploying new, possibly better technology is not feasible due to the bets that have been placed on the current infrastructure. If and when a neolab serves us a new, tasty technology that is more efficient, the McDonalds and Burger King guys are likely to swoop down and apply their nation-equivalent wealth to squashing any attempt to improve the state of AI yumminess.
OK, I went off on a different metaphorical arc there, I'm just making sure you're still paying attention. Back to internal combustion automobiles, which run on petroleum.
I'm naïve enough to believe that if 5% of weight and power of the petrochemical industry had been diverted to fusion research, we would be in a very different energy state right now. I'm not as smart as Vaclav Smil, so you should perhaps forgive me for this uncharacteristically optimistic view. Let me just say that if it works in the universe (I believe that there is a surfeit of evidence that it does) humans can technologically apply it. The question is how long will it take?
And as long as I'm playing woulda coulda shoulda here, if the romantic idea of cruising around in your own private transportation module instead of crowding into a larger metal container with the rest of the herd were less attractive, we would also be in a different world.
Buildout apologists will no doubt be thinking that compute is neutral, that if newer models based on different architectures are released, they will run just as comfortably on the hardware we're building now. True enough.
However, it is increasingly evident to me as a user and observer of the current tech stack that the buildout is a tool to gain control. $7 trillion was never required. But the bet is large enough that governments will back it, even in the face of NIMBY pressure.
At the same time, the increase in model efficiency (it has a cap, but nobody knows where that cap is) is rapidly heading to a state where most of the AI tasks can be performed on a laptop locally.
So, along comes a cute little baby dragon. And a Frank Frazetta character comes along and slays it with a labrys, likely through the economic expedient of buying the dragon's parent company. Or the political expedient of convincing the political structure that it is too dangerous to even consider. Bummer.
There are advantages to being optimistic. I'm pessimistic by nature, so I think the most likely scenario is less magical than my preferred outcome. But I'm a sucker for the underdog narrative, so we can join hands, sing Kumbaya and hope for the best.
1 The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain
