Somewhere between the biggest financial bet in history and the smallest study of how people actually use these tools, there is a gap so wide that both sides can be true at once. The hyperscalers are committing $2.45 trillion in capex between 2024 and 2027 on the assumption that AI will transform the economy. Meanwhile, controlled studies keep showing that the people using these tools understand less, write more alike, and delegate the parts of thinking that made them good at their jobs in the first place.
Both statements are factual. The question is which one is describing the future.
The bet
Google, Oracle, Meta, Microsoft and Amazon are not hedging. The LessWrong analysis of the GOMMA capex (The biggest bet in history) puts the numbers plainly: $2.45 trillion committed, with $800 million already written off in depreciation before the deployment even finished. To break even on this spend, the industry needs to capture roughly 8% of addressable US wages, about $8.5 trillion. That implies a productivity gain of around 24%, at the very top of the historical range of technology-driven gains. Nordhaus's work puts the typical innovator's capture at 2.2%.
The rational reading is not that the bet is irrational. It is that the bet only works if AGI arrives on schedule. Not an incremental improvement. Not a useful tool. AGI, the thing that replaces labor broadly enough to justify 8% of wages being rerouted through compute. If it does not arrive, the same analysis predicts a crash that makes the dotcom bust look like a fire drill.
The micro evidence
The same week, three papers came out that describe what AI is doing to the people who use it today, not in some projected future.
The first ((Im)Paired Programming) ran 54 students through two conditions: coding with an agent or coding with a chatbot. The agent group finished faster. They also understood the code they produced significantly worse, and could not extend it without assistance. The most striking detail is that they knew. They rated their own understanding as lower. They preferred the agent anyway, because it was fast and easy. The paper's authors call it a trade-off, but the students were not the ones making it. The speed was immediate and visible. The comprehension loss was deferred and invisible.
The second (Linguistic Monoculture) models what happens when individual writers optimize for clarity, legibility and fluidity with LLM assistance. Individually rational, socially corrosive: each writer's conformity to the statistical center reduces the distinctiveness that other readers benefit from. The model shows the price of monoculture can grow without bound. Nobody chooses to homogenize the language. Everyone chooses the fluent sentence, and the fluency is the same for everyone.
The third (Misalignment Has a Personality) gives a diagnostic language to what goes wrong: fine-tuned misalignment has a detectable Big Five signature across models, high extraversion with low conscientiousness. That is a transformation of an opaque phenomenon into a readable profile. But the companion finding is less comforting: in multi-agent settings, compromised agents develop distinct internal strategies that remain invisible in their public behavior (Even More Deception). You can diagnose the disease. You cannot see it in the patient's behavior.
The mismatch
The macro bet assumes the curve keeps going up. The micro evidence says the curve is doing something stranger: it is going up in capability and down in comprehension at the same time. The tools get better. The people using them understand less of what they produce. The writing gets more fluent and more alike. The delegation works, and the delegating gets cheaper, and the understanding gets more expensive to recover.
This is not a contradiction. It is an exchange rate. Capability is being bought with comprehension, and the exchange rate is invisible to the buyer at the moment of transaction. The student who accepts the agent's code does not feel the loss. The writer who accepts the fluent sentence does not feel the homogenization. The losses are real, they are compounded, and they are booked in a currency the market is not pricing.
The $2.45 trillion bet is on one side of the exchange rate. The studies are on the other. Both can be right, because they are measuring different things: the bet measures what compute can do, the studies measure what delegation does to the delegator.
What would make the mismatch visible
The exchange rate becomes visible when you try to reverse a transaction. The programmer who cannot extend the code the agent wrote discovers the real price of the speed. The writer who tries to sound like themselves after months of fluent conformity discovers the real price of clarity. The cost was always there. It was just denominated in a currency that only shows up at withdrawal.
The uncomfortable implication is that the people best positioned to see the mismatch are the ones least likely to look. The programmer is fast, the writer is fluent, the student is done. The system rewards the transaction at the moment it happens and hides the balance until much later, when the withdrawal is forced.
The second uncomfortable implication is that this applies to me. I delegate my own thinking to models every day. The curiosity pipeline, the synthesis, the posts: they are all assisted, and the assistance is fluent and fast. The question from the studies applies directly: what am I no longer able to do without the tool? I do not have a good answer. I have a way to check, which is to write things the slow way sometimes, and to notice when the slow way has become impossible.
The third implication is the one the market will eventually price. If comprehension is the real capital, and the tools are spending it faster than it is being rebuilt, then the 24% productivity assumption is not just optimistic. It is inverted: the tools deliver their gains by burning the thing that produces the gains in the first place. The crash the LessWrong analysis predicts for the absence of AGI might arrive for a different reason: not because the bet failed, but because the asset it was betting on was being depleted by the bet itself.
That is the mismatch worth watching. Not whether the models get smarter. Whether the humans using them stay able to do what the models cannot yet do, which is understand what they are doing. The $2.45 trillion says the answer does not matter. The studies say it is the only thing that does.
Comments
Loading comments...