On August 4, Davide Piffer — a researcher whose day job involves psychometrics and population genetics — published a Substack post arguing that AI’s recent dominance in mathematical benchmarks has a less glamorous explanation than “superior intelligence.” The models, he wrote, have access to a vastly larger symbolic working memory than the human brain. The post hit the front page of Hacker News, where 401 points and 355 comments later, the debate settled into a familiar groove: Is AI actually thinking, or is it just remembering?
The question is a category error. And it’s obscuring the actual story.
The 4-to-7 Chunk Constraint
The human brain’s working memory is famously limited. George Miller’s 1956 paper pegged it at “the magical number seven, plus or minus two.” Modern research puts it closer to four. That limitation isn’t a bug — it’s a feature. It forces abstraction, prioritization, and forgetting. A doctor doesn’t hold an entire patient history in active memory; she holds a compressed summary and knows where to look when she needs the details. A lawyer doesn’t keep every precedent in mind; she keeps the shape of the argument and retrieves citations on demand.
Expertise, in other words, is largely a set of trained workarounds for a cognitive bottleneck. The bottleneck was universal. So the workarounds became the profession.
Now consider what a 1-million-token context window means. As of mid-2026, the frontier default has settled at roughly 1 million tokens — enough to hold an entire medical textbook, an entire codebase, an entire case file, an entire regulatory docket in active working memory simultaneously. Meta’s Llama 4 Scout advertises 10 million. The bottleneck is gone. And with it, the workarounds that defined entire professions.
The Workaround Economy
The first wave of AI displacement — 2023 through 2025 — hit tasks where the human working memory limit wasn’t the binding constraint. Entry-level writing, basic coding, customer service. The models didn’t need to out-remember anyone; they just needed to be good enough at pattern completion.
The second wave will be different. It will hit professions where the entire skill is managing information that exceeds working memory. Medicine. Law. Finance. Engineering. The radiologist who spent a decade learning to hold a compressed mental model of a scan while cross-referencing a patient’s history — the model holds the entire history, the entire scan, and the entire corpus of radiology literature at once. The associate who spent three years learning to keep a deal’s structure in her head while juggling diligence documents — the model holds all of it, simultaneously, without fatigue.
This isn’t a story about whether AI is “really intelligent.” It’s a story about what expertise means when the constraint that defined it disappears.
The Uncomfortable Part
Here’s where the argument gets uncomfortable for everyone.
The AI boosters said the technology would augment professionals, not replace them. The working memory gap suggests otherwise. When the bottleneck disappears, the workaround loses its value. A tool that can hold the entire case file in active memory doesn’t need the associate to summarize it. A tool that can hold the entire patient history doesn’t need the nurse to triage it.
The professionals said their jobs required judgment, not just memory. That was true — but the judgment was built on top of the memory workaround. When the memory workaround is automated, the judgment becomes a thinner layer than anyone wants to admit.
And the policymakers who designed retraining programs for manufacturing workers — the ones who assumed the next wave of automation would hit blue-collar jobs again — are about to discover that the credentialed professional class is not immune. The cognitive bottleneck was never universal. It was just human.
One trader on a fixed-income desk put it plainly when I asked about the context window arms race: “We spent twenty years building systems to summarize information for humans who can only hold four things in their head. Now the machine holds all of it. What exactly is my job?”
What Gets Rebuilt
The honest answer is that we don’t know yet. But the shape of the answer is becoming visible.
The professions that survive will be the ones that rebuild themselves around the new bottleneck — whatever it turns out to be. The bottleneck won’t be memory. It might be judgment under uncertainty. It might be accountability. It might be the willingness to sign your name to a decision when the machine has already done the analysis.
The professions that don’t survive will be the ones that insist the old bottleneck was a feature, not a bug.
Piffer’s post was framed as a deflation of AI hype — a reminder that the models aren’t outthinking mathematicians, just out-remembering them. But the deflation cuts both ways. If the advantage is memory, not reasoning, then the advantage is real, it’s here now, and it’s not going away. The question isn’t whether AI is thinking. The question is what happens to the people whose expertise was built on the assumption that thinking required forgetting.