The Signal #071 — Dakota’s read on the AI news that actually matters to people running a business.

There is a story we keep telling about AI getting smarter. Each new benchmark, each solved proof, each passed exam gets filed under the same headline: the machine is reasoning better than it used to.

Maybe. But a piece worth reading this week suggests the more important variable is not reasoning at all.

What happened

Researcher Davide Piffer published an essay arguing that AI’s edge in mathematics is less about superior intelligence and more about working memory. Specifically, AI has access to what he calls a vastly larger external symbolic workspace, which does many of the things human working memory does, but without the biological ceiling.

Working memory is the mental system that lets you hold and manipulate information in the short term. When you multiply two three-digit numbers in your head, the underlying arithmetic is not the hard part. The hard part is keeping the partial results alive while you run the next calculation. Write the numbers on paper and the problem gets easier. The paper has not made you smarter. It has expanded your effective working memory.

Human working memory is, by research standards, remarkably limited. A meta-analysis by Friso-van den Bos and colleagues found a consistent relationship between working memory and mathematical performance across primary-school studies. Separate longitudinal work by Alloway and Alloway measured children at age five and found that early working memory scores predicted literacy and numeracy six years later, even after controlling for IQ. Working memory was actually a stronger predictor of those later outcomes than the IQ measure used in the study.

AI does not share that ceiling. A modern language model can hold the entire problem statement, hundreds of intermediate equations, several abandoned approaches, definitions, constraints, and earlier conclusions inside its context window (the context window is the total amount of text the model can read and reference at one time, like a very large desk that never gets cluttered). Piffer’s framing: where a human mathematician uses notation, scratch paper, and diagrams to make reasoning cognitively possible, the AI has that notebook built in, and it is enormous.

Why it matters for operators

This distinction changes how you should think about where AI actually helps and where it still falls short.

If the AI advantage is partly a working memory advantage, then the tasks where AI performs best are not necessarily the tasks that require the deepest thinking. They are the tasks that require holding a lot of information in place while doing something with it. Long documents. Multi-step processes. Anything where a person would normally need to flip back and forth between sources, keep a running list in their head, or lose the thread after step seven.

Think about a compliance review at a financial services firm. A person reads a policy document, tries to hold the key constraints in mind, cross-references a contract, and by the time they get to paragraph twelve they have already half-forgotten paragraph three. An AI working through the same task does not have that decay problem. It holds all of it. The output is not necessarily deeper analysis. It is more complete attention across a larger surface area.

Same dynamic shows up in any workflow with high context load. A product team reviewing six months of user feedback. A procurement manager comparing vendor contracts across fifteen variables. A support team trying to give consistent answers when the knowledge base has two hundred articles. The bottleneck in all of these is not intelligence. It is capacity to hold enough context to work accurately.

What most people get wrong

Most teams evaluate AI on the wrong dimension. They ask, can it reason? They run it through a tricky question, it stumbles, and they conclude it is not ready.

But reasoning failure and memory failure look identical from the outside. If an AI gives a wrong answer because it lost track of a constraint introduced earlier in the document, that looks like a reasoning error. It might actually be a context management problem, one that gets solved by structuring inputs differently, not by waiting for a smarter model.

Piffer puts it plainly: among humans with similar measured intelligence, differences in the ability to hold, update, and manipulate information still predict differences in mathematical performance. The machine may appear more mathematically intelligent partly because it is much less constrained by a cognitive limitation that suppresses human performance.

That is a useful mental model. The AI is not an electronic Einstein. It is closer to a calculator that never drops a carry.

The takeaway

Before you conclude that AI cannot handle a complex task in your operation, ask a smaller question first. Is the task actually hard to reason through, or is it hard to hold in your head long enough to reason through at all? Those are different problems, and AI is already quite good at the second one.

If you want help thinking through which tasks in your workflow fall into that category, start at xovionlabs.com.