The Long Bet · Ongoing Series

The Floor Keeps Rising

The scarce skill in data and computing has shifted seven times since the 1880s. The foundation underneath it has never moved.

In 1925, Ronald Fisher was working on experimental design for agricultural trials at Rothamsted Research Station. Small plots of land, crop yields, soil variations. The tools he built – analysis of variance, randomised controlled trials, the logic of statistical significance – were so specialised that almost nobody outside mathematics and biology had heard of them.

A hundred years later, every clinical trial, every A/B test, every machine learning model uses the same foundations. Not because the field moved on from Fisher. Because it couldn't.

That is the thing about old tools that survive. They survive because they correspond to something the newer eras keep needing.


The pattern nobody is teaching

Every decade since the 1880s, the scarce skill in data and computing has shifted. A new bottleneck appears. The previous one dissolves into the baseline – into what everyone is assumed to have.

It has happened seven times.

Statistical intuition was the bottleneck until software made it easy to run the calculations. Programming was the bottleneck until databases and cloud platforms abstracted the infrastructure. SQL was the bottleneck until ORMs and no-code tools handled the queries. Distributed systems were the bottleneck until AWS commoditised the infrastructure. Classical ML was the bottleneck until sklearn and AutoML handled the model-building. Deep learning was the bottleneck until PyTorch and Keras abstracted the architecture.

Each time, the people who mastered the scarce skill built the tools that replaced it. Each scarce skill digs its own grave.

This is the pattern. Not a coincidence, not a trend – a mechanism. Every time a skill gets tooled, the value moves to whatever the tool cannot yet do.


The number that changes the argument

Here is how long each skill stayed scarce before the next tool arrived.

Years each skill stayed scarce
Before the next tool arrived and democratised it
*Boundaries between eras are approximate – Hinton's foundational work on neural networks predates the Deep Learning era by decades. The exact dates matter less than the direction. The coding cycle is not yet complete – 4 years is the current duration, not the final one.

A distinction before going further. The coverage and prerequisite findings earlier in this series are measurements – checkable against a 100-university dataset and admissions pages. This table is not a measurement. It is a reading of history: the era boundaries are constructed, and reasonable people can draw them differently. And the conclusion built on it – mathematics as the layer that holds – is a bet, not a finding. I think it is the right bet. But it belongs in a different category from “30 of 31 programmes require Mathematics”, and it should be read that way.

Two companion pages go further than this table can. For the full detail behind each era – what was abundant, what was scarce, who was elite, and a longer look at whether the next layer up (domain judgment) can compress the same way – see The Scarcity Ladder — A Timeline. For a single chart showing how each era's floor builds cumulatively on the last, see The Rising Floor.

60 4
Years of scarcity, first era to now
The cycle is not just repeating – it is compressing. Every iteration, the gap between "scarce" and "assumed" gets shorter. A four-year degree is now longer than a full scarcity cycle.

The implication is uncomfortable. Choosing what to study based on what is currently scarce is a bet against a curve that has won every time.


What the old foundation actually is

Here is the part that looks like a coincidence but isn't.

Across every transition – across seven different bottlenecks – one capacity has never been commoditised. Not statistical intuition as a set of procedures. Not programming. Not SQL. Something underneath all of them: the ability to reason rigorously about the limits of a system – whether that system is a statistical sample, a cluster of servers, or a neural network. They knew how things break, and why.

Fisher had it. The distributed systems engineers who built Hadoop had it. The ML researchers who designed transformers had it. At every era, the people who were genuinely elite were not just the ones who could use the tool – they were the ones who understood the limits of the tool. Who knew when the output was wrong. Who could see what the model couldn't see.

That capacity kept being needed because each new era produced outputs that could fail in ways that looked like success. A regression can be statistically significant and meaningless. A distributed system can be highly available and consistently wrong. A language model can be fluent and confidently incorrect. The tool cannot evaluate itself. Something outside the tool has to.

Today – in most companies, unremarked
LLM analysis
A product analyst asks Claude to review A/B test results. The LLM reports variant B is statistically significant – p = 0.03 – and recommends shipping. Confident. Correct-looking.
What the LLM missed
The team ran 20 simultaneous tests that week. At p < 0.05, one false positive is expected by pure chance. The "winner" is probably noise. An analyst with statistical training applies a multiple comparisons correction and kills the recommendation. The analyst without it ships a feature that does nothing – or quietly causes harm.
LLMs can run the regression and produce a confident, correct-looking analysis. What they cannot do is know whether to trust it – because that requires knowing which questions to ask before the data even arrives.

The old foundation – the thing Fisher was building in 1925 – is not a historical curiosity. It is what every subsequent era rests on, and what gets more exposed as the implementation above it gets automated.


The three layers

That capacity – reasoning rigorously about the limits of a system – is the first of three layers that make someone genuinely irreplaceable as AI automates the implementation beneath them.

Layer 1 – built at degree stage
Mathematical foundation
The ability to evaluate whether the AI's method is sound. Can you tell when the model is wrong? Can you see what kind of error is plausible? This comes from working at the level beneath the tools – from understanding what .fit() actually does, not just that it works. Without this layer, you cannot validate the output.
Layer 2 – built through years of practice
Domain expertise
The ability to tell when a technically sound result doesn't make sense in context. A model can pass every statistical test and still produce a conclusion that anyone in the field would recognise as wrong. Domain expertise is what lets you make that call. A degree programme is where it begins.
Layer 3 – built through accountability
Accountability
The social and legal legitimacy to make the final decision. A doctor who operates bears the consequences. An engineer who signs off on a design is professionally liable. An AI is neither. This is not a skill – it is a relationship with consequences, built through years of being in the room where the decision lands.
All three are needed to act responsibly on AI output. None alone is sufficient. The mathematical foundation lets you evaluate the method. Domain expertise lets you evaluate the conclusion. Accountability gives you the right – and the obligation – to decide.

The first layer is built early, at the point of choosing a degree or an early career path. The other two follow from years of practice, and they compound on top of whatever foundation was laid first. Which is why the foundation matters so much – not because it is sufficient, but because without it, the other two have nothing to stand on.


The same force, two different moments

The compression curve hits differently depending on where you are in your career. But it is the same force, and it raises the same question for everyone: which of the three layers do you actually have?

If you are choosing what to study

Don't choose based on what is currently scarce. The subjects that build the capacity underneath the tools – rigorous reasoning about uncertainty, mathematical foundations, the ability to frame a problem correctly – are not the ones that feel most immediately relevant. They never have been. In 1990, the immediately relevant skill was databases. In 2010, it was Python. In 2024, it is prompting. None of these is the foundation. All of them rest on something older.

The degree you choose is where the first layer gets built – or doesn't. The previous article in this series examines which subjects build that foundation most durably – and why the answer has been stable for longer than most people expect.

If you are mid-career

A data engineer who has spent eight years building pipelines opens Claude one morning and watches a junior analyst do in twenty minutes what used to take them a day. That is not a hypothetical. It is happening now, in most companies, unremarked.

The question is not "will I be replaced?" It is: what proportion of your value has been in the implementation layer, and what proportion has been in the layers above it?

If it was primarily implementation – the SQL, the pipeline, the model, the code – that position is dissolving faster than any previous transition. The moat is draining. If it was in knowing which problem is worth solving, knowing when the output is wrong, knowing what the model cannot know – that layer did not compress. It has never compressed.

The person who feels most exposed – who has spent years developing deep domain expertise and can see the disruption coming – is often the one with the most durable position. The person who feels fine is often the one who should worry: purely technical roles are exactly what the compression is replacing, and the demand signal just hasn't turned negative yet.


The open question

One thing the pattern doesn't settle: will domain judgment itself compress?

Every previous scarce skill could be encoded because it had a verifiable correct answer. Judgment in domains where being wrong is irreversible – a doctor deciding whether to operate, an engineer deciding whether a design is safe enough, a researcher deciding which hypothesis is worth five years of their life – may be different. These involve accountability that AI does not carry, trust that is earned in a specific domain over years, and the particular kind of knowledge that comes from having been wrong before.

Or it may not be different. The counterargument is that people may simply defer to AI anyway – which might already be happening in radiology and legal research.

If it does compress, what replaces it as scarce is something we cannot name yet.
The floor keeps rising. The question is where you are standing when it does.
Also in this series
The Problem Statement The four-constraint optimisation problem and why every obvious answer fell apart.
How I Built It Why ARWU over QS and THE, why 100 universities, and what 1,200+ normalised pairs made possible.
What the World Actually Offers CS at 95 universities. AI at 12. France above Imperial. What the coverage matrix actually showed.
The Unfashionable Four Four A-level subjects. Stable across 31 programmes, 100 universities, 19 countries.