Subtitle: A note on puzzle scores, Roman aqueducts, and the question nobody is asking about AI training data
There is a game. You may have encountered
it, or one of its cousins – tiled puzzles, colour-matching challenges, logic
sequences that resolve in a satisfying snap of completion. They are, in their
own terms, elegant. They reward pattern recognition. They are calibrated to
take between two and eight minutes.
What they are calibrated to produce, at
the end of those minutes, is a score. And the score is calibrated to be
shareable.
This is not an accident of design. It is
the design.
I want to be careful here, because the easy
version of this argument is boring. “People are distracted by their phones.”
This has been true since 2007 and it stopped being interesting as an
observation around 2012. That is not what I am describing.
What I am describing is something more
specific, more contemporary, and more structurally interesting: the collapse of
the membrane between performance and work, at precisely the moment when a
technology has arrived to do the work itself.
For most of professional history, there
were two fairly distinct populations. The people who worked, and the people who
gamed. The overlap was small. They occupied different hardware, different
social spaces, different self-concepts. The word “gamer” carried connotations
that had nothing to do with corporate ladders.
That membrane has dissolved. The most
earnestly professional social space we have – the network where people list
their credentials, their promotions, their certificates of achievement – is now
where game scores are posted. And the scores are posted not as admissions of
leisure but as demonstrations of intellect. The game has been reframed as a
cognitive benchmark.
The score is not entertainment. The score
is evidence.
This reframing is interesting because it
reveals something about what the professional class has lost. The old sources
of daily cognitive proof – the formula that worked, the argument that landed,
the strategy that proved itself – were private. They happened in documents and
meetings and decisions. They didn’t need to be posted because they had
real-world consequences that were observable to the people who mattered.
When AI absorbs that work, the
consequences remain. The business outcome happens. But the cognitive proof – the
moment of having done something difficult – has been redistributed. The AI did
the difficult thing. The human reviewed it, approved it, and forwarded it.
Review and approval are not the same as
doing. The brain knows this. And so it goes looking elsewhere for the feeling.
The puzzle provides it.
In 476 AD, the Western Roman Empire
formally ended when Romulus Augustulus was deposed. This is the date historians
give, but as historians will tell you if you ask them, dates like this are a convenience.
The actual process was longer, more diffuse, more ordinary.
What is perhaps more interesting is that
the aqueducts – the great engineering achievement, the infrastructure that made
Roman urban life possible – stopped functioning around 537 AD, more than 60
years later. Not because anyone decided to turn them off. Because the people
who knew how to maintain them gradually died, and the people coming up behind
them had not been taught, and the people who might have taught them had been
doing other things.
The knowledge didn’t disappear
dramatically. It was practised less. Then less. Then not enough. Then it was
gone.
Here is the question I find myself
returning to.
The AI systems currently capable of
writing arguments, drafting strategies, summarising complexity, and generating
analysis – these systems were trained on data produced by humans who did those
things. Decades of documents, papers, reports, decisions, arguments. The
accumulated cognitive output of a professional class that was, at the time,
doing cognitive work.
That is the inheritance. That is what the
models learned from.
The professional class that produced it
is retiring. The next generation is working alongside AI that handles the hard
parts. The cognitive output being produced now is, in some meaningful sense, a
collaboration between humans and systems – and the human’s contribution,
increasingly, is oversight rather than origination.
When these current models age – when they
are replaced, retrained, improved – what does the next generation of training
data look like?
I don’t know. I’m not sure anyone has
asked seriously. It is the kind of question that sounds paranoid until it doesn’t,
and by the time it doesn’t, the aqueducts have already stopped working.
None of this is an argument against AI.
The tools are extraordinary. The argument is simply this: using a tool and
maintaining the capacity to function without it are different things, and right
now, we are not clearly distinguishing between them.
Use the AI. It is not the same as
thinking.
Play the game. It is not the same as
sharpening your mind.
And do the difficult thing occasionally –
specifically the one the AI offered to do for you and you declined.
That muscle will not maintain itself.
Filed as a Quarterly Civilisational Progress Report. All
observations are historically accurate. All implications are yours to make.
LinkedIn Newsletter Article
No comments:
Post a Comment