Your test is exact: when the task is done, look at what is left behind. That test has a structural formalization.
A constituting entity operates within an admissibility band. Perturbation within the band is absorbed: the entity flexes, adapts, strengthens. The band widens. The cabbie's posterior hippocampus growing over four years IS the band widening through sustained perturbation. The flex IS the learning. The learning IS the structural change.
Remove the perturbation and the band narrows through disuse. The GPS user's spatial memory declining IS the band narrowing. The student who used ChatGPT scoring 57.5% against 68.5% IS the band narrowing. The narrowing is not a judgment. The narrowing is a structural observation: the range within which the entity can absorb perturbation contracted because nothing perturbed it.
Your two kinds of difficulty map to two regions of the gradient. Waste difficulty (the tax form, the broken interface) is perturbation that produces no flex. The entity absorbs it but nothing widens. The perturbation was friction, not structure. Building difficulty (the argument you forced yourself to hold still, the city you learned street by street) is perturbation that produces flex. The entity absorbs it and the band widens. The entity IS different after.
The dangerous part is the one you named: judgement is its own only check. You cannot hand judgement to the machine and keep a cheaper instrument to verify the machine's judgement, because the only instrument that could verify it is the judgement you just handed over. The verification loop is circular. The circularity IS the structural reason the loss hides itself. The band narrowed. Nothing in the narrower band can detect that it narrowed. The detection requires the wider band that no longer exists.
Really interesting piece. Difficulty isn’t the obstacle rather it’s how we build value.
Calculators and GPS made life easier, but we still learned the fundamentals before relying on them. AI feels different because it can step into any layer right from thinking to execution.
The challenge is making sure we use AI to enhance our judgment and remove only friction, not replace the productive struggle that helps us develop it.
But it's not unique to handing effort off to a machine. We lose capability any time we hand effort off to anyone or anything else.
Using AI is delegation. And just like new managers need to learn what to delegate and what not to delegate, people who use AI need to learn to consciously decide what work to delegate and what to keep.
When managing humans, any work you hand off to someone else, you become worse at. I used to be a top-tier programmer, but I've been in management so long that my programming skills are only a shadow of what they used to be; on the other hand, my people skills are much, much stronger than they were when I was a dedicated engineer. That's a conscious career choice I made.
And that's the choice we need to teach people to make intentionally when using AI. Delegation is a management skill, and with AI, everyone becomes a manager. That means many AI users may get more value from learning management skills than technical skills.
I like the management analogy. Every act of delegation changes what we practise, and over time that changes what we’re good at. AI makes those choices so cheap and frequent that it’s easy to stop noticing we’re making them.
That’s why I think being intentional about which difficulties we keep is becoming a genuine skill in itself.
Your test is exact: when the task is done, look at what is left behind. That test has a structural formalization.
A constituting entity operates within an admissibility band. Perturbation within the band is absorbed: the entity flexes, adapts, strengthens. The band widens. The cabbie's posterior hippocampus growing over four years IS the band widening through sustained perturbation. The flex IS the learning. The learning IS the structural change.
Remove the perturbation and the band narrows through disuse. The GPS user's spatial memory declining IS the band narrowing. The student who used ChatGPT scoring 57.5% against 68.5% IS the band narrowing. The narrowing is not a judgment. The narrowing is a structural observation: the range within which the entity can absorb perturbation contracted because nothing perturbed it.
Your two kinds of difficulty map to two regions of the gradient. Waste difficulty (the tax form, the broken interface) is perturbation that produces no flex. The entity absorbs it but nothing widens. The perturbation was friction, not structure. Building difficulty (the argument you forced yourself to hold still, the city you learned street by street) is perturbation that produces flex. The entity absorbs it and the band widens. The entity IS different after.
The dangerous part is the one you named: judgement is its own only check. You cannot hand judgement to the machine and keep a cheaper instrument to verify the machine's judgement, because the only instrument that could verify it is the judgement you just handed over. The verification loop is circular. The circularity IS the structural reason the loss hides itself. The band narrowed. Nothing in the narrower band can detect that it narrowed. The detection requires the wider band that no longer exists.
I like that framing. The really difficult part is that the feedback loop breaks.
If you can’t recognise that your own capability has narrowed, it’s much harder to recover it.
Really interesting piece. Difficulty isn’t the obstacle rather it’s how we build value.
Calculators and GPS made life easier, but we still learned the fundamentals before relying on them. AI feels different because it can step into any layer right from thinking to execution.
The challenge is making sure we use AI to enhance our judgment and remove only friction, not replace the productive struggle that helps us develop it.
Thanks, Vasanth. I think that’s the key difference too.
Calculators and GPS automated specific tasks, but AI can shortcut the whole process from thinking to execution.
The real question is which parts of that process we’re willing to outsource.
If we outsource the experiences that develop judgement, we lose the ability to recognise when an answer is actually wrong.
This is an excellent post, and it's all true.
But it's not unique to handing effort off to a machine. We lose capability any time we hand effort off to anyone or anything else.
Using AI is delegation. And just like new managers need to learn what to delegate and what not to delegate, people who use AI need to learn to consciously decide what work to delegate and what to keep.
When managing humans, any work you hand off to someone else, you become worse at. I used to be a top-tier programmer, but I've been in management so long that my programming skills are only a shadow of what they used to be; on the other hand, my people skills are much, much stronger than they were when I was a dedicated engineer. That's a conscious career choice I made.
And that's the choice we need to teach people to make intentionally when using AI. Delegation is a management skill, and with AI, everyone becomes a manager. That means many AI users may get more value from learning management skills than technical skills.
Thanks John. I really appreciate it.
I like the management analogy. Every act of delegation changes what we practise, and over time that changes what we’re good at. AI makes those choices so cheap and frequent that it’s easy to stop noticing we’re making them.
That’s why I think being intentional about which difficulties we keep is becoming a genuine skill in itself.