I am finishing more than I ever have. That is what makes the problem so strange.

AI has made it so cheap to learn, test and rebuild that I can outgrow an idea while I am still implementing it. Sometimes I finish the work and discover that the person who began it no longer agrees with all of its assumptions.

The work is not old because the wider field has moved on. It is out of date against my own understanding, which has continued moving while I built it.

That is an extraordinary gift. It is also how AI made me so productive that the shelf life of my work collapsed.

I have always learnt by doing

One of the persistent threads in my life is a need to understand how things work. Unfortunately, it is paired with very little patience for learning only in the abstract. I want to try the thing.

That is why I fell in love with technology. I built my first website in Notepad before discovering Notepad++. I learnt SQL through a terminal. These were not the best tools available; being self-taught often meant finding the difficult route shortly before finding the sensible one.

As life became busier, I tried to compress the process. Books became audiobooks. Talks and courses ran at twice their normal speed while I exercised or took a break. Blog posts multiplied into twenty open tabs. There was always another useful idea to absorb and never quite enough time to absorb it.

AI changed that relationship. It did not merely give me more information. It made information interactive and experimentation inexpensive.

AI turned explanations into experiences

I now use AI to explore unfamiliar codebases, question my assumptions, break down difficult systems and quiz me on what I think I understand. I have built tools to manage the videos and articles I once kept open in tabs. More importantly, I can turn a question into a working experiment while the question is still fresh.

Books can let us borrow an author's experience. AI adds something different: the ability to create a small experience of our own. I can build two versions, change one assumption, benchmark the result, debate the interpretation and throw the whole thing away if it teaches me enough.

That distinction matters. Reading about a design choice is not the same as living with it. An explanation can tell me what a trade-off is; an experiment lets me feel where it becomes inconvenient.

I am careful not to mistake generated work for understanding. If I stop paying attention, AI can remove the very struggle through which the lesson would have formed. Used deliberately, though, it lets me reach that struggle much sooner.

Then my work began ageing in real time

A cycle that once took weeks can now happen in a day. Instead of carrying one or two experiments, I can have seven or eight moving across software, AI systems, investing and whatever else has caught my attention.

Each experiment creates an insight. I can turn that insight into an alternative, refine it, test it and gather evidence that it is better. This is not merely another idea competing for my attention. It can become a working proof while the original is still underway.

By the time I reach the end, I may be completing a decision I would no longer make—and already have the evidence to explain why.

AI shortened the distance between curiosity and experience. It also shortened the life of the decisions that experience produces.

My AI Aquarium experiments made the pattern impossible to ignore. Before I had finished one approach, I had learnt enough from it to see its weaknesses and begin a second, more controlled system in parallel. I am still completing the first. The second is not an arbitrary distraction; it is a reasoned response to what the unfinished experiment has already taught me.

That is what makes the loop difficult to manage. The newer design usually is better. Its architecture is clearer. Its boundaries are more deliberate. It is easier to maintain, test or move. By the measures I care about, the refinement is real.

The awkward measure is how long the resulting work remains the best expression of what I know. Increasingly, the answer is: not very long.

The constraint moved

The Goal, by Eliyahu M. Goldratt and Jeff Cox, gave me useful language for this. Improving one stage of a system does not necessarily improve the whole. It often moves the constraint somewhere else.

AI dramatically improved my ability to discover, build and test. Those were once the slow parts. Now the constraint is my ability to feed what I have learnt back into work without continuously destabilising it.

I had mistaken local acceleration for progress through the complete system. I was producing more experiments, more finished work and more informed possibilities, but the useful life of each outcome was becoming shorter.

There is a subtler mistake too. The work had not necessarily become obsolete for anyone else. My understanding had changed. I was treating every improvement in my mental model as an immediate obligation to redesign the artefact.

The limitation is me. So is the cause

It would be easy to call this a problem with AI. It is more accurate to say that AI found one of my existing tendencies and removed the friction that used to contain it.

I like learning. I like doing a good job. I find it difficult to leave a design alone when I can see a better one. Before AI, rebuilding imposed enough time and effort to force a choice. Sometimes that friction protected me from myself.

Now a redesign can begin before the emotional excitement of the discovery has faded. Several can begin in parallel while the original work continues towards completion. Each carries new terminology, decisions and unanswered questions. One lesson becomes several live branches, all asking to be compared with the work that produced them.

I have always struggled to sleep when I have too much to process. AI can now give me a week's worth of new decisions, contradictions and what-ifs in a day. The work may have stopped, but the comparison between those branches has not. I go to bed and keep redesigning them.

AI did not create my curiosity or my impatience. It scaled both.

I am still learning when new knowledge should take effect

I do not have a neat solution yet. That uncertainty is part of the note.

I am beginning to distinguish discovery from delivery. During discovery, rebuilding may be the work: the point is to expose assumptions and improve the model. During delivery, a better idea does not automatically need to enter the current version. It may belong in the next one, unless it changes the value, safety or basic viability of what I am building.

That distinction sounds obvious. It is much harder to maintain when the newer understanding arrives before the current work is complete, and acting on it is cheap, interesting and probably worthwhile.

The positive result remains enormous. I can learn and gain practical experience at a rate that would previously have been impossible for me. I would not willingly give that up.

But learning and delivery run on different clocks. A better possibility is not automatically a reason to invalidate useful work. My new constraint is no longer how quickly I can explore. It is deciding when new understanding should change what I am already building—and when it should wait for the next version.