Premise
We've got insanely powerful tools.
LLMs. Agents. Autocomplete on steroids. Research on demand. Code that writes code. Plans that write plans.
The speed of progress went up. Waaay up.
Looks pure on the first sight.
It is not.
Bandwidth
Your brain has a limit.
Not metaphorically. Literally.
Attention is capped. Working memory is tiny. Cognitive energy drains. Context switching hurts. Multithreading thoughts is expensive (and mostly fake, let's be honest about it).
Yet modern tooling encourages the opposite:
- more tabs
- more threads
- more options
- more analysis
- more optimization
Result: mental swap file thrashing.
Then vs now
Before:
You solved similar problems repeatedly.
Patterns formed.
Decisions got cached.
Now:
We solve new problems faster.
But we re-evaluate old ones again.
Because we can.
We can ask:
- "Is this still best?"
- "What do others use?"
- "Is there a better abstraction?"
- "Should I redesign this?"
It's like an infinite improvement loop.
The hidden tax
Tooling removed execution friction.
It did not remove decision cost.
Actually it increased it.
Because now every choice has:
- alternatives
- benchmarks
- opinions
- comparisons
- benchmarks of comparisons
Our brain pays for all of it.
Context poisoning
LLMs have context windows.
So do us.
The difference? LLM throws away tokens after request.
Our brain don't.
We accumulate them as:
- half-finished ideas
- tools to try
- architectures to revisit
- notes to review
- systems to redesign
We walk around with something like 200 open mental tabs.
The menu problem
You open a restaurant menu.
It has 12 pages.
Optimal strategy is not analysis.
Optimal strategy is:
pick something reasonable.
If bad - order different next time.
If good - remember it.
I do this in real life too.
For years I have been buying the same shoes I wear daily and the same t-shirts from a chain store. I know them. I know they fit. I know (at least how long) they last.
At this point I just order them online. No browsing. No trying. No decision fatigue. It saves time and, more importantly, mental bandwidth.
Humans historically survived by satisficing, not optimizing.
Modern builders forgot that.
Local optimum is fine
You don't need best tool.
You need good enough tool that:
- works
- is stable
- you understand
Switching cost is real.
Re-evaluation cost is real.
Learning cost is real.
Optimization is not free. It is a loan with interest.
Rule
If it works - keep it.
If it solves the problem - ship it.
If it is understandable - don't replace it.
Save brain tokens for problems that actually matter.
Practical heuristics
Things that reduce cognitive load:
- reuse decisions
- standardize stacks
- limit options
- write less notes
- close tabs more aggressively
- stop researching once solution is viable (enough)
- avoid tool tourism
Senior engineers don't know more tools.
They ignore more tools.
Real bottleneck
The limit is not models.
Not GPUs.
Not tooling.
It's attention.
The scarcest resource in modern development is not compute.
It is focus.
Final thoughts
You don't need bigger context window.
You need stricter context pruning.
Best,
-- ł.