Notes from the first page
How the Workbench Happened
This wasn't supposed to become a workshop.
There was no master plan, product roadmap, market analysis, or whiteboard covered in arrows pointing toward THE FUTURE.
It started with small problems.
Something took too long. Something was unnecessarily confusing. An idea was harder to explain than it should have been. A pile of information required far too much digging to find the one useful answer buried inside it.
And increasingly, the same question kept coming up:
Could this be better?
For a long time, answering that question with software meant the idea had to be important enough to justify actually building it. Development takes time. Time costs money. And “I wonder if this would be useful” is traditionally a somewhat flimsy business case.
AI changed that equation.
Suddenly, the distance between:
“I wonder if this would work…”
“Well, let's find out.”
became remarkably small.
So we started finding out.
One experiment explored whether AI could help someone think through a problem instead of simply handing over an answer.
Another asked whether learning might stick better if you had to make decisions instead of memorize them.
Another started looking at the same news story through several very different minds.
Bullshit Detector came from the considerably less academic question:
Can we please just say what we actually mean?
Some ideas worked.
Some didn't.
Some worked just well enough to reveal a much better idea hiding underneath them.
And a few went from “this is kind of ridiculous” to “wait... this is actually useful” surprisingly quickly.
Eventually, a pattern became hard to ignore.
The experiments we kept coming back to weren't really about AI.
They were about friction.
The friction between information and understanding.
Between knowing something and explaining it.
Between having choices and knowing what to do next.
Between what someone meant and what somebody else heard.
Between curiosity and the amount of effort traditionally required to follow it somewhere interesting.
So the experiments got a bench.
AI is the material, not the point.
Everything here uses AI.
AI also helped build a lot of it.
The code, research, design iterations, testing, writing, debugging, and plenty of the product thinking have been developed collaboratively with increasingly capable AI systems.
We're not trying to build the next foundation model.
The companies spending billions of dollars on that problem appear to have it adequately covered.
We're interested in something much smaller—and, to us, much more fun:
What happens when those models become building materials?
What can a curious person or a small team make when the distance between an idea and a working prototype suddenly collapses?
What becomes worth trying when you don't need a six-month development cycle just to discover that your clever idea was, in fact, terrible?
And what happens when a prototype is cheap enough to throw away...
…but useful enough that sometimes you don't?
That's where the Workbench gets interesting.
The bench is deliberately unfinished.
There isn't a fixed roadmap for what belongs here.
That's not an oversight.
An experiment earns its place by being useful, interesting, surprising, or by helping someone see something a little more clearly than they did five minutes ago.
And the things on the bench aren't monuments.
People change them.
Someone gets confused on mobile. We change it.
Someone expects an analysis to go deeper. We dig deeper.
Someone points out that two modes supposedly behave differently but don't actually feel different. We test them.
Sometimes we discover a prompt problem.
Sometimes we discover a design problem.
And occasionally we spend an unreasonable amount of time discovering that the raccoon simply wasn't invited into the request.
These are the hazards of modern software development.
But that's also the point.
The Workbench isn't meant to demonstrate that AI can generate things.
We already know that.
It's meant to explore what happens when curiosity, judgment, real feedback, and increasingly capable tools can work together fast enough that an idea doesn't have to remain an idea for very long.
The process is pretty simple:
- Notice friction.
- Get curious.
- Build something.
- Use it.
- Learn from it.
- Make it better.
Then see what deserves a place on the bench next.
There was no master plan.
Depending on how much you value master plans, this is either reassuring or deeply concerning.