Using AI /
How AI lowers the cost of testing an idea
Greg Brockman wrote recently that AI lets people ‘build things you would not have attempted before’.
Of all the claims in his note about a compute-powered economy, that is the one I recognise most clearly.
AI has made me faster. More importantly, it has changed which ideas survive the first ten minutes.
Every idea arrives with an invoice
An idea for a small piece of software used to carry a hidden cost.
You needed to learn the framework, understand the API, set up the infrastructure and work through all the mistakes that come with entering an unfamiliar field. Sometimes you needed a designer, developer or analyst before you could learn whether the original idea was useful.
The cost was not only time or money. It was the number of unknowns you had to accept before seeing anything work.
Most ideas were never disproved. They were abandoned before the first honest test because discovering whether they were good was too expensive.
The first test got cheaper
I have built a vessel tracker that parses cargo documents and plots ships on a live map. A tool that lets someone question an insurance policy. A health chatbot operating in five South African languages.
None of these projects proves that AI can do everything. They were also possible before modern AI tools.
The difference is that each crossed enough unfamiliar territory that I may not have started it alone. Mapping libraries, document parsing, insurance language, translation and deployment all created separate reasons to stop.
AI did not remove those fields. It helped me move through them quickly enough to reach something real.
The question changed from ‘Can I justify spending months learning this?’ to ‘Can I learn something useful by tonight?’
Cheap experiments replace arguments
When testing an idea is expensive, people debate it.
They write plans, make forecasts and argue about what users might do. Everyone is working from imagination because reality costs too much to consult.
When the first version becomes cheap, you can ask reality earlier.
Parse the document. Call the API. Put the rough workflow in front of one person. Find the point where it breaks.
A failed prototype is often more useful than another week spent polishing the idea. It gives you evidence. The feature nobody needs can die before becoming a roadmap. The awkward workflow becomes visible before a team is hired around it.
Cheaper execution does not make every idea good. It makes it less necessary to guess.
Compute is not the only constraint
Brockman argues that the scale and sophistication of the problems we solve will increasingly depend on the compute available to us.
Compute matters, but I do not think it becomes the only meaningful limit.
Context still matters. So do permissions, access to good data, trust, distribution, capital, taste and responsibility for the result.
AI can help produce ten possible systems. It cannot take responsibility for deciding which problem deserves to exist, which failure would hurt someone or when the rough prototype is safe enough to become a real product.
As producing gets cheaper, these other constraints become easier to see.
The bottleneck moves from making something to choosing well.
Starting is cheaper. Finishing well is not
There is an obvious failure mode here.
The same tools that make a useful experiment cheap also make it cheap to produce piles of plausible rubbish. A generated application can look finished while leaking data, failing silently or relying on assumptions nobody checked.
A prototype answers a question. A production system needs security, monitoring, maintenance, clear ownership and a plan for when it fails.
The dangerous mistake is to confuse the disappearance of first-step friction with the disappearance of work.
AI lets more ideas reach the starting line. It does not carry all of them to the finish.
The identities get less rigid
We used to treat job titles as permission.
‘I am not a developer.’
‘I do not know design.’
‘I have never worked with that kind of data.’
Those statements still describe real gaps. The model does not make someone an expert by association. But the gaps no longer need to prevent the first experiment.
I do not know whether AI will allow everyone to become whoever they want. Access, time, money and luck still matter.
But it does make it harder to dismiss an idea because one part sits outside your current skill set.
That is freeing in a smaller and more practical way. You can follow your curiosity further before deciding that the work belongs to somebody else.
The rule
When the cost of being wrong is low, stop asking whether you could build the whole thing.
Build the smallest version that could prove you wrong.
Use AI to lower the cost of finding out. Then let the result, rather than the excitement, decide whether the idea deserves more work.