Yuxuan Cao

AI Founder and Physicist
PhD candidate at UT Austin

[Yuxuan Cao @ACongroo]

AI could make society much richer, while making it much harder to escape the social economic class you were born into
--That possibility gets far less attention than job loss.

For a long time, one of the most important forms of upward mobility was simple: you could turn your skill into economic leverage.
You did not need to own a factory.
You could become a doctor, engineer, lawyer, scientist, programmer, accountant. You could spend years becoming good at something scarce, and the economy would pay you for it.
Your skill was an asset.

AI changes that equation because it attacks scarcity.
If a company can buy the equivalent of thousands of competent programmers, analysts, designers, researchers and paralegals as compute, then being competent at one of those things may no longer give an individual the same bargaining power.
The important question is not whether every profession disappears.
It is what becomes more valuable when human capability becomes easier to reproduce.
The answer is probably ownership.
Compute. Energy. Models. Robots. Infrastructure. Distribution. Capital.
Assets.
People sometimes respond that humans can still become athletes, creators, actors, influencers, entrepreneurs.
Of course.
But look at the economics of those careers.
They are brutally winner-take-most.
A tiny number of people become extraordinarily wealthy. Most make very little.
That is not a replacement for a society where millions of reasonably talented people can study engineering, medicine or law and reliably build a better life.
You should not need to be LeBron James, MrBeast or Elon Musk to move up the economic ladder.

This is where the AI debate becomes uncomfortable:
AI does not have to make ordinary people poor to weaken them.
It only has to reduce their leverage.
You could imagine a future with cheap goods, incredible entertainment, personalized AI companions, perhaps even some form of basic income.
Life might be materially comfortable.
But comfort and power are not the same thing.
If the productive assets are overwhelmingly owned by a small number of corporations and individuals, while everyone else rents access to intelligence from them, then the economy may become more productive while class mobility gets worse.

That is a very different version of the AI future from the one we usually hear.
Not “robots kill everyone.”
Not even “AI takes all the jobs.”
Something quieter:
what you know matters less, and what you own matters more.
That is why I think “democratizing AI” cannot simply mean giving everyone access to the same chatbot.
Access is not ownership.
If intelligence itself becomes productive capital, ordinary people need some ability to own, control and accumulate it.

We may build the most powerful productivity technology in history, but we also discover that: it also removed one of the main ways ordinary people ever gained economic power.

This is what is called Cyberpunk
High tech. Low life.

[Yuxuan Cao @ACongroo]

I think we’ve been designing AI coding workflows backwards.

My cofounder and I recently spent way too long debating a familiar question:

Should an agent write code?
Review it?
Test it?
Merge it?
Or should a human approve every step?

Eventually we realized—we were asking the wrong question entirely.

The real question is:

What should humans still need to remember to do?

Because if the answer includes things like:

“Remember to classify the issue”
“Remember to request tests”
“Remember to verify architecture impact”
“Remember to re-check every batch”
“Remember to ask for review at the right time”

…then we haven’t built an AI-native workflow.

We’ve just wrapped a traditional engineering process in an AI wrapper.

And for startups, that’s a problem. The promise of agents is simple:
one person should be able to take an issue from idea → implementation without turning every step into coordination overhead.

So we simplified the model.

We found there are really only three types of work:

1) Bugs (objective truth)

Login works or it doesn’t. Data persists or it doesn’t.
Here, the agent should be fully autonomous: investigate, implement, test, and verify.

2) UX / product judgment

“Does this feel confusing?” “Is this flow intuitive?”
The agent can build and propose—but a human must decide what’s actually good.

3) High-risk or irreversible decisions

Auth systems, data models, core architecture.
The agent should pause, surface tradeoffs, and escalate.

And here’s the key insight:

The human should not have to remember which category something falls into. The agent should handle that.

We also learned something painful:

“The agent says it’s done” is not the same as “it’s done.”

We once gave an agent four issues. It confidently reported success.
Later, we found at least one wasn’t actually fixed.

So now we require proof per issue:

#201: PASS — evidence attached
#202: PASS — evidence attached
#203: FAILED — explanation provided
#204: PARTIAL — remaining work listed

It sounds obvious in hindsight—but it’s easy to forget how often we accept summaries instead of verification.

The broader principle is becoming clearer:

Automate mechanical work aggressively. Preserve human judgment ruthlessly.

Agents should:

classify work
investigate issues
write and test code
reproduce bugs
gather evidence
detect architectural impact
escalate decisions when needed

Humans should focus on:

whether the product is actually good
whether the problem is worth solving
whether tradeoffs are acceptable

Not “human-in-the-loop” for everything.

More like human-on-the-loop.

The agent runs the system.
The human steps in only when judgment truly matters.

That feels much closer to what AI-native software development should look like.