How to Build Taste and Opportunity Sense in the Agent Era

TL;DR: In the agent era, the important skill is not memorizing prompts. It is building a small factory that keeps producing value, then developing the judgment to decide which outputs deserve to survive.

---

On June 1, Naval published a long conversation called The AI Industrial Revolution.

On the surface, it is about AI, agents, vibe coding, hardware, regulation, healthcare, and art. Underneath, the real message is simpler: stop defining yourself as the person who does the work. Become the person who builds the production line.

The subtitle, Build your own factory, is the key.

It does not mean everyone should open a literal factory. It means one-off output is becoming cheaper. The valuable thing is turning a class of work into a reusable, expandable, verifiable production system.

§From Shipping Output to Building Factories

In the old model, a person was valuable because they could ship an output: code, a plan, a video, a report.

Agents make many of these outputs cheaper.

The difference moves upstream. Can you define the goal, break down the process, choose the tools, set checkpoints, add human verification, and make the system keep producing useful results?

So the ordinary question should not be:

"How can AI help me write one piece of copy?"

A better question is:

"Can I turn this repeated kind of work into a small factory?"

For example:

• ·A sales lead research factory
• ·A short video topic factory
• ·A product feedback analysis factory
• ·A contract pre-review factory
• ·A customer follow-up factory
• ·An industry report factory
• ·A publishing and formatting factory

Using AI once saves time. Turning AI into a process creates an asset.

§Do Not Save Tokens, Save Your Time

Naval makes a counterintuitive point in the conversation: do not obsess over token cost. Look at your time and the final output.

The deeper view is simple. The industrial revolution was not mainly about saving coal. It was about saving scarce human time.

The agent era works the same way.

Many people hesitate over a few dollars of model usage, then waste hours manually searching, editing, copying, and fixing. The accounting is backwards.

A better pattern is: in exploration, run models in parallel, ask them to criticize one another, and generate alternatives. In delivery, verify aggressively: facts, edge cases, tests, user feedback, and responsibility.

Spend tokens during exploration.

Raise verification before release.

Do not mix these two stages.

§AI Amplifies Your Judgment

Another important point in the conversation is that AI often reflects the user's level in a domain.

An expert gets expert leverage. A beginner still gets help, but also exposes beginner-level judgment.

This does not mean beginners should avoid AI. The opposite is true. The entry barrier has dropped. In the past, you could get stuck in setup, syntax, errors, and search. Now agents can carry you across many of those obstacles.

But the ceiling still comes from you.

Can you see where the result is wrong? Where it is generic? Where it is risky? Where it looks complete but misses the real problem?

The agent era does not reward prompt templates as much as four deeper skills:

First, problem definition. Can you turn a vague wish into a clear task?

Second, evaluation. Can you judge whether the result is good?

Third, feedback. Can you tell the agent what to change next?

Fourth, boundary awareness. Do you know when the model will bluff, become lazy, default to templates, or sound too confident?

That is the real meaning of taste and judgment.

§Where Ordinary People Should Look for Opportunities

The most useful lesson for ordinary people is not "start an AI-native software company."

The more realistic opportunity is in old industries.

Blake Scholl talks about hardware engineering workflows that still rely on Excel, email, manual spreadsheet handoffs, no source control, and no automated tests. Once AI enters the process, small teams can make old engineering work more software-like, automated, and iterable.

This matters because many industries are not poor. Their processes are just old:

• ·Factory quotes still live in Excel
• ·Renovation budgets move through chat and spreadsheets
• ·Clinics still do follow-up manually
• ·Logistics exceptions still depend on phone calls
• ·Legal drafts still come from copy and paste
• ·Training centers still schedule by hand
• ·Cross-border sellers still rely on scattered tools and gut feel
• ·Small manufacturers still depend on senior workers for inspection

The opportunity is not necessarily competing with model labs.

The better question is: which industry still has too much Excel, email, PDF, screenshots, group chats, and manual copy-paste?

Those are factories that have not been softwared yet.

Turn them into agent workflows, and you may have an opportunity.

§Wealth Is Not "Making Money With AI"

We should be sober here.

This is not investment advice. It is not another story about getting rich from an AI side hustle.

The more reliable wealth logic is: use AI to lower production cost, then turn the freed-up capacity into ownership, cash flow, customer relationships, or equity.

An ordinary person can follow this path:

First, find a process that is frequent, expensive, repetitive, and under-softwared.

Start inside an industry you understand. AI amplifies domain judgment. It does not magically give you one.

Second, build the smallest factory that actually runs.

Do not start with a big platform. Start with one workflow: what is the input, what does the agent do, where does a human verify, and what is delivered at the end?

Third, sell the result, not the AI.

Customers do not care which model you used. They care how much time you saved, how many errors you reduced, how much money you made, how many hires you avoided, or how many days you removed from delivery.

Fourth, productize the service.

At first, you may need human fallback. That is fine. Run an AI plus human verifier service. Once you have cases, data, error samples, and industry rules, gradually turn the service into software.

Fifth, build a real moat.

The moat is not "I know how to prompt." It is private data, industry workflows, customer trust, distribution, verification systems, brand, taste, and the ability to take responsibility.

§Humans Become Verifiers

One theme keeps returning in the conversation: AI generates, humans verify, sign off, and take responsibility.

That is not low-value work.

It may become more valuable because the verifier carries trust and consequences.

AI drafts a contract. A lawyer verifies the key clauses.

AI writes code. An engineer verifies security and maintainability.

AI summarizes medical research. A doctor verifies applicability.

AI produces compliance documents. A professional verifies risk.

AI proposes a brand direction. A founder verifies whether it still carries the soul of the brand.

One future moat is not "I generate better than AI." It is "I can be responsible for what AI generates."

§Why the Conversation Ends With Art

The later part of the conversation asks a beautiful question: What’s Your Definition of Art?

This is not a random detour.

The real question is: if AI can generate images, music, movies, software, essays, and product prototypes, what is still uniquely human?

Max's answer is close to meaningful out-of-distribution behavior. Art is not random weirdness. It is something that escapes an existing distribution and changes your future trajectory.

Naval's answer emphasizes emotion and intent. Someone felt something, created an object, and wanted another person to feel it too.

Guillermo's observation is more product-oriented. After you have seen enough Claude-generated websites, you notice a default taste: a certain font, color range, spacing, and mood. Once a style is mass-produced by a model, it stops feeling creative and becomes a template.

Put the three together, and art in the agent era has three conditions:

First, surprise. It cannot just be the average of training data and popular templates.

Second, meaning. It has to change someone's feeling, judgment, or action.

Third, intent. It has to feel like a real subject is expressing something.

My working definition is: in the agent era, art is an intentional form made by humans with machines that cuts through an existing distribution and changes another person's emotional trajectory.

It does not have to be a painting, film, or song.

It can also be a product, a company, a brand move, a workflow, or a way of living.

§Taste Becomes a Core Advantage

When generation becomes infinitely cheap, selection becomes infinitely important.

You can ask AI for 100 headlines, 100 pages, 100 images, and 100 product ideas. The question is: can you tell which one has life, which one is merely correct, and which one carries the model's default flavor?

That is taste.

Kant's account of aesthetic judgment contains a useful tension: the judgment comes from subjective feeling, but it also carries a claim that others should be able to see what is good in it. Taste is not simply "I like it." It sits between private preference and public judgment.

Hume's view is more useful for training. Good taste comes from delicate perception, practice, comparison, freedom from prejudice, and long calibration.

In today's world, taste is not decoration. It is business judgment:

Product taste: knowing what to cut and what to keep.

Brand taste: knowing what sounds credible and what sounds greasy.

Design taste: knowing what has order and what merely looks premium.

Content taste: knowing what has force and what is just AI-generated correct prose.

Founder taste: knowing which problems are worth solving and which ones are just technical toys.

§How to Train Taste

Training taste is not about consuming more.

The better method is comparison, naming, imitation, and deviation.

First, study classics and the present.

Only studying classics can create museum taste. Only studying the present makes you a servant of the feed. You need both durable standards and contemporary language.

Second, name differences.

Do not stop at "good" or "bad." Say whether it is proportion, rhythm, whitespace, tension, material, narrative, light, tone, structure, contrast, restraint, motion, or information hierarchy.

If you can name differences, you can improve judgment.

Third, compare side by side.

Each day, put three works of the same type next to each other: one classic, one commercial hit, and one AI-generated version. Ask: which has more intent, which is more generic, which is more memorable, and which is just a template?

Fourth, copy strong work.

Do not rush toward originality. First imitate a good essay, poster, web page, short video script, or film scene.

Imitation is not for publishing. It is for understanding the choices behind the work.

Fifth, deliberately deviate.

After copying, change one important variable: make the cold palette warm, slow down the pacing, cut the information density in half, turn commercial language into poetic language, or turn a Japanese composition into a Bauhaus structure.

You will start seeing how one variable changes the whole mood.

Sixth, build an anti-slop checklist.

Every model has a default flavor. Record the aesthetic errors it repeats: excessive symmetry, excessive smoothness, over-explaining, empty emotion, clickbait, fake premium feeling, template colors, and lack of real detail.

Before generating next time, ask the agent to check against the list.

Seventh, return to the physical world.

If all your experience comes from screens, it is hard to create anything beyond screen culture.

Go to exhibitions, travel, cook, exercise, talk to people, observe streets, visit factories, touch materials, and hear live music. Reality gives you the raw material to resist the model average.

Eighth, use AI as a coach, not a ruler.

Ask AI to explain composition, history, genre, and failure. Ask it to generate 20 variations. Ask it to play a harsh critic.

But the final decision has to be yours.

Taste is not asking AI which one is good. Taste is using AI's many variations to train your own judgment.

§Final Judgment

The lesson for ordinary people is not to chase every new model or memorize more prompts.

The more important bets are:

First, use the strongest models on real problems. Do not save tokens by wasting your life.

Second, choose an old workflow in a domain you understand and turn it into a small agent factory.

Third, understand basic systems: inputs, outputs, data, permissions, tests, logs, and feedback.

Fourth, become a trusted verifier in a specific domain.

Fifth, keep training taste, because when production becomes cheap, selection, deletion, judgment, intent, and style become expensive.

In the agent era, humans should not stay mere executors.

The better position is director, architect, verifier, and owner with taste.

Wealth will flow to people who have leverage, judgment, distribution, trust, and aesthetic standards, then turn those things into production systems.

§References

• ·The AI Industrial Revolution
• ·Aesthetic Judgment, Stanford Encyclopedia of Philosophy
• ·Hume’s Aesthetics, Stanford Encyclopedia of Philosophy