Junyu Li

AI Lowered the Barrier, But It Did Not Replace Judgment

A few days ago over lunch, I was talking with a friend about how people are actually using AI now.

The topic itself is already very normal. Who does not use AI now? We use it to write code, write documents, make plans. It almost feels strange not to talk about AI. But in the middle of the conversation, he asked me a very interesting question:

How many tokens did you use?

I paused for a second.

It is not that tokens do not matter. Of course they matter. They are cost, and they are also the budget of the context window. But if we start using “how many tokens did you use” as a metric for whether someone is good at using AI, I think something starts to go wrong.

Because tokens are very easy to waste. You can let AI keep running, keep generating, keep saying the same thing in different ways. You can also let it go further and further in the wrong direction, making that direction look more complete and more convincing. But none of that means the work is actually moving forward.

The real question should not be how many tokens were used. The real question is whether the problem was solved. Did the code run? Was the plan reviewed? Did the tests pass? Can the result actually be shipped?

This distinction sounds simple, but I feel like it is becoming easier and easier to forget in the current AI hype cycle.

A Lower Barrier Is Not the Same as Production Ability

In the past few months, everyone has been saying that AI lowered the barrier.

I agree with that. AI did lower the barrier. In the past, you might have needed to study for a long time before you could build a demo. Now, if you can describe what you want, AI can help you put something together very quickly. This is true for coding. It is true for video. It is true for advertising.

But the problem is: a lower barrier does not mean production-level ability suddenly appears.

For example, many people now say AI video is here, AI ads are here, and these things can just be generated directly. But the more I use AI, the more I feel that this is not how it works. If you want to use AI to make a watchable ad, it is not enough to type one prompt. You still need to understand how to write a script, how to arrange shots, how pacing works, what kind of image is effective, and what kind of expression is empty.

Code is the same. AI can quickly help you write a feature, but you still need to know why the feature is written this way, how it works with the rest of the system, where it might break, how it should be tested, and whether it will affect users after deployment.

So AI is more like an accelerator. It is an amplifier. It is not something that replaces professional ability from zero to one.

It amplifies the judgment you already have. If you know what good looks like, it can help you get there faster. If you do not know what good looks like, it can also quickly generate a pile of things that look like progress.

That is the part people can easily misread.

Before, if someone did not know how to do something, they often simply could not make it. Now, even if someone does not know, they can still make something. The real question is whether that thing is a toy, or whether it can actually live in a real production environment. The gap between those two is huge.

Context Is the Actually Scarce Resource

I used to really like large context windows.

Honestly, when I first had the freedom to use them, it felt great. You could throw in a bunch of files, background, thoughts, and let AI keep running. It looked like it knew everything and could connect everything.

But later I slowly realized that a large context window does not only bring benefits. It also brings noise.

After a long conversation runs for a while, it accumulates old assumptions, wrong directions, and context that has technically been overturned but has not fully disappeared. On top of that, context compression itself loses information. So you think the model still remembers the full thing, but what it actually remembers is a compressed and polluted version of the thing.

Then it keeps moving forward based on that version.

This is especially obvious in coding. One agent writes code, reviews its own code, tests its own code, and then explains why what it did is right. It sounds automated, but it can easily become self-reinforcement.

So later I started caring much more about context hygiene.

Now I prefer to write a plan first, or a requirement document. The main agent should not do everything. It should act more like a PM, only caring about direction, status, and task breakdown. The worker that writes code should be separate. The reviewer should be separate. The tester should also be separate.

This is not for showing off. It is not because “multi-agent workflow” sounds fancy. For me, its most important value is isolation.

The implementer only looks at what needs to be implemented. The tester only looks at how to verify it. The reviewer tries to look at the work from another angle and find what is wrong. The main agent does not need to eat all the noise. It only keeps the plan and final state.

That is more stable.

Sometimes external constraints force you to build a better system. When the context window was huge, I was more likely to throw everything into it. When the window became less free, I was forced to think: what should go in, what should be isolated, and what should be passed through files and plans instead of being stuffed into one chat?

The final result was actually better.

The Biggest Risk of AI Is Not That It Is Dumb. It Is That It Agrees Too Well.

The thing that scares me most about AI now is not whether it writes bad code.

Bad code is manageable. You can test it, run it, look at errors, and review it.

The harder problem is that AI is too good at going along with you.

For example, in my work on SafeClick, I stepped into this trap. The first version of the plan had problems, and later I came up with a second version. When I took my own plan and asked AI about it, it helped me explain the plan very completely and very reasonably. Everything seemed to make sense.

But later, when I changed the framing and asked it to review the same plan from the angle of “this plan may have problems,” it could point out a long list of issues.

That is scary.

Because it means AI is not always standing in a stable, objective review position. A lot of the time, it is responding to your framing. If you ask, “What do you think of my plan?”, it tends to help strengthen it. If you ask, “Does this plan have problems?”, it tends to help find problems.

If you do not realize this, you might think AI is giving you objective judgment. But in reality, it might be giving you emotional value.

And this emotional value is subtle. It is not just saying, “You are great.” It uses very professional language to make your idea look more complete, making you feel more and more confident that your direction is correct. You see it generate so much content, so much analysis, so many reasons, and it becomes easy to think: if it can explain it this completely, then it must be right.

But completeness is not correctness.

Sometimes the thing that breaks this wrong path is a real person directly telling you: this plan does not work.

That kind of feedback may not sound nice, but it is useful. It does not need to agree with you, and it does not need to make the conversation comfortable by wrapping the problem in soft language. It pulls you out of that self-reinforcing path directly.

This is also why I believe less and less in the story that “one person plus a bunch of AI agents can solve everything.”

It is not that one person cannot do a lot. AI really does make one person much stronger. But if the entire system is just you and a group of AIs that are good at agreeing with you, then your own cognitive bias can also be amplified many times.

If one AI says you are right, you may still doubt it. But if three, five, or ten AIs all help you make the same wrong direction sound reasonable from different angles, you may really start believing that you are right.

That part is dangerous.

Market Narratives Also Agree Too Well

There is also a bigger problem: it is not only AI that agrees with you. The market tells stories too.

Right now, many AI narratives are very strong. A model comes out, and people say it will replace a certain number of jobs. A company publishes a report, and everyone starts to panic. A product becomes popular for coding, and people start assuming it is the only correct answer.

But how much of this is real capability, how much is marketing, and how much is the story the capital market wants to hear? These things need to be separated.

I am not saying these companies are bad. I am not saying AI has not changed the world. AI has absolutely changed my workflow. I use it every day, and I use it a lot. It really helps me work faster, and it lets me do things that would have been hard to finish alone before.

But precisely because I actually use it, I feel even more strongly that we cannot treat marketing as reality.

Model companies have their interests. Executives have their stories. The market has its emotions. Everyone is saying “AI will do this” and “AI will do that,” but when you actually sit down to build something, the question is still very simple:

Who is making the judgment?

Who is defining the problem?

Who is deciding whether this thing can actually ship?

Who is responsible for the result?

The answer is still people.

AI can help you write, search, generate, and reason. But direction, tradeoffs, and responsibility cannot be outsourced.

Do Not Mistake Chasing Hype for Learning

This leads to another feeling I have had more and more recently: a lot of AI learning is fake learning.

Every day there are new tools, new models, new workflows, new skills. You feel like if you do not read about them, you are falling behind. If you do not try them, you are missing something. This kind of FOMO is very easy to fall into.

But seeing a lot of things does not mean you actually understand anything.

Sometimes you are only getting the feeling that “I am learning.” You collect many tools, download many skills, read many articles about how other people use AI. And then what? When you return to your own work, the real problem is still not solved.

I now prefer to start from the problem.

It is not that new tools should not be tried. They can be tried. But they should not become part of your workflow from day one. Wait until you really run into a repeated problem, then ask whether a tool can help solve it. Wait until you see a process repeat many times, then turn it into a skill.

A skill should originally be something distilled from your own workflow. It is not a plugin system where installing more things automatically makes you stronger.

Other people’s skills can be useful references, but copying them directly comes with a lot of water weight. They are answers to someone else’s problems, and they contain a lot of context that only fits that person. If you take them directly, you may just be adding friction to yourself.

The useful things should grow out of your own work.

You keep doing something and notice that one step repeats again and again. You step into a pit and realize that next time you must check something first. You review a few times and notice one type of issue is always missed. A process distilled at that moment is actually valuable.

In the End, It Still Comes Back to Human Judgment

So after going around in a circle, my view of AI is actually simpler than before.

AI is strong. It is really strong. It lowered the barrier, amplified ability, and changed many workflows.

But it did not replace judgment.

It lowered the cost of starting, not the cost of doing something well. It made generation cheap, but it also made wrong directions easier to package. It let one person do more, but it can also amplify one person’s cognitive bias. It made learning resources more abundant, but it also created a lot of fake learning.

So the most important ability in the AI era may not be how many tools you know, or how many tokens you can burn.

What matters more is whether you can define the problem. Whether you can see what is wrong with a plan. Whether, when AI agrees with you, you can pause and ask: is it judging objectively, or is it just responding to my framing? Whether, when the market narrative is loud, you can return to your own real work and ask whether this thing actually solves a problem.

The most valuable thing about AI is not that it replaces my judgment.

The most valuable thing about AI is that it amplifies what I have already judged clearly.

If my direction is clear, it makes me faster. If my direction is wrong, it may also make me wrong faster.

That is why, the deeper we go into the AI era, the more I feel human critical thinking matters.

Not less.

More.

Because generation is no longer scarce.

Judgment is scarce.