Day 8 | In the AI era, what matters most is judgment and taste

What matters most in the age of AI?

Here's my ranking: taste = judgment > asking good questions > willingness to actually build things.

Asking and building both matter. But AI now handles more and more of the work, from organizing material and researching problems to carrying out tasks. That leaves a different set of questions: which of these outputs are worth using? Which ones can't be trusted yet? What should I get right first?

I run a lot of projects at once and switch between very different fields. Working this way has made me care more and more about one thing: everything needs its own yardstick. Without one, I end up with finished work and no standard for telling whether it's any good.

Output is no longer the hard part

With AI, one person can stretch what they handle to something close to a team's workload. More things become possible, and so do more trade-offs.

AI will happily keep writing, keep listing options, keep filling in material. My time and attention are limited. Which questions will shape the next decision and need answering now? Which are polish that can wait? I still have to decide that order myself.

So taste, as I mean it, includes a bar for done: what result am I willing to actually use? Where does something look finished but still fall short?

The travel booklet from my last post is a concrete example.

I used AI to make a travel booklet for a friend. I assumed it would be easy. The text inside the images alone was hard to control: fix one word and something else breaks, and the image gets blurrier with every pass. Then the resolution was too low for print, and after that came color shifts.

If "there's an image" counted as done, I could have stopped long before. What I wanted was a booklet that prints properly and ends up in my friend's hands. The text, the sharpness and the printed colors all decide whether it's usable.

That project gave me more respect for print shops and designers. Only after doing it myself did I see how many judgment calls go into a finished piece. AI can generate the images now, but someone still has to notice what's wrong with them.

Several fields in one day

A typical day for me: one moment I'm discussing PCB placement, the next I'm troubleshooting a DNS issue on Cloudflare.

These fields have little in common, and one set of answers doesn't carry over to the other. Each switch means quickly pinning down the actual problem, understanding the constraints, and deciding where to look next.

Working across fields also means "sounds reasonable" isn't enough. When AI hands me a complete-sounding explanation, I still need to know: under what conditions does it hold? Does it match the situation I'm actually in? If I follow it, which of my problems does it solve?

PCB placement means going back to the requirements this board has to meet. DNS troubleshooting means going back to what I actually observe and what the lookups return. The vocabulary is different, but in both cases I first have to define the scope of the problem.

That's why I put taste and judgment together. I need to know what a good result looks like, and I need to know what evidence I have that's strong enough to move in that direction.

With many projects running, I can't put the same effort into everything. Questions that affect direction get settled first. Details that don't change the decision in front of me can wait. Telling those apart is how I decide where my attention goes.

How I judge what AI gives me

My method is first principles, combined with checking AI output against experimental results.

For me, first principles means going back to the problem itself: what am I actually trying to solve? Which conditions are known? Which are assumptions I'm working with for now? Once those are separated, I can see what the conclusions are built on.

Whether AI gives me a report, research or a summary, I ask it to leave the reasoning in. What the data supports, where speculation begins, why it recommends this approach: all of that has to be spelled out, or I can't check it.

A smoothly written explanation is easy to accept. I still take it back to the results. Do the actual results support it? If not, which assumption needs another look?

The travel booklet works the same way. How an image looks on screen and how it looks printed are two separate checks. Fixing the resolution didn't fix the color. Solving one problem doesn't mean the whole booklet is fine.

I want AI reports to let me do that kind of check. When it says a problem is solved, I want to know which part it confirmed, how it confirmed it, and what hasn't been verified yet.

Keeping the reasoning also makes the next round easier. If a conclusion turns out to be wrong, I can go back to the step in question, add data or rerun the experiment. With only a one-line answer, I have no idea where to start checking again.

One habit

The habit I'd pass on: ask AI to explain what its conclusion rests on, what it's assuming, and how it could be verified. Then check each point.

You can ask it directly:

What data or experimental results is this conclusion based on? Which parts are still assumptions? Walk me through the key steps of the reasoning, and give me a way to check it myself.

Once you have the explanation, go back to the data and the results. An extra paragraph of explanation from AI doesn't mean that explanation has been verified. For anything important, you should be able to find what supports it.

This also gives "asking good questions" a concrete direction. When you know what you're checking, you can ask about the missing conditions and the unconfirmed assumptions, instead of just asking AI for more detail.

From PCBs and DNS to a travel booklet, I care about the same thing: what makes this result good enough to use? The more AI can help with, the more I need to keep that judgment for myself and decide what's good enough and what still needs work.