AI Changed How We Build. It Did Not Change What Matters (opens in new tab)

(piljoong.dev)

3 pointspiljoong27d ago4 comments

4 comments

But it did. What used to be unworthy of building is now totally doable. I can roll out one off ad-hoc UIs for our customer to navigate their bespoke data sets without having too much worry about expending a lot of development time to throw up a visualization page that may get discarded when its usefulness expires. So it has expanded the realm of the worthwhile if not necessarily the realm of the possible. At least not yet.

piljoongOP27d ago

I agree. AI clearly expanded the set of things that are worth building, especially small or previously unjustifiable work. What I was trying to say is that once those things become real systems, the old constraints show up again. You still have to understand boundaries, failure modes, and how to operate what got produced.

abraxas27d ago

Yes, no denying this. Now in time this will change too. I'm slowly letting Claude do some target maintenance work in AWS in non-critical environments. Over time I will let it perform reversible operations in production but that time has not come yet. And it still too often takes silly approaches to problems that have more elegant solutions. These days it has a personally of an ambitious, restless junior who wants to always get stuff done but doesn't have the depth of thought to just handle it all by itself. But the times they are a changing.

piljoongOP27d ago

I've been using coding agents a lot recently.

Building got faster.

But when the system actually has to run, the same issues come back: failure modes, unclear boundaries, things going wrong in ways that are not always visible.

So it no longer feels like the problem is "can we build this".

It feels more like the real problem is whether we can still understand and operate what gets produced.

I wrote this to make that framing clearer.

Curious if others are seeing something similar.

j / k navigate · click thread line to collapse

4 comments

abraxas27d ago

piljoongOP27d ago

abraxas27d ago

piljoongOP27d ago

I've been using coding agents a lot recently.

Building got faster.

But when the system actually has to run, the same issues come back: failure modes, unclear boundaries, things going wrong in ways that are not always visible.

So it no longer feels like the problem is "can we build this".

It feels more like the real problem is whether we can still understand and operate what gets produced.

I wrote this to make that framing clearer.

Curious if others are seeing something similar.

j / k navigate · click thread line to collapse