> If you weren't able to build it before then you aren't able to properly verify it
GP said "couldn't dream of building". To make an analogy, I wouldn't dream of building a house if I had to do every bit of brick-laying myself. But with others doing what I can't do, I can indeed dream of building a house and I do know how to verify that the house was built to satisfy my dream.
This analogy is ironic to me. For most people, building a dream house and verifying it means hiring someone else who is accountable for it being built to code etc.
This is exactly the naive gap that is happening with AI. There is no trusted middleman who does the verification and takes the burden of accountability. The AI vendor tries to sell it like magic but with a disclaimer that is is merely entertainment. The user tries to pretend they are reviewing, but often is just implicitly trusting the tool to do right by them on things they have no real capacity to evaluate.
> There is no trusted middleman who does the verification and takes the burden of accountability.
It's a great example of a project that needs accountability, but there's also thousands of other projects that need absolutely no accountability.
It's all relative to the stakes. The lower the stakes the less informed verification is needed.
The real trick is being informed enough on the boundaries to know where it matters. In the construction analogy, you need to be informed enough to know that a house is a bad idea since there are safety concerns. However building something small and non-load-bearing is probably fine to "vibe".
Eg there's no expertise needed to judge a garden trellis or 2 foot picket fence. It either works or it doesn't, and if it fails down the road there's no harm.
A house is a good analogy here. You can verify that a house is built to satisfy your dream. You can't likely verify that a house won't collapse in the middle of the night or during a storm/quake/etc. Once no human understands how the house is built anymore, we will need other ways to verify that or just trust it's done right.
Ok, so before I couldn't (/didnt have the time to) build cross-platform native desktop applications with minimal amount of 3rd party dependencies, but now I "can". What exactly am I unable to verify here?
I still don't have the skill/knowledge to build cross-platform native applications the way I'm doing it with an agent now. I am still able to verify it's correct. How is that possible if what you say is correct?
> Clicking around a finished UI is not verifying
Definitively a part of it, if you care about having mouse interactions and having graphical UIs.
Depends on the scope. I’m slopping some projects (graphics related) that I’d never have learned before. Mostly because of the huge lift, huge time investment, etc. is it quality code? I’m sure not. But verification of the application is quite easy, even if the underlying graphics and gpu shaders could be done better.
What I’ve learned is quite minimal due to the slop nature. However I’d have never even started this project in the past due to time, so it’s still fun being able to learn small bits of projects I’d otherwise not even touch.
GP said "couldn't dream of building". To make an analogy, I wouldn't dream of building a house if I had to do every bit of brick-laying myself. But with others doing what I can't do, I can indeed dream of building a house and I do know how to verify that the house was built to satisfy my dream.