That's always been the justification to skip this phase :). Tools have just changed. One-person to small-team wonders that could code and build directly made the same arguments.

My point isn't to stitch things to Figma, that's abhorrent to me as well. My point is to not get bogged down on the implementation details, in this case an actually working DB, those tables, etc, but rather less fidelity actual full flow concepts that can be generated and iterated.

Then that can be fed into a magic genie GPT that generates the front-end, back-end, and all that good jazz.

If the effort to produce websites goes tends to zero, the value of websites will surely tend to zero. Either issues with security and maintainability will be a break on this tendency or we will get to a point where generating a custom website will be something trivial that will be done on demand.

The thing is, the cost of producing websites is already pretty low, but the value of websites mostly derives from network effects. So a rising flood of micro crud saas products will not be likely to generate much added value. And since interoperability will drive complexity, and transformer based LLMs are inherently limited at compositional tasks, any unforeseen value tapped by these extra websites will likely be offset by the maintainability and security breaks I mentioned. And because there is a delay in this signal, there is likely to be a bullwhip effect: an explosion of sites now and a burnout in a couple of years in which a lot of people will get severely turned off by the whole experience.

If you need a website that needs prototyping in 2025, you're probably doing it wrong (eg launch on insta or something). But anyway, you can vibe iterate, and not just small iterations, but wholesale different value props and approaches, so why not. it's tangible, easier to test, and you get more meaningful feedback. I do this and it's 3-4x faster than working with a designer. And to be sure, we're not making websites, but protyping features into a saas app to test with users and ourselves.

The value of Amazon.com is not the cost to produce the HTML and JavaScript you see when you visit that website. It is a component of the Amazon business, and to Amazon it is extremely valuable, and to everyone else it would be almost worthless.

If someone has the idea for the next Amazon, as well as everything else you need beyond the idea, and tools like Supabase and Lovable allow them to get it off the ground, those tools are incredibly valuable to that person.

If someone’s ideas are worthless, their websites will be worthless.

Don’t get me wrong, I love Supabase, but

> you can vibe code a fully working UI+backend

…is gonna bring a lot of houses crashing down sooner or later.

I couldn't agree more. "Vibe coding" is pretty cool, but it's not sustainable at least with with current technology. You're much better off being a knowledgeable developer who can guide an an LLM to write code for you.

One thing I will agree on though is that LLMs make it easier to iterate or try ideas and see if they'll work. I've been doing that a ton in my projects where I'll ask an LLM to build an interface and then if I like it I'll clean it up and or rebuild it myself.

I doubt that I'll ever use Figma to design, it's just too foreign to me. But LLMs let me work in a medium that I understand (code) while iterating quickly and trying ideas that I would never attempt because I wouldn't and be sure if they'd work out and it would take me a long time to implement them visually.

Really, that's where LLMs shine for me. Trying out an idea that you would even be capable of doing, but it would take you a long time. I can't tell you how many scripts I've asked ChatGPT or similar to write that I am fully capable of writing, but the return on investment just would not be there if I had to write it all by by hand. Additionally, I will use them to write scripts to debug problems or analyze logs/files. Again, things that I am perfectly capable of doing but would never do in the middle of a production issue because they would take too long and wouldn't necessarily yield results. With an LLM, I feel comfortable trying it out because at worst I'd burn a minute or two of of time and at best I can save myself hours. The return on investment just isn't there if it would take me 30 minutes to write that script and only then find out if it was useful or not.

LLMs are better search. Google burned down the house to keep itself warm and held off on LLMs until it was inevitable and are now pulling up ahead. This is the logical conclusion. LLMs will be monetized and enshitified by ads.

Soon, some free smart LLM code generators will stop generating certain outputs and instead suggest using commercial components that have paid for promotion.

The whole point of Supabase is not needing to vibe code the backend part.

PostgREST is quite boring, open source, proven tech.

So they say; I still haven't seen any high quality vibe coded software, and I'm pretty sure I never will.

I’ve seen tons of low quality successful software (concur anyone). Clearly being well built is not a requirement for success.

I'd split the difference and say the cons that low quality software creates only matter sometimes.

E.g. Concur is primarily feature-complete and will only ever need to evolve gradually.

So the drawbacks of being brittle, kludged-together, and incapable of making rapid feature changes doesn't really matter.

In some other products, that matters a huge deal.

So the tl;dr is, as always, optimize for the things that actually matter for your particular situation.

This can only be true of some products. Often there are a lot of concerns like privacy, white labeling, legal consequences that need to be considered _before_ you vibe code.