Earlier quoted context omitted.
But the architecture is the important (and hard) part!!! Not the UX!
A customer doesn't care about architecture. They want a good UX.
The customer wants a working product!
Nobody knows how to build with AI yet
411–420 of 432 posts
Re: Nobody knows how to build with AI yet
#412Earlier quoted context omitted.
An unhealthy attachment to determinism will turn out to be a career-limiting hangup, I suspect. You already lack insight into how 100% of the code in your project works, unless you only work on trivial projects. Did you think that state of affairs was going to get better with time? As usual, TDD covers a multitude of sins. As for "autocorrect," let us know when your "autocorrect" takes gold at the International Math…
Talk is cheap. Give your LLM/agent your badge and let it turn in 100% of your job.
That'd be awesome. Not going to happen this week or this year, but it will.
Re: Nobody knows how to build with AI yet
#413Earlier quoted context omitted.
Yeah; there's still a massive chasm between "I spent hours precisely defining my requirements for this greenfield application with no users and the AI one-shot it" and "million line twenty team enterprise SaaS hellscale with ninety-seven stakeholders per line of code". The fact that AI can actually handle the former case is, to be clear, awesome; but not surprising. Low-code tools have been doing it for years. Retool…
> Dario Amodei says we'll have the first billion dollar solo-company by 2026 [1]. I lean toward this not happening. Why not? Not like companies have to actually do anything beyond marketing to get insane evaluations… remember theranos?
Theranos peaked at ~800 employees, FYI.
Reason 1: Because any company with one individual who leverages AI to achieve a billion dollar valuation would trivially & obviously be more valuable and be able to achieve more if they had two people leveraging AI. And, at a billion dollar valuation; why not pay that extra salary? Why not add a third? There's a huge difference in potential output by building with a founder team of 2 or 3 people versus 1; with not that much difference in cost.
Reason 2: The most capital efficient valuations in history are B2C companies (Instagram & Whatsapp are maybe the two best examples) during the VC boom era of the 2010s. B2C is naturally very capital efficient; you can do inbound marketing, no sales teams, just build. But, B2C success stories are more-and-more rare; the name of the game in the 2020s era has generally been B2B. Cursor might be the fastest company to reach $500M in revenue, and it got there on B2B, Notion isn't building its AI tools to sell to its consumer customers, etc. B2B is a lot harder; it requires outbound marketing & sales, as well as customer-led product development that usually requires interacting with people. AI will absorb some of those roles, but to suggest it can widdle down to one person per billion dollars in valuation in a year feels too accelerated to me. The world is complex, old, and crusty.
Just look at the YC batches for 2025 [1]: Of the 375 companies in the three batches, 20 are Consumer tech (~5%).
Reason 3: Weaker, but something I think about: AI is weirdly non-differentiating/egalitarian. You see some people try to differentiate it with crazy prompts or context engineering, but then next month Cursor ships an update and suddenly those aren't differentiating anymore. If you're an investor in a single-person unicorn-moonshot, you have to ask: Why can't some other company just come in and do the same thing you're doing? If that much can be automated through off-the-shelf systems. My feeling is that this concern would lead to lower revenue multiples on the valuation, which just makes it that much harder to hit unicorn status.
[1] https://www.ycombinator.com/companies?batch=Summer%202025&ba...
Re: Nobody knows how to build with AI yet
#414this isn’t vibe coding. This is something completely new. I call it “flex coding.” heck I built a full app in an afternoon AND I was a good dad? > I'd wander into my office, check what Claude had built, test it real quick. If it worked, great! Commit and push. "Now build the server connection UI," I'd say, and wander back out. Made breakfast. Claude coded. Played with my son. Claude coded. Watched some TV. Claude cod…
This is all very emotive and I'm sure is a dream many of us would love to live. But does Claude's code work? Does it work to the level where you'd depend on it yourself; where you'd bill customers for it; where you'd put your reputation behind it? I say no. And it's because I use Claude. Two events changed how I use Claude: now it's an advisor, and I mostly type the code myself. Because I don't trust it. First, I cau…
I had Claude, as an agent, get into an infinite loop trying to resolve a diagnostic in Java test that any human would easily figured out and fixed in a second.
Re: Nobody knows how to build with AI yet
#415Earlier quoted context omitted.
in my experience only a limited part of software can be done with just really clear specs, also at times in my career I have worked on things that became more "clear" what was really needed as time went on the more we worked on it, and in those cases really clear specs would have produced worse outcomes.
Which is the real reason agile is so much more effective than waterfall. The beginning of the project is when you know least about your project, so naturally you should be able to evolve the specification.
You are confusing waterfall with BDUF.
Re: Nobody knows how to build with AI yet
#416Re: Nobody knows how to build with AI yet
#417Earlier quoted context omitted.
I definitely did a double take when I got to this section. I am neither an AI optimist nor an AI pessimist (probably slightly on the optimistic side of the midpoint) but this sounds insane to me for any software that people might truly depend on. Five minutes of review for 10,000 lines, happening multiple times per day?
I did some digging, and this seems to be the project the author is using AI to work on: https://github.com/sublayerapp/sublayer You look at the PRs... there are 786(!) AI generated pull requests and an associated AI generated code review for each one. Each PR is about ~20-100 lines of Ruby (including comments) that implements an "action" for the sublayer system as a Ruby class. So probably something that could be han…
He deleted the PRs? Struck a nerve
Re: Nobody knows how to build with AI yet
#418My theory on AI is it's the next iteration of google search, a better more conversational, base layer over all the information that exists on the internet. Of course some people will lose jobs just like what happened to several industries when search became ubiquitous. (newspapers, phone books, encyclopedias, travel agents) But IMHO this isn't the existential crisis people think it is. It's just a tool. Smart, clever…
Yeah; there's still a massive chasm between "I spent hours precisely defining my requirements for this greenfield application with no users and the AI one-shot it" and "million line twenty team enterprise SaaS hellscale with ninety-seven stakeholders per line of code". The fact that AI can actually handle the former case is, to be clear, awesome; but not surprising. Low-code tools have been doing it for years. Retool…
AI is still in an experimental phase for many teams, especially when it comes to handling complex, long-term projects. For PMs and EMs, the cost-benefit analysis of AI credits vs. manual tasks is a big concern before fully committing to AI adoption. Some teams have seen great success, particularly in areas where speed and flexibility are key, but others are still waiting for clearer ROI before diving in. It’ll be interesting to see how the balance of risk and reward evolves as AI tools mature.
Re: Nobody knows how to build with AI yet
#419Reminds me of Sussman’s “We don’t know how to compute” which is essentially maligning that we haven’t gone full Smalltalk with some Lisp sprinkled on top to get “soft computers.”
None of that is actually what makes my favorite tools work. It’s usually some nerds that never stopped using C/C++ and really know hardware.
Re: Nobody knows how to build with AI yet
#420Earlier quoted context omitted.
Isn’t that like „the dog speaks English, but makes occasional grammar mistakes“? Give it two years.
It’s been two more years for the past three years. What’s going to happen in those two years? Are we going to find 10 more internets to train on?
two years ago it didn’t do anything that made me want to use it during work.
The data wall idea of needing 10 more internets is like the peak oil theory of the 2000s, imho. I don’t think 10 more internets are required, and all those tasks like proper security and i18n, none of these require creativity of any kind. These are ideal candidates for LLMs to solve quite soon.
But to talk about the data wall more: Current SOTA LLMs are trained on circa 30-50T tokens. First, that is not a full internet‘s worth; it is estimated that Meta owns more like 200T tokens of user data etc. Second, the work in optimizing for data efficiency is just beginning.