4 ms·

Maybe you dashed this off and I'm digging into a comparison that wasn't meant for deep scrutiny. But ripping wood is an extremely small part of carpentry. The A

Maybe you dashed this off and I'm digging into a comparison that wasn't meant for deep scrutiny. But ripping wood is an extremely small part of carpentry. The AI hype is literally "have an LLM eat a bug report on GitHub, wait a little while, review the bug fix PR, also this is the least capable this will ever be." That's kind of the whole thing. Sure there's system design, requirements gathering, blah blah blah, but the whole product is legitimately "SWE too expensive? Get Claude!"

simonw 1y ago
That's my point: I'm pushing back against the AI hype here, and arguing that the part of our job that can be done by an LLM is far from being everything we do.
- camgunz 1y ago
  Fly.io just shipped Phoenix.new which is legitimately "do the whole thing super fast (as long as you could already mostly do the whole thing at regular speed)". I really don't think this is hype.
  - simonw 1y ago
    Phoenix.new is great, but you still need to know a lot about web development to use it effectively. https://simonwillison.net/2025/Jun/23/phoenix-new/ https://simonwillison.net/2025/Jun/23/phoenix-new/
    - camgunz 1y ago
      You're right and I agree: in order to successfully use LLMs to build software products of any significant complexity, you need an experienced dev at the helm. My guess is this is the case for another couple years max, but who knows really. Reading your (as always, great) post, it seems like Phoenix.new did like 95% of the work? You had to coach it around FTS, but I also have to look that up every time so :shrug:. It's pretty bananas; I still maintain this is nothing like "traditional software engineering but one of your tools got way more efficient", i.e. you got a table saw. This is "I typed 'build a table' into a box, it couldn't figure out dovetailing, I taught it dovetailing, and I had a table." Regardless of whether or not you think that's good or bad, I think it's clear it's a huge change.
      - simonw 1y ago
        Yeah, it did almost all the work for that demo but that's mainly because I let it - I didn't have the time or inclination to take over and knock the code into the actual shape I wanted. I also think that even the few prompts I gave it require pretty deep knowledge and experience - many professional software engineers don't know that SQLite has good FTS baked in!
        camgunz 1y ago
        Wow it's one of the best things about SQLite, truly tragic. Another point in your favor I think is like, the coding (deployment, etc) isn't usually the bottleneck. Like in a given 2 week sprint I'm probably spending... max 24/80 hours coding, so something like a 30% ceiling. The rest is like, reqs, design, shop around for buy in, mentoring, politicking, bug hunting, planning, etc etc. The problem is most of that stuff isn't nearly as enjoyable as the coding part. Why didn't we invent a bot that talked engineering leadership out of terrible ideas? That's where the real gains are haha.
globular-toast 1y ago
> also this is the least capable this will ever be Don't be so sure about this. I can think of countless examples of technologies that are less capable than they used to be either in specific areas or in general. For example, telephones used to be very low latency with graceful quality degradation. Now they are high latency and just cut out with a poor signal. Cars used to be a luxury affording high levels of personal freedom, now you just get stuck in traffic and struggle to park everywhere. If AI actually happened then of course it would change everything. It could even destroy us, and I don't see why it wouldn't. But LLMs are not that. We don't know how this is going to play out after several iterations of LLM generated code is pumped out and we can't distinguish LLM-generated from human-generated. This might well be the best it will ever be.

simonw 1y ago

That's my point: I'm pushing back against the AI hype here, and arguing that the part of our job that can be done by an LLM is far from being everything we do.

camgunz 1y ago
Fly.io just shipped Phoenix.new which is legitimately "do the whole thing super fast (as long as you could already mostly do the whole thing at regular speed)". I really don't think this is hype.
- simonw 1y ago
  Phoenix.new is great, but you still need to know a lot about web development to use it effectively. https://simonwillison.net/2025/Jun/23/phoenix-new/ https://simonwillison.net/2025/Jun/23/phoenix-new/
  - camgunz 1y ago
    You're right and I agree: in order to successfully use LLMs to build software products of any significant complexity, you need an experienced dev at the helm. My guess is this is the case for another couple years max, but who knows really. Reading your (as always, great) post, it seems like Phoenix.new did like 95% of the work? You had to coach it around FTS, but I also have to look that up every time so :shrug:. It's pretty bananas; I still maintain this is nothing like "traditional software engineering but one of your tools got way more efficient", i.e. you got a table saw. This is "I typed 'build a table' into a box, it couldn't figure out dovetailing, I taught it dovetailing, and I had a table." Regardless of whether or not you think that's good or bad, I think it's clear it's a huge change.
    - simonw 1y ago
      Yeah, it did almost all the work for that demo but that's mainly because I let it - I didn't have the time or inclination to take over and knock the code into the actual shape I wanted. I also think that even the few prompts I gave it require pretty deep knowledge and experience - many professional software engineers don't know that SQLite has good FTS baked in!
      - camgunz 1y ago
        Wow it's one of the best things about SQLite, truly tragic. Another point in your favor I think is like, the coding (deployment, etc) isn't usually the bottleneck. Like in a given 2 week sprint I'm probably spending... max 24/80 hours coding, so something like a 30% ceiling. The rest is like, reqs, design, shop around for buy in, mentoring, politicking, bug hunting, planning, etc etc. The problem is most of that stuff isn't nearly as enjoyable as the coding part. Why didn't we invent a bot that talked engineering leadership out of terrible ideas? That's where the real gains are haha.

globular-toast 1y ago

> also this is the least capable this will ever be Don't be so sure about this. I can think of countless examples of technologies that are less capable than they used to be either in specific areas or in general. For example, telephones used to be very low latency with graceful quality degradation. Now they are high latency and just cut out with a poor signal. Cars used to be a luxury affording high levels of personal freedom, now you just get stuck in traffic and struggle to park everywhere. If AI actually happened then of course it would change everything. It could even destroy us, and I don't see why it wouldn't. But LLMs are not that. We don't know how this is going to play out after several iterations of LLM generated code is pumped out and we can't distinguish LLM-generated from human-generated. This might well be the best it will ever be.