One nit: they’re good at writing code. Specifically, code that has already been written. Software Engineers and Computer Scientists still need to exist for technology to evolve.
This. Was setting up a new service and it scaffolded all the endpoints after the swagger and helped me setup tooling, tests, within a few hours. Also helped me research what has happened in the area since my last ms.
Now when adding the business logic I’ll be doing most of it myself as it tends to be a bit creative about what I’m trying to achieve and tends to forget to check my models etc.
It’s great at generic code, has issues on specifics.
I feel like if your code is so generic a generator can make it, you could achieve tge same results faster, more reliably, and more energy-efficiently with a shell script or two.
A specific tool should definitely beat a generic one. If I was doing these things all the time I would consider building something like that, scaffolding based on a swagger seems pretty easily achievable but since I do this every other year tops, and the setup will need to be updated with new techniques it’s fast from a valuable time investment to write for me.
It is pretty bad at things that are “black boxes” that require documentation to analyze. For instance, I was trying to debug an SSL issue with DB2 (IBM database) and chatgpt and copilot gave conflicting answers. They frequently gave commands that didn’t work, with great confidence of course. I had to keep feeding errors back to it. I even had to remind it that I was working in Linux and not Windows.
FWIW, ChatGPT and Copilot are two of the worst AIs out there for things like this. At many gigs I’ve had they’re outright banned for use because of how garbage they are.
Claude Code, or Claude in general, notably Sonnet 4.5 and Opus 4.5
Gemini also solid, though for coding found it lesser than Claude, but for heavy inference and reasoning it can be great and also supports a larger context window
It’s undeniable that AI is great at problems with tight feedback loops, like software engineering.
Most jobs don’t have the tight feedback loops that software engineering has
I, CandleTiger, do hereby deny that AI is great at software engineering.
it is totally deniable. Because it’s simply not true. It’s been studied.
One nit: they’re good at writing code. Specifically, code that has already been written. Software Engineers and Computer Scientists still need to exist for technology to evolve.
This. Was setting up a new service and it scaffolded all the endpoints after the swagger and helped me setup tooling, tests, within a few hours. Also helped me research what has happened in the area since my last ms.
Now when adding the business logic I’ll be doing most of it myself as it tends to be a bit creative about what I’m trying to achieve and tends to forget to check my models etc.
It’s great at generic code, has issues on specifics.
I feel like if your code is so generic a generator can make it, you could achieve tge same results faster, more reliably, and more energy-efficiently with a shell script or two.
A specific tool should definitely beat a generic one. If I was doing these things all the time I would consider building something like that, scaffolding based on a swagger seems pretty easily achievable but since I do this every other year tops, and the setup will need to be updated with new techniques it’s fast from a valuable time investment to write for me.
It is pretty bad at things that are “black boxes” that require documentation to analyze. For instance, I was trying to debug an SSL issue with DB2 (IBM database) and chatgpt and copilot gave conflicting answers. They frequently gave commands that didn’t work, with great confidence of course. I had to keep feeding errors back to it. I even had to remind it that I was working in Linux and not Windows.
FWIW, ChatGPT and Copilot are two of the worst AIs out there for things like this. At many gigs I’ve had they’re outright banned for use because of how garbage they are.
Which ones have you had recommended?
Claude Code, or Claude in general, notably Sonnet 4.5 and Opus 4.5
Gemini also solid, though for coding found it lesser than Claude, but for heavy inference and reasoning it can be great and also supports a larger context window