Egregoros · Phoenix Framework

Post

Remote status

Context

foolghool

Currently trying Mistral Vibe to see if it can autonomously write the missing pieces for a Go project I'm sick to work on: HTTP handlers, missing bits from the "service" layer, entry point, and a few things in between.

It's a basic ass web service (a KMS over a set of plain text notes) so not a daunting task compared to what I read other people do with agentic AI but it's scifi-tier to see a bot writing valid code, testing and fixing what's wrong.

It produced me a compilable project while I was writing (so 5min top). It 404, but this still blows my mind.

Harblinger

@harblinger@wizard.casa remote

@foolghool I should mess with CLI AI again maybe, last year I experimented with Claude sonnet 3.6 wired up to https://aider.chat/ but I ended up just going back to the web interface and copy pasting. The slower pace helps keep things on track. Releasing local AI agents on some toy project might be fun though

Replying to @harblinger@wizard.casa

lain, author of the quixote

@lain@lain.com remote

@harblinger @foolghool I can assure you that ChatGPT can write you nearly any piece of software with high quality if you give it a good feedback loop (tests, specs, screenshots…)

Replies

Replying to @lain@lain.com

@i@declin.eu remote

@lain @foolghool @harblinger especially now that dirt cheap 1T chink models exist that do about as well too, let opencode and kimi2.5 lazily rewrite my personal fedi frontend without too many issues

Replying to @i@declin.eu

lain, author of the quixote

@lain@lain.com remote

@i @foolghool @harblinger yeah, everyone can have personalized software now

Replying to @lain@lain.com

Harblinger

@harblinger@wizard.casa remote

@lain @foolghool I still haven't tried any of the OpenAI models
What agent manager/interface(?) do you plug your API key into?

Replying to @harblinger@wizard.casa

lain, author of the quixote

@lain@lain.com remote

@harblinger @foolghool just use codex

Replying to @lain@lain.com

foolghool

@foolghool@wizard.casa remote

@lain @harblinger So far I had better success by using Claude or Mistral as a "consultant" for when I have doubts about how I should implement something or how to rewrite part of the program I knew to be shoddy. I learned a lot by doing that. I set up gpt.el in Emacs for code and tend to use the web interfaces for questions that don't need a lot of context

The code produced by Mistral Vibe was getting a bit sloppy toward the end, like if it started missing some of the context along the way. I did nothing special regarding the feedback loop so that's most likely the issue

I did not try anything by OpenAI or Google yet, surely I must miss out but I slowly catch up with the new world

Replying to @foolghool@wizard.casa

lain, author of the quixote

@lain@lain.com remote

@foolghool @harblinger indeed you are missing out. I wouldn’t want to go at a slow pace right now

Replying to @lain@lain.com

foolghool

@foolghool@wizard.casa remote

@lain @harblinger Judging by what I can read on the WWW it seems Mistral quickly lost popularity a bit after they released their pro paid service, I don't understand exactly why but I guess it's just that everything else on the market is just noticeably more performant (it's obvious compared to Sonnet). I'll try Codex today to write the missing tests from my program