I have a side project which is an experiment to build an interesting quick UI for local AI. As part of it I want a very very specific, interesting look involving shaders, animations, and so on.

I was trying to just get a prototype in place by prompting and it was going nowhere, just constant yo-yo'ing and never really getting what I wanted. This also was quite de-motivating and I found myself "yelling" at the model.

So I told Codex:

- Make this API first-class in our framework, with easy parameters (it had been sort of a hacked low-level thing)

- Add hot reloading to our system so I can edit it without any state loss or refresh

- Give me more knobs (X, Y, Z) so I can tune everything here as I need

- Add a HUD that lets me also drag sliders to tweak the same things

And I got my desired look within a few seconds.

The principles of good design and products have always been this btw, you need your feedback loop to be as tight as possible. Good design has always come from the ability to iterate incredibly fast, your brush needs to move precisely with your hands, and can't have delay from the time you put it down to the time the stroke shows up.

They released more major features and breaking changes in their last patch release than most software sees in two major versions.

I've been using it just as a script runner and npm package manager basically, and it's incredible the amount of work you have to do to find "good" versions. We've had patch versions suddenly freeze on install more than once, we couldn't upgrade for quite a while due to this. I think they broke postinstall scripts with trustedDependencies entirely two minor versions ago - not a mention in release notes, and somehow no one reporting it in GH issues. In 1.1 or so you could get Bun to do trustedDependency builds in postinstall, and then after that you couldn't. I looked around for release notes and saw nothing mentioned. It's been broken for months.

Comments URL: https://news.ycombinator.com/item?id=47927186

Points: 1

# Comments: 0

https://gist.github.com/natew/fce2b38216edfb509f7e2807dec1b6...

I've had 0 issues with Codex once it adopted it. I use it for Claude too, which seems to also improve its continuation.

It was revised for friendliness based on the Anthropic paper recently, I'd have been a lot less flowery otherwise.

And so if you ask it to do something big it will do a very surface level implementation. But if you have it iterate many times, or give it small pieces each time, you’ll end up with something closer to what a human would do.

I imagine the pelican test but done in a harness that has the agents iterate 10+ times would be closer to what you’d expect, especially if a visual model was critiquing each time.

.6 is some sort of quantized or distilled .5 with a bit more RL, and the current .5 is that same cost reduced model without the extra RL.

One example is I have a multi-stage distillation/knowledge extraction script for taking a Discord channel and answering questions. I have a hardcoded 5k message test set where I set up 20 questions myself based on analyzing it.

In my harness Minimax wasn't even getting half of them right, whereas Sonnet was 100%. Granted this isn't code, but my usage on pi felt about the same.