Discussion
Oy! Remember "GLSL Optimizer"? Yeah last time I did anything there was 10 years ago. Well, just spent like four days adding compute shader support as well as moving underlying Mesa compiler/optimizer code in there from 2014 to 2026 (that's... alot of changes). Now it also is based on their "NIR" optimization pipeline.
Changelog: https://github.com/aras-p/glsl-optimizer/blob/main/Changelog.md
Now onto controversial parts: ...
... there were a lot of moments when I was "geez why the heck did you do this", but there also was a handful moments of "oh, nice, I would have not thought of that".
Now, yes, sure, the whole lollum industry is still plagued with issues from business models to ethics to concentration of power to <insert all the other reasons>. Still, for doing this type of work, it would have taken me, several weeks perhaps? A month? Now it was casual babysitting for four days, and maybe 100 euros.
…would I have done it at all “manually”? Definitely not, I don’t need glsl optimizer, and TBH probably no one else does at this point either. So for me, this was more of a “ok let’s see what the lollums can do” at a task like this. Impresions: far from ideal, but quite positive. Feelings: very mixed and conflicted.
Oh well, time to get food ;)
@aras I'd argue boycotting is a good strategy at the moment, because IMHO success for these companies is a dystopian outcome.
This includes the doom loop we're looking at right now with the Internet and information collapse, in addition to the climate issues: they are building *fields* of gas turbines.
It also includes a world where these companies act as gatekeepers for any sort of software development even if they work, and the world in which software quality plummets if they don't work well.
@aras thanks for the transparency and honesty.
If I'm honest this all speaks to how something "working" is so incredibly distracting to us as engineers. The thing clearly can produce solutions to technical problems, and that simple fact pulls in the attention of a problem-solving brain like gravity. Every single other aspect that is hidden by that pull is dangerously tainted, and yet we must validate if it works first.
Where does that leave us? Well, that's an exercise for the individual.
...I used the lollums (opus 5.5 and gpt 6.1-sol specifically) to help me do that "let's integrate 12 years of upstream changes, and move to whole new optimizer pipeline, and cover it all with tests, and evaluate the results" part. And while the lollums definitely are way less annoying and "stupid" than they were a year ago, this was very far from "one shot / vibed" effort. ...