Max Liani

The Food and Drugs Administration initiated a review to request adoption of nutrition facts labeling for any catalogued celestial body.

@abbatordev Hehe, it’s part of my series of incomplete publishing. The way I imagined it, part 2 would have been a list of utilities to deal with containers, like adding/removing an element at an index in a vector. 5 years? Uff I feel old.

@maxliani I'm wondering if part 2 of 'Undo, the art of' is ever coming. I've only been waiting 5 years. Really enjoyed part 1, though.

@svadrb @ShinjiOgaki It’s like teaching a junior dev to not do what they were taught in college. With the difference the AI won’t learn, so you have to re-brief every single time.

@svadrb @ShinjiOgaki That depends a lot on your codebase, they domain of the problem, your coding style, etc. For instance, I don’t like abstractions, helper functions, accessors, and a lot of the design patterns stuff… by default the AI produces that galore.

It's pathetic, but at this rate, the day I can no longer code—much like how smartphones made me forget Kanji—is just around the corner. I'll stop using coding agents for my hobby projects.

@ShinjiOgaki 💯 I think the productivity boost has to be found in everything we don’t want to do but still is part of the job. For the core competence we care about, that has to be done by us, otherwise there is no craft anymore.

@maxliani Thanks for the comment. I loved using it at first too, but later struggled to fix bugs because I didn't remember the code the agent wrote. I lose attachment to it, and it feels less fun. It's a very strange experience.

@ShinjiOgaki I still code the vast majority of what I produce. I let the agents find out information in other people’s contributions, or try to debug for a bit, but then I rewrite the majority or it. I have found that the code I produce tends to be 1/3 or less LOC of what the agent generated.

@ShinjiOgaki For a period I thought AI for personal project was the factor that would have given me the satisfaction to see the project progress. Truth is the opposite happened. Now I am stuck and haven’t opened my project in months.

@RHeescher @skdh @yinzer_g

@skdh @yinzer_g and? What's the point? I hope you didn't need a pocket calculator for it. lol

99.99% of the world's population can't tell you which of these flags is Belgium.

@dansanburu それらは焼いたものではなく、中にツナが入っています。一方、先週作ったこちらはひよこ豆入りで、フライパンで焼いたものです。外側のご飯の、少しカリッとした食感が気に入っています。

@maxliani 具も入っているんですか。本格的ですね。 BBQグリルなどで焼いて焦げ目をつける 焼きおにぎりも美味しいですよ

Coming back from Japan, I got in the habit of cooking onigiri.

@dansanburu はい、醤油とごま油で味付けしたおにぎりで、具はツナです。毎回違うバリエーションで作るようにしているんです。🙏

@maxliani 醤油おにぎりですか？ 形も良くて美味しそう。 最初この写真を見た時に CGかと疑ってしまいました。 とても良い質感だなあと

@renderwonk Relativistic, that’s all.

Holy crap. Codex just implemented a camera zoom-out by *dynamically shrinking the scene* - rebuilding the scene slightly smaller each frame, instead of just adjusting the camera matrix. And that’s with gpt-5.5 high. Sigh.

@cmuratori 0xDEADBEEF — A bold scent with hints of corrupted memory.

Everyone who subscribes to my Substack knows that we do not accept advertising or sponsors. But I would absolutely make an exception for this. I think we can all agree the world needs more designer fragrances formulated with the low-level programmer in mind!

@esmolikowski 💯 Personally, I still think that using Cursor to “codevelop”, while letting agents run free to “find out and come back with information” is still the sweet spot for me.

Having said that it is really difficult to know when to not use AI tools. And it is sometimes super exhausting to use them. Chatting with claude the whole day and reading code is actually much harder than writing the code yourself. Mostly because on how the brain works.

For some tasks, I find using AI tools to be extremely productive. I get about 100-300x speed up to write small automation scripts, parsing logs, simple refactoring. But there are some areas where AI is actually much slower. Mostly because it keeps writing the wrong code.

@MadMaraca Seeing that box, my mind went immediately to render scaling tests I was doing back in the days (when I was working on rendering tech for artists). It looks like a box from far... pixel size, high resolution models... hundreds of millions of them :) Screenshots from 2014.

Sketch >>> ( Guess the Final ) 🙃

Hollywood has spent nearly $600 million trying to bring Matt Damon home. Have they considered maybe he just doesn’t want to come back?

@SebAaltonen Once I had two sessions on the same presentation (sponsored sessions) one was a 20min the other 40. Half way through the 20min session I realized I was presenting with the 40min deck 🤦🏻‍♂️

The most painful thing about SIGGRAPH presentations: Your slot is 40 min. Your presentation is 48 mins. You are still missing the last slide... Making a presentation about a 50 page blog post is PITA :D

@esmolikowski If I may ask, what type of model/harness/claw do you use?

I’m running into more issues these days where the agents are straight up gaslighting me. They tell me they’re done with a task. Tests fail. They ‘forgot’ to write half the code.

@ocornut I tend not to use goto at work, mostly to avoid lengthy discussions during code review… But in my own code, where I know nobody can mess it up, I do and sometimes in unconventional ways. Example I posted in the past:

Stumbled on the only 'goto' in my code today. Recalled how it took me an unreasonable amount of minutes to accept adding it back then. Partly swayed by decades of unwarranted bad rep; partly because I didn't fully grok what it implied for the compiler dealing with constructors.

@Jiayin_Cao It’s the right technique. You let the door open to use SIMD to vectorize shader execution of sorted batches and get real performance improvements. I did some Siggraph presentations with Alex Welsh a while ago about this (when I was still leading the RenderMan XPU effort)

@maxliani Exactly what I did. I took some detour by making an attempt to SIMD the whole shading language first, only to realize that I had to introduce too many new concept to deal with divergency. The AD solution is quite neat. Though it has its problem, but it should be good for my use

After years of thinking about it, I finally got a reasonable approximation for mipmap selection working in my renderer. The biggest challenge, by far, was supporting procedural UV generation in my shading language. I believe I have enough content to put down a blog post next.