Max Liani

@savage309 LoL. History on repeat

codex has issued a salutation, cancel or allow youtube.com/watch?v=8Cwolu…

@SegfaultMax @marcosalvi The trick may be in managing the transition. I mean, I am ok if the AI produces untamable code IF the AI can also debug it, profile it, optimize it etc, at that point I don’t care what code looks like. But right now I need code that works for “me in the loop”.

@maxliani @marcosalvi same, but I know that these high standards will be untenable soon, outside of hobby projects the demand to herd agents to output 100x sufficient/mediocre code is already deafening

Pair vibe coding workflow: Codex writes ~99% of the code, Claude reviews and improves it. Claude also handles most of my docs and specs.

@marcosalvi Often times the code after my cleanup is 1/3 or less of the amount of code generated.

@marcosalvi I have to say I am demanding on the quality of the result. Often time what comes out automatically seems to work, but I take a look and say “not a chance I’d commit that”. Again, some other times is “looks awesome”

@tomcr2100 I was trying to be motivational 😆

@maxliani youtu.be/XtIuC0NAF_E?si…

Now is when the future becomes history.

@iquilezles @tr_babb @wookash_podcast The hierarchical subdiv edits never took on. No modeling/rigging tool supported it. It’s impossible to rig a very low polycage and get fine details for character animation, so with time polycages became more and more complex.

@maxliani @tr_babb @wookash_podcast Culturally you were supposed to work with CC subdivs (low-res-ish cages). Not that an artists couldn't still create a super dense cages, so you are right.

The concept that with Nanite-style virtualized geometry you get to the point, where triangles approximate pixels, with overdraw reduced to ~zero = which means you have ~constant body of work per frame, no matter the scene, is just crazy. I can't get over it, how cool this is

@tr_babb @iquilezles @wookash_podcast the big problem in REYES wasn’t occlusion culling, was shading dense visible grids. You had densely modeled asset “sitting” in one bucket at the distance, or lots of visible grids due to camera motion, etc.. and the renderer couldn’t handle it.

@maxliani @iquilezles @wookash_podcast Front to back rendering order + depth buffer rejection can help, but isn't perfect

@iquilezles @wookash_podcast For those not familiar with REYES rendering architecture: start with simpler geometries and tessellate/displace all the finer details. But eventually all objects in the distance will be over modeled, requiring hand-modeled discrete LODs or the renderer will fail.

@iquilezles @wookash_podcast Perhaps the difference that Nanite was designed to simplify to pixel-size triangles, while REYES was to tessellate up to pixel-size, a simpler problem to solve. With dense input geometry, REYES had the same overdraw problem of rasterization.

Also, because the scene graph is a DAG, not a tree, one can add parents by holding Ctrl during drag & drop, and creating instances that way.

Last weekend I refined the scene tree view in Workbench. I used the drag & drop support in dear imgui to implement DAG reordering and reparenting. Custom behavior and drop region draw to show where the node would be inserted, and display where and why it cannot.

PS: don't look at the viewport :D and I hope the tiny colored lines will show through the compression.

@bosmadev @ssh4net @cursor_ai I was charged $130 of cellular data roaming for a short stopover in Dubai, because my phone operator has agreements with virtually every operators around the world, but those in Dubai. I mean shit happens when you let the guard down 🤷🏻.

Oh boy, I am letting Cursor run automated mutation testing in Workbench... 🍿

@ssh4net I stopped it, selected the content with the marquee selection in the viewport, deleted it. Boom. I got the stack trace printed out, but the condition looks tricky.

@maxliani Why crashed?

100,000 fuzzer iterations

@ssh4net No, it’s just my stuff that run in the app loop

@maxliani Also do you using ImGui test engine for workbench tests? 🧐

@ssh4net @cursor_ai Yes I am aware. That little experiment from today burned down my monthly allowance :D Understandable though. I find extra cool that I can even do these things. So 20 bucks well spent for the learning.

Be careful with @cursor_ai i was charged for about 100$ because accidentally enabled On Demand mode (that I’ll never do, in a right mind). A friend of mine got 400euro bill, because forgot to turn it off. When Open Ai in Codex CLI give insane token limits that even on Plus can be used for a pretty big projects. At 200usd Pro, you need to run multiply big projects in parallel to barely touch of available limits.