The killer feature for me is not drawing TikZ visually, but being able to touch old TikZ without turning the source into generated-looking soup.

Exactly, I wanted to avoid that. In contrast, if you open an SVG in (for example) Inkscape and make a minimal change and save, the resulting file has little to do with the original.

The part that would bother me most is not the ban itself, but having no idea what to change. “Policy violation” could mean VPN, billing risk, location, or actual usage.

All online services are like this. I used to sell stuff on eBay. Had multiple accounts randomly closed. Never did find out why.

It sucks, but it comes from the learning that when you do provide specific explanations it’s really just coaching chronic abusers how to get around bans. The collateral damage of legit users who accidentally tripped the detection algorithm is accepted as the less-bad of options.

Not defending it, just sharing perspective from having worked on some fraud/abuse mitigation projects.

Using employees instead of contractors because they produce higher-quality data sounds less clever after you leak the employees' data.

The gift card analogy breaks down a bit: a gift card doesn’t get more expensive for the store to redeem later. Compute can.

Back when I experimented with getting ChatGPT by paying for API access rather than the official subscription, the credit was denominated in money, not in tokens.

Compute does whatever it does, but $5 sitting in an account to be spent on tokens at some arbitrary time in the future, earns money for OpenAI etc. even when they didn't make it expire.

A voxel world is a pretty good sales pitch for APL: the weird-looking part is the notation, not the model.

The part about context discipline feels underrated. Larger context windows don’t remove the need to decide what the model shouldn’t see.

Totally!

"Under-16 social media ban" sounds narrow. In practice it means building an age-checking layer for the whole web, then hoping it only gets used for children.

The headless WebKit on DRM approach is interesting. How heavy is it in practice on the flipper one's soc, whats the idle and active memory footprint? Asking because WebKit's memory behavior on constrained devices tends to be the thing that bites you later, not the CPU

Immediate-mode in pure C is a nice constraint. how does it handle text rendering, do you bring your own atlas or is there something built in? Thats usually the part that balloons the dependency footprint.

The demo uses a simple prebuilt fixed-size font atlas texture and renders the entire UI (including character quads) via batched glDrawElements calls (one draw call per clip-rect).

But the way how text rendering is delegated to the user is quite flexible, microui basically calls these three user-provided functions:

    int r_get_text_height(void)
    int r_get_text_width(const char* text, int len);
    void r_draw_text(const char* text, mu_Vec2 pos, mu_Color color);

...how r_draw_text() is implemented is then entirely up to you.

If anyone wants it here is an old fork of mine that uses SDL3 - https://codeberg.org/krapp/microui_sdl3_demo

it does not, it basically delegate that part to lower level library for which you have to write the glue code for, there is an example for SDL.