Luca

I just published my first article on @handmade_net on how to write better (game) libraries. Thanks to @cmuratori and everyone else for the feedback. Check out the article here: handmade.network/wiki/7138-how_…

@SebAaltonen That's kinda crazy that you managed to run out of quota on a gpt pro sub. I share one with a friend and we havent managed despite using it daily for all our work.

I am paying $200/month for the ChatGPT Pro subscription. I have had weeks when it runs out of tokens, but when I switch to personal account. Never had to pay tokens yet for LLMs. Video and mesh generators are quite expensive to run. But you can save meshes to the disk and use them again :D

I would be worried if a dev in my team consumed $250,000 worth of tokens yearly. Of course Jensen would love that every developer was willing to pay that much for tokens processed on Nvidia hardware. Also: A RTX Pro 6000 96GB ($8000) or a M5 Max 128GB ($5000) runs LLMs locally. Open source models today are roughly half a year behind commercial ones, but they have been catching up rapidly. If you really want to run the biggest models, a H200 256GB costs around $30k. Put couple of them in your local server room. That's much less than $250k/year per dev.

@Rasmic The ben davis mind virus is spreading

man im really fighting to not record a whole video on svelte

@TheGingerBill More companies should do this

@wookash_podcast @TylerGlaiel Omg yes please

Who would be interested in Mewgenics source code review with @TylerGlaiel? :)

@SebAaltonen Yeah I don't think it's a one shot for games like it is for game dev, but it is still a meaningful boost depending on the kind of game you're making. Im making some games in flutter atm and there it enables me to iterate super super fast.

I have been testing a lot of AI app/gamedev tools lately, and frankly none of them has good enough quality to have any impact. They are focusing way too much on one-shotting to impress investors. But real game-dev is an iterative process. If your AI platform can't beat Roblox UGC games in quality, then your platform will fail. Who's going to spend time playing shit like that if there's better alternatives around? More is not better. Quality is what matters in competitive market. Better quality than Roblox games is the minimal bar. And it can't be 5% better. Nobody will notice that. Needs to be clearly better quality to lure people jump the ship.

@wookash_podcast Otoh Jonathan Blow's upcoming game is a great example of a project where LLMs could not have helped much, the complexity and experimentation required to develop it is incredible. But maybe they could've made his editor UI prettier if he could get them to work in his engine.

@wookash_podcast They are also great search engines, tons of expert programmers use them for that too. I think people like @mitchellh, @SebAaltonen and @theo showcase a lot of great usage of LLMs for development across a variety of domains. I find them very transformative too personally.

So as far as I understand AI close-my-eyes-and-dont-look-at-the-code landscape looks like this: > it's super duper awesome, and Claudes and GPTs are phenomenal, AGI achieved > ship 10k loc a day, every day, even today! > but of course nobody would do that to process payments, way unsecure, for that ideally you would have a careful engineer implementing, and making sure it's ok > auth is also risky, probably shouldn't vibe code that part > if autoscaling is messed up, we will be charged for tons of $$$ so ideally, config should be done manually, this and cloud infrastructure in general, those terraforms baby > core logic of the service, should be as smooth as possible, after all it's *the* product we ship, so we want to have the knowledge, understand tradeoffs, and not outsource it > of course "make it secure" doesn't work 100% time so ideally here, we would do a proper review of the security to make sure it's done correct So what's left? Tests? Apart from twitter demos, of things that have been already done, which are recreated, with the AI trained on the programs it's recreating?

@convex Finally!

ok we finally did it 🇪🇺

@bcherny Crazy, I was just asking for this x.com/SasLuca/status…

A small quality of life improvement. Hope you find it useful!

@opencode Best approximations I can think of now: 1. undo the chat but not the code 2. tell the llm you had to make some fixes and describe them briefly 3. continue with task 2 You could also use separate agents for the task with an md for context sharing between agents I suppose.

Does any AI agent tool like @opencode have a way to do a "partial compact"? Agent does task 1 from a todolist, I test it and it's buggy so we talk to fix it. I would like to continue task 2 in the same convo and compact just the portion of the chat about fixing the task 1 code.

@vkrajacic Incredible, file pilot manages to embed a keylogger and full blown crypto miner all in just 2mb with no internet connection, and still feels way faster than file explorer! What more could you want from your software really.

I've been exposed. The real reason I spent 4 years crunching, burning my savings, and staying away from family was to eventually ship a closed source file explorer… so I could secretly steal your data. I even made it standalone with no internet access to avoid suspicion. The internet never disappoints. reddit.com/r/Windows11/co…

@valigo char *c is what the creators intended but char* c is what I prefer because chopping your type in half and mixing it with the variable names seems crazy to me. In an ideal world we'd all write types like in Jai or zig `c: *u8`

It's crazy how many people, including C programmers, don't know where does the star go. Give your option, solution is in reply

@ainexdev @raysan5 On android you need cmake for native code, as gradle calls into it and integrates into android studio projects. On ios cmake just has xcode support, so you get an xcode project generated from it with everything setup properly, which seems rpeferable to messing with xcode.

@SasLuca @raysan5 Actually, I see that as best practice. Platform-native setups (Gradle/Java on Android, Xcode on iOS) make debugging, profiling, and long-term maintenance much healthier than forcing everything through CMake or other “one-size-fits-all” hacks.

Hope raylib gets enough sponsors to buy a MacBook someday 😂 iOS is basically the only platform raylib still can’t cover.

@ainexdev @raysan5 I know some big game studios manage to integrate android and ios in a visual studio workflow, which seems cool but I never discovered how to do that. I tried the vcremote xcode integration and it just didn't work. And the official android templates in vs are very limited.

@ainexdev @raysan5 For example with SDL3 for android, you need a whole bunch of java files, and gradle likely has to get in the mix, and if you wanna debug and profile native code well, idk, without android studio that's gonna suck. With iOS theres also a lot of xcode specific stuff you might want

@ainexdev @raysan5 So, from my brief experiments, all the gles code in raylib should largely still work on ios even tho the drivers arent updated anymore. In Skia for example I did hit driver bugs, and ANGLE fixed those since it reimplements gles in metal.

@SasLuca @raysan5 For rendering, Metal support would be ideal long-term. But GLES is fine to start with — features first, optimization later.

@ainexdev @raysan5 I got raylib working on iOS years ago. I've been debating recently if I should try and contribute a port but there are a lot of questions about how that would work. Do we provide an xcode template or have people generate it from cmake? Do we bundle ANGLE or use gles straight? etc

@raysan5 Hope raylib can find an iOS expert to join in 2026 and push iOS support forward.