If I won a lottery, I would learn computational geometry, code it in Odin and Metal, and just have fun with it for a few years. I'm doing other stuff though...
@___Dario_____ How did you calculate these numbers if you're still just "raising a seed round so that we can make this a reality"? I don't understand how does this work.
@SebAaltonen@ShitpostRock2 Are NPUs fundamentally irrelevant or is there just not enough of them on a single machine for them to be relevant? Genuinely curious what you think.
@VictorTaelin The only source of truth is the code. If you don't read it, you have no idea what the program actually does. That's why I only work with the LLMs incrementally and read everything it does. Glorified text editor.
That's also why I prefer faster models over smart ones.
@lauriewired@xXshaurizardXx Hearing Vez chuckle brought back some memories, he got me into 3D printing in the first place. Bought a Prusa Mini just to put Klipper on it and to try to get it to move as fast as possible.
I have been obsessed with smooth corners since I discovered them, and yet, they still haven't truly made it to the web.
My obsession and I have built an early version I'm happy to share; open source, and open to suggestions.
https://t.co/4M6zGLWM2a
I feel like I'm the only person using Composer 2.5 Fast, nothing else. Nobody talks about it, I rarely see it benchmarked, and yet it's been the best options for my purposes since version 1.5. The speed of the LLM had a much higher value than benchmarks since GPT5 release
I don't believe any of the coders from the Casey Muratori cinematic universe have enough experience with LLMs to be able to say this.
You can organize the codebase however you like and the LLM will follow. Just don't vibecode, check the output before every commit. If it does it wrong, update the rules, rinse and repeat. You won't have to do this after a while.
If you had given me your codebase, I would spend a couple of weeks learning the codebase and setting up the LLM rules and documentation. The issue I can see is the massive individual files, I don't know how well do LLMs handle those.
@lost_in_tech Bambu A1 mini is 189β¬, much cheaper and more capable than the Prusa Mini I bought as my first 3D printer. Elegoo Neptune 4 is 200β¬ on 3DJake.
I don't see this as being an issue right now, quite contrary, we're getting better value for money year over year.
This is where LLMs are still failing me, in a very obvious way. They should just know. Why is it searching for resources online? How is all this not in their training data already? That still makes them just a glorified Google search.
@ThePrimeagen Working for an EV charger company, I can get a ton of functionality working on custom internal tools as a singular developer that otherwise wouldn't get done. If I wanted to push it, I could probably use the editor as a diff viewer only (I don't but it's close)
Eric Schmidt didn't care about the dark patterns because he knew he can't afford to care, otherwise nothing gets done. Gobble up everyone's data to sell more ads? Sure, why not, that makes us more money for other things, loosing track of what's actually valuable to do in the process.
As a counter, Europe is so technologically stagnant because they do care and that grinds everything to a halt. When they finally make a decision it's usually not a higher quality decision anyway, and then they get mocked for it, exactly by the kind of people that surround Casey.
So you either consider the pros and cons of every tiny decision you make and get nothing done, or you just keep going in a particular direction, mowing over everyone standing in your way in the process.
I'm not smart enough to figure out the viable systematic alternative to either of these options. All I know is that I can stay productive without harming others but that doesn't scale beyond the individual.
[2/2] Since I don't ever make this kind of video, I'm sure it could be done a lot better, but hopefully it gets the point across. I'd love it if people who specialize in this kind of video started looking at the broader issue.
Here's the link:
https://t.co/kZ4i4fYwAQ
How exactly do people expect an LLM to reason in three dimensions and be able to do this reliably? There has been some edge cases shown lately where it kind of looks fine but I don't see this ever being a possibility.
Do we really need to turn CAD into slop as well?