It seems like I need to optimize the offscreen buffer logic Runs smoothly here on arm mac, Safari and Firefox, but seems to stutter on Chrome Windows….
EDIT: Replaced with optimized version
EDIT: Just for fun, these tracks aren’t played randomly (no chiptunes), but kinda nostalgic:
It’s quite helpful now, but to be clear here, you still need to be a (quite experienced) coder, it will not at all do the whole work for you. But I’ve tried something similar a year ago, and the training data improved a lot. It knows clearly very good approaches for each problem most of the time - of course only, if such a problem was published or they stole its contents, in multiple ways.
What I’ve learned from this:
The AI still will never solve a new or complex combined problem, since it is not smart at all (and I think never will be).
Best is to break down the tasks in little fragments, for example components that can be solved each in their own context.
You need to control the generated code each time and know exactly what it does. The AI will produce a lot of errors, and forget context details all the time. So you need to know how to structure, build and read complex algorithms, how good code looks, know the design patterns, etc.
The AI clearly is a mono tasker. You need to exactly describe what should be changed, and be as precise as possible. Often you need to explain the AI the technical approach exactly, like “use an offscreen buffer, for the font, then generate a sine scroller buffer by copying each vertical pixel to another buffer using a sine function”, etc.
Since the AI is a mono tasker, it will very likely remove other parts in the code, and introduce bugs. For example, if you ask it what could be optimized, it will also remove conceptual code or introduce bugs, since it cannot differentiate here.
The AI does not know what it is generating. It can compare contents from an insane amount of data, obviously, and then decide which might fit best to the asked prompt and the prompt history. Yet it has no concept of the code and simply is trusting its sources / the valuation of the trained data.
It’s very helpful for scaffolding ideas into code.
Using perplexity seems to be a good idea, because it provides multiple advanced models at once. If one model will fail with a task multiple times, you can switch the model. The difference of answers also contains useful information, even if both did it wrong (since I can think like a human being, right).
The AI never produces 100% exact and reliable results, and never will. That’s why it’s a very, very bad idea to give an AI the power to decide about important real world stuff, like “is the human being allowed to get insurance”, or “is it a good idea to drop a bomb here”… Once again, human stupidity remains undefeated.
@Garf I’ve watched quite a lot recent videos about claude vibe coding where it becomes clear that the generated code looks like good code, but actually is using a slow, essy “gaming” like engine under the hood. So it’s like an interpreter/engine which mimics good code, but that interpreter itself is coded in a horrible way, it’s totally crap code. I didn’t try myself, but the PrimeTime channel gives some nice insights often: https://www.youtube.com/@ThePrimeTimeagen
do you know any good chiptunes in .AY format (AY-3-8912 dsp chip)? So I guess from either Amstrad CPC, Spectrum ZX or Atari ST…? Seems to be hard to find these days. Couldn’t find too many in chiptunes.app either…
I remember that there are some really awesome tunes, where specialists tested the barriers of/hacked that chip and created impressive stuff… I think it was really some chiptune, so not for a game, but either for a demo or cracktro….
Runs smoothly here on arm mac, Safari and Firefox, but seems to stutter on Chrome Windows….