OS scrapes through the test, scoring deceptively respectable five out of nine.
This isn’t much of an own. They vibe coded an OS. The fact that it works at all is enough. Days are numbered.
Submitted 8 hours ago by BrikoX@lemmy.zip to technology@lemmy.zip
OS scrapes through the test, scoring deceptively respectable five out of nine.
This isn’t much of an own. They vibe coded an OS. The fact that it works at all is enough. Days are numbered.
For minor personal computing, “eh it partially works OK” is extremely annoying, but maybe worth it if it’s dirt cheap. For anything more complex or important, that level of unreliability is an enormous net negative no matter how low the cost.
We have excellent shit that already cists nothing.
If they let it “improve” itself it will be like Windows ME that had to be re-installed every 6 months after it optimized itself to death.
Screenshot showing half an application window destroyed by entirely transparent areas
Windows 12.1 screenshot leaked?
www.youtube.com/watch?v=m3K7wUTX-SY
Meanwhile actual coders managed to make Doom playable in an Excel spreadsheet.
Public articles like this will likely be used to train future models, and end up crushing their self-confidence.
You do you, little LLM. Here, have a cookie.
The fact this even almost works remains fascinating. Someone got a full-featured bare-metal operating system to program itself, and it’s halfway between early ReactOS and that Flash game for Windows Really Good Edition.
We forget that computers are for everyone. They’re supposed to be a bicycle for the mind. Needing to rely on existing mature applications, created by teams of experts, is an obstacle we’ve been fighting since BASIC. Anyone should be able to slap together some program that suits their needs. And not just theoretically capable - able. ‘You could learn to do it!’ means you cannot yet do it.
There’s a handful of languages intuitive enough to make people dangerous in matter of days. Until recently, suggesting that English was one of them would be laughable - but now you really can describe what you want and swallow the elephant. It doesn’t work, but it’s alarmingly close, for amateur use of this fiddly tool. The computer literally did your work for you, and people still complain about limitations.
An actual intelligence trained on StackOverflow would look at ‘make me a new OS’ and respond, ‘but why do you want that instead of modifying BSD?’ An intelligence that’s also been trained to avoid backsass would start from structure and complete its own todos. Which some guy claims to have done, for compiling a new language that mostly uses stupid names for things. He’s got the right idea, for weaponizing an LLM into something useful, but that idea is to make it act like diffusion.
I mean if it made that after one prompt that would be amazing.
givesomefucks@lemmy.world 8 hours ago
I hate these “AI is this dumb” posts because most of the time it’s someone telling the AI to build a bad product just to laugh at it.
But they’re still using AI.
And it’s still training the AI.
It’s like burning a tree down in Cali and then saying it was down to raise awareness that Forrest fires are bad…
Everyone knows it already, and you’re just making an existing problem worse.
It’s just clickbait, and I hate that Tom’s post so much about it because it gets clicks.
Aatube@thriv.social 4 hours ago
I mostly agree, but Tom’s/the reporter didn’t vibe the product. Ze reviewed something someone else vibed.
givesomefucks@lemmy.world 4 hours ago
Them promoting it makes more watch/read whatever original.
Just like OP posting this.
ShinkanTrain@lemmy.ml 7 hours ago
Can they pull a Wimp Lo and train it wrong as a joke
Telorand@reddthat.com 4 hours ago
There’s a handful of projects out there that are trying to do exactly this, by programmatically poisoning potential training data.