- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
You must log in or register to comment.
Big news: My physical AI slop doesn’t do anything. Amazing!
e.g. blind
Why? Is looking at the damn thing before you pay money on manufacturing that hard?
This baffles me about vibe coders too. You’re already saving a lot of time just look at the damn code and see if there’s any glaring mistakes.
Why are we treating AI assistance like it’s all or nothing? Why can’t we just have it help a little and still use our own skills?
Why?
He says why in the next sentence. “see what happens”
Well, I now feel better about using OpenCV to fish in Minecraft…
Thought something this stupid was just for shits and giggles. Then I saw this is LinkedIn and he’s a senior product manager.
Haha the joke is on the rest of us
But is he doing that as a kind of joke to show how awful AI is for that task, while being an actually decent product designer/manager himself? I hope it’s the latter, but I wouldn’t be surprised if he drank the kool-AId
deleted by creator
Do this is ALMOST an actual thing. Seems like this guy tried to use an LLM to do something useful, lol. Meanwhile actual scientists are using actual AI to design actual conputer chips that actually perform better than their human competitors. That being said, the new chips are not “end-to-end” AI, theyre designed by AI and a human does the final touches. Theyre also highly unique, i cant find the article now but their AI designed a chip that seemed to use extra parts, i remember there was a diode and transistor that were literally completely seperated from the traces of the rest of the chip, and yet they were functional pieces and the chip wouldnt work if you removed them.
i remember there was a diode and transistor that were literally completely seperated from the traces of the rest of the chip, and yet they were functional pieces and the chip wouldnt work if you removed them.
If that was true they would be getting the Nobel in physics for discovering some incredible new quantum phenomena, it would be front-page news everywhere. I highly doubt it’s true.
Frustratingly, that article you linked doesn’t actually link to the paper. But it is in Nature Communications. That’s a respectable journal but not that prestigious, and it publishes a lot of over hyped stuff. Not that any journal doesn’t. But if they had really found new physics with AI chip design that would go to Science, Nature, or maybe PRL.
Edit: ah, I found it.
https://www.nature.com/articles/s41467-024-54178-1
Chip design isn’t at all my specialty so take my opinion about this with a grain of salt. But I think it’s notable that
Prior works in nanophotonics have demonstrated the class of inverse methods for specific dielectric-based passive structures through gradient based optimizations such as adjoint method
So, there are already known algorithmic approaches to solving for these. I think it’s also notable that these are for signal transformation and antennae, relatively simple operations.
This seems like a vaguely useful result but I don’t expect it’ll be breaking any new ground any time soon.
Confidently incorrect about how fragile some circuits can be. Simple functionality is a convenient illusion we’ve beaten into various squiggles of metal. Electricity is secretly also a radio and a magnet, and even in wires it can’t know there’s nothing at the end until it gets there. Sometimes things just happen.
They’re fabricating micrometer components with a 90 nm process. That’s pretty well in the classical regime. If they’re seeing substantial tunneling at that scale it would be rather noteworthy to say the least.
It’s not tunneling, it’s capacitance. Even truly disconnected components can accidentally become a transformer.
A single diode a micrometer away from anything else is not suddenly a transformer without which a 400 um2 antenna stops working.
Looks like a vital component of next Tesla car.
A new single AI chip controls the brakes, acceleration, and door locks:
($200/month subscription to keep it activated)
see I get doing this in a kind of research way but then if the V1 is this bad surely this should tell you to quit
Has he tried not wasting his time on stupid ideas?
“I’ll just have it make a Hello World to validate my workflow”
“Oh no”
I fully believe this story.
I can imagine the guy having to print this at the factory just pictures someone with schizophrenia making the board somewhere out there.
That C3 is vital to keep the voices out of the PCB.
“我们现在必须建立圈子。”
ah, see here, the problem is, once again, rust. As usual. Claude probably isn’t trained on PCB like it’s trained on rust, and probably PCB doesn’t have something that validates correctness like rust, so that’s your problem. rust rust rust. this message brought to you by the Rust Defense Force.
PCBs are basically designed using graphic editors. The editors all have their own standards and don’t really import pure text pass the netlist. Simply importing from one to the other is quite painful.
It’s basically trying to build a really really complicated SVG and llms struggle with simple svgs. And that’s only the Gerber, there’s also the whole schematic that needs to be done first from which the Gerber is built from.
PCB design tools are basically made to make our lives easier and avoid mistakes, but it makes it incredibly difficult for any kind of LLM, even before we hit all the intricacies involved in it.
It’s only a matter of time until the PCBs match the usual vibe-code colours that you get in vibe-coded HTML, haha
Luckily, it’s a lot easier to train electronics inspectors than decent coders who can follow the whole system or module design. That should help for a bit but, man, to be the QA person looking at that atrocity…
