It is possible for AI to hallucinate elements that don't work, at least for now. This requires some level of human oversight.
So, the same as LLMs and they got lucky.
this post was submitted on 07 Jan 2025
108 points (82.5% liked)
Technology
60560 readers
3558 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 2 years ago
MODERATORS
It's like putting a million monkeys in a writers' room, but super charged on meth and consuming insane resources.
load more comments
(10 replies)
Tim Harford mentioned this in his 2016 book “Messy”.
They just wanna call it AI and make it sound like some mysterious intelligence we can’t comprehend.
It sorta is.
A key way that human intelligence works is to break a problem down into smaller components that can be solved individually. This is in part due to the limited computational ability of the human brain; there's not enough there to tackle the complete problem.
However, there's no particular reason AI would need to be limited that way, and it often isn't. Expert Go players see this in AI for that game. The AI tends to make all sorts of moves early on that don't seem to be following the usual logic, and it's because it's laid out the complete game in its "head" and going directly for the goal. Go is basically impossible for humans to win against the best AIs at this point.
This is a different kind of intelligence than we're used to, but there's no reason to discount it as invalid.
See the paper Understanding Human Intelligence through Human Limitations
Except we can't build what we can't comprehend that also works.
The problem here is that people with power to direct funds are, more often than not, utterly ignorant in building anything.
I think where all this is generally directed is a society, like in Asimov's Foundation or Plato's Republic (with additional step), where people competent in building something are reduced to a small caste, most of them with local, not professional, competencies, like priests, and with a techno-religion centered on that "AI". This is a hierarchical structure very vulnerable to, well, that kind of powerful people.
The majority will work non-essential jobs (like in Heinlein's Door Into Summer), which do not give them any kind of power, the soldier caste will work the military, and the builder caste will work the technology, and the philosopher caste will be those powerful people. The difference with Plato is in having that first group of people which does not fit into any main caste. By Plato they would all be builder (worker) caste, but that would create a problem with the attempt to make it a religion and a hierarchical monopolized structure. The builder caste should be small.
You might see a whole lot of problems with that idea (which still seems to be attempted), that's because the people from whom it comes don't understand how civilization works and that instruments change the rules constantly, not just to the point they can understand.
Recommend reading: Jodorowsky’s Technopriests
This isn’t exactly new. I heard a few years ago about a situation where the ai had these wires on the chip that should not do anything as they didn’t go anywhere , but if they removed it the chip stopped working correctly.
That was a different technique, using simulated evolution in an FPGA.
An algorithm would create a series of random circuit designs, program the FPGA with them, then evaluate how well each one accomplished a task. It would then take the best design, create a series of random variations on it, and select the best one. Rinse and repeat until the circuit is really good at performing the task.
I think this is what I am thinking of. Kind of a predecessor of modern machine learning.
It is a form of machine learning
Which is just stochastic optimisation.
Which yes is exactly what evolution does, big picture. Small picture the genome evolves a bit more intelligently, using not random generation and filtering but an algorithm employing randomness to generate, and then the usual survival filter because doing it that way is, well, fitter. Also what you can see under a microscope.
I don’t know about AI involvement but this story in general is very very old.
I thought of this as well. In fact, as a bit of fun I added a switch to a rack at our lab in a similar way with the same labels. This one though does nothing, but people did push the "turbo" button on old pc boxes despite how often those buttons weren't connected.
My turbo button was connected to an LED but that was it
Some weren't connected? For most PCs that had it, it was a real thing, though counterintuitive and marketing-speak, because enabling "turbo" was just normal speed and disabling would run in a slower mode for compatibility.
After the 486, there were pentiums built at shops that still used 486 cases. In my experience the button wasn't plugged in.
I remember that as well.
Edit; moved comment to correct reply.
Sounds like RF reflection used like a data capacitor or something.
The particular example was getting clock-like behavior without a clock. It had an incomplete circuit that used RF reflection or something very similar to simulate a clock. Of course, removing this dead-end circuit broke the design.
Yeah, that probably sounds so unintuitive and weird to anyone who has never worked with RF.
Yeah, I've stumbled upon that one a while back too, probably. Was it also the one where the initial designs would refuse to work outside the room temperature 'til the ai was asked to take temps into account?
Flashback to the 1960s, Magic and More Magic
I remember this too, it was years and years ago (I almost want to say 2010-2015). Can't find anything searching for it
You helped me narrow it down. I expect Adrian Thompson's research from the 90s, referenced in this Wikipedia article is what you're thinking of.
Yes! Exactly this thank you
For example, one group of gates has no logical connection to the rest of the circuit, yet is crucial to its function
(I should have gone with my gut, I knew it was ages ago. 30ish years by the sound of it!)
Perhaps you're an AI who only hallucinated a circuit design.
:)
It's been found. Adrian Thompson's research from almost 30 years ago..
So the wires did something
"We are coming up with structures that are complex and look randomly shaped, and when connected with circuits, they create previously unachievable performance. Humans cannot really understand them, but they can work better."
Great, so we will eventually have black box chips running black box algorithms for corporations where every aspect of the tech is proprietary and hidden from view with zero significant oversight by actual people...
The true cyber-dystopia.
This has been going on in chess for a while as well. Computer can detect patterns that human cannot because it has a better memory and knowledge base.
load more comments
(10 replies)
See? I want this kind of AI. Not a word dreaming algorithm that spews misinformation
Read the article, it's still 'dreaming' and spewing garbage, it's just that in some iterations it's gotten lucky. "Human oversight needed" they say. The AI has no idea what it's doing.
Yeah I got that. But I still prefer "AI doing science under a scientist's supervision" over "average Joe can now make a deepfake and publish it for millions to see and believe"
I wonder how well it could work to use AI in developing an algorithm to generate chip designs. My annoyance with all of this stuff is how much people say, "Look! AI invented something new! It only took a few hours and 100x the resources!"
AI is mainly the capitalist dream of a drinking bird toy keeping a nuclear reactor online and paying a layman slave wages to make sure the bird does its job (obligatory "Simpsons did it").
load more comments
(1 replies)
This is what most all ai is. Gpt models are a tiny subsect.
Subset
You are correct but I like subsect better.
I like the subtlety of it tbh.
I want AI that takes a foreign language movie, and augments their face and mouth so it looks like they are speaking my language, and also changes their voice (not a voice over) to be in my language.
You want AI that makes chips that run AI faster and better?
You've fallen into its trap!
Idk, kinda the same, but instead of misinformation we get ICs that release a cloud of smoke in a shape of a cat when presented with specific pattern of inputs (or smth equally batshit crazy)
They are all of the same breed and it's an ongoing field of study. The megacorps have soiled the use of them but they are still extremely strong support tools for some things, like detecting cancer on xrays and stuff
view more: next ›