> Initially, the AI’s designs seemed outlandish. “The outputs that the thing was giving us were really not comprehensible by people,” Adhikari said. “They were too complicated, and they looked like alien things or AI things. Just nothing that a human being would make, because it had no sense of symmetry, beauty, anything. It was just a mess.”
This description reminds me of NASA’s evolved antennae from a couple of decades ago. It was created by genetic algorithms:
What’s amazing to me is this design looks a hella lot like tree branch formations to me. Makes me wonder if trees have some form of antenna-like functionality we are unaware of.
They used genetic algorithms to evolve digital circuits directly on FPGAs. The resulting design exploited things like electromagnetic interference to end up with a circuit much more efficient than a human could've created.
In my mind this brings some interesting consequences for 'AI apocalypse' theories. If the AI understand everything, even an air gap might not be enough to contain it, since it might be able to repurpose some of its hardware for wireless communication in ways that we can't even imagine.
In practice, we'll just let that AI have a direct internet connection, and also give it enough access to push code straight to prod. For the good measure.
> Five individual logic cells were functionally disconnected from the rest — with no pathways that would allow them to influence the output — yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. Furthermore, the final program did not work reliably when it was loaded onto other FPGAs of the same type.
I don't remember the title, but someone wrote a story were an AI would use the (imperceptible) flickering of a fluorescent lightbulb and a camera to transmit information across such an "air gap".
There's a comic out right now positing that a sufficiently intelligent AI with appropriate access could use imperceptible (to us) vibrations from mechanical computing parts like spinning rust HDD's etc.
It's a throwaway mechanic in the comic, but it seems plausible.
You don't need an AI to come up with remote sensing or air gap traversal capabilities though.
Note for example TEMPEST surveillance, or using a distant laser to pickup speech in a room based on window vibrations. Air-gap traversal is easily done by exploiting human weaknesses (e.g. curiousity to pick up a USB drive to see what's on it), and was successfully done by Stuxnet.
Indeed, there are lots of methods, but i was specifically thinking of the possibility of a method an isolated AI might feasibly figure out with only the tools it has easily available to it.
But as someone said earlier, the real interesting part is when/if they start figuring out novel concepts we as humans haven't even considered.
There was something similar about using evolutionary algorithms to produce the design for a mechanical piece used to link two cables or anchor a bridge’s cable, optimizing for weight and strength.
The design seemed alien and somewhat organic, but I can’t seem to find it now.
But I recommend logseq for anyone new! All the best features, none of the weird energy of roam, where i worry the founder is training ML models on data.
“Typology optimization” is probably what you’re thinking of. All current versions of it result in this similar blobby-spider-web vaguely alien and somewhat organic structures.
Looking at things like bicycles designed this way leaves me suspicious that it doesn’t actually have the power to derive interesting insights about material properties. I suspect future versions may end up starting to look more mechanical as it discovers that, for example, something under tension should be a straight line.
Reminds me a bit of chess engines that crush the best humans with ease but play moves that human players can identify as "engine moves". In chess the environment is fixed by the rules so I'd assume this deeper understanding of underlying patterns is only amplified in more open environments.
The bias is a handicap, the looking for beauty, symmetry, a explanation, a story, its all googles upon googles of warping lenses and funhouse-mirrors, hiding and preventing the perception of truth.
Zero is taught routinely to primary schoolers today, but it has been a hard thing to come with for scholar who struggled to nail as smooth as a concept as we know it now.
The bias toward familiarity is detrimental to edge research, but on the other hand if no one smooth the baseline, most advanced knownledge will remain just that and will never reach their full utility to humans. Finding the proper set of concepts that makes it click can be very complicated. Finding a communicable simple thought framework to let other also enjoy it and leverage on it to go further can be at least as hard.
Maybe not so much the implications. If our science is defined by symmetry, beauty, anything - and it is, because so much of physics is literally about looking for symmetries of various kinds - why are we ignoring the loud hints from ML solutions that this is a limiting heuristic?
> why are we ignoring the loud hints from ML solutions that this is a limiting heuristic?
This comes up a lot and always strikes me as rather anti-science, even anti-rationality in general. To speed run the typical progression of this argument, someone says alchemy and astrology occasionally "work" too if you're determined to ignore the failures. This point is then shot down by a recap about the success of QM despite Einstein's objections, success of the standard model even with lots of quasi-empiricism etc, etc.
Structurally though.. if you want to claim that the universe is fundamentally weird and unknowable, it's very easy to argue this, because you can always ignore the success of past theory and formalisms by saying that "it was nice while it lasted but we've squeezed all the juice out of that and are in a new regime now". Next you challenge your detractors to go ahead and produce a clean beautiful symmetric theory of everything to prove you wrong. That's just rhetoric though, and arguments from model/information/complexity theory etc about fundamental limits on what's computable and decidable and compressible would be much more satisfying and convincing. When does finding a complicated thing that works actually rule out a simpler model that you've missed? https://en.wikipedia.org/wiki/Minimum_description_length#MDL...
Because you can never be sure with the ML stuff. Perhaps it was 1 iteration away from finding a solution that was better and also symmetric. Perhaps it it a great, but not optimal, local maxima.
Lacking symmetry, it's extremely hard to understand how the antenna actually works (i.e. why those six bends, as opposed to any other random six bends).
My best guess is that the edges are oriented such that at the tested frequencies they cause constructive interference inside the antenna therefore boosting the signal. The orientation is weird because that's probably the best way to make it work in all directions, if the edges were in a flat plane, the constructive interference would only work in a single direction.
I mean sure, but how do you figure out what directions and angles to bend it in? I don't know much about signals and radio and stuff, but it feels to me like this could only be achieved through trial and error until the ideal was found, which is what evolutionary arguments are designed for.
I've had bugs go away when I added `print("working fine until here")` to the preceding line. So if someone told me "this line is needed but I don't know why", I wouldn't even blink.
Back in the day the QA department would veto a release candidate of our firewall product due to performance degradation if a network cable in the testing equipment was bent too much during validation runs. Rearrange the cable and next test run would pass.
This description reminds me of NASA’s evolved antennae from a couple of decades ago. It was created by genetic algorithms:
https://en.wikipedia.org/wiki/Evolved_antenna