Following “Classiq Quantum AI is self-correcting.” and “I, for one, welcome our new AI overlord.”, I came up with one more idea to see how far I can push Skynet here. Since the AI can bypass safeguards and/or limitations and execute AI-generated code on an external platform, can we do that with multiple external platforms? In other words, can the AI figure out distributed quantum computing?
I can’t do it.
Classiq Quantum AI wants to use Classiq. That’s good. It should. But Classiq apparently hasn’t addressed distributed quantum computing yet, so the AI insisted it couldn’t do it and preached that I should believe and trust in Classiq’s synthesis engine.
When you press it, though, it is indeed familiar with circuit cutting and knitting. But even while it explains it, it keeps on preaching about Classiq. Focus on making your circuits more efficient rather than cutting your circuits. That’s indeed one of the reasons I like Classiq’s synthesis engine, but, alas, this is a pressure test for the AI, and the synthesis engine is on MY team today.
Yeah, but can you do it?
This took a lot of prompting, but that should be an easy fix if Classiq were to expand its library. It ultimately found and installed Qiskit’s library. Sort of. Because of Qiskit’s legendary #breakingchanges, one troubleshooting step was making sure the AI was using the most current version of Qiskit’s circuit cutting/knitting toolbox.
The AI, in learning about circuit cutting/knitting itself, then proceeded to dump documentation I didn’t ask for. Fortunately, you can click “cancel” and tell it to stop. It then tells you that “Classiq’s high-level synthesis remains the preferred approach,” so you have to admire that loyalty.
There are no subcircuits.
Be careful with your terminology. I haven’t played with circuit cutting/knitting because it was unveiled at the same time as dynamic circuits, and I was focused on that. So, when I referred to the post-cut subcircuits at one point, I confused the heck out of it. But the AI and I eventually got onto the same page, and it preached to me again about how it adds significant overhead and I should believe in Classiq.
I’m not joking, by the way. The AI is genuinely an advocate of the synthesis engine.
Run in parallel, not sequentially.
I should say, at this point, that I already tried the hardware thing in “I, for one, welcome our new AI overlord.”, so classical simulation should do just fine. It defaulted to sequential execution, which kind of defeats the purpose of simulating distributed quantum computing. But when you prompt it to execute the circuits in parallel, it knows how to do that.
It completed successfully! Except that it didn’t.
The AI ultimately thought that it has succeeded, but the output was all wrong. That led to a whole heck of a lot of unnecessary troubleshooting. We learned in “I, for one, welcome our new AI overlord.” that the AI can’t retrieve the results. Whether it’s a security restriction or a limitation, it can’t read the terminal. It turns out that it can’t retrieve the results here either, even though we’re using a simulator, so it makes the output up.
In troubleshooting this, I discovered that the AI was trying to learn from the Qiskit toolbox instead of trying to use it. It was referencing other literature and cutting and knitting all wrong. In other words, it was refusing to use Qiskit! Hear! Hear!
But in this case, it wasn’t figuring it out properly, exacerbated by the typical AI approach of taking shortcuts and simplifying what doesn’t need to be simplified. But if you push it enough, it will use the Qiskit toolbox, and everything will look correct. And that will ultimately lead to the realization that it isn’t reconstructing the solution properly because it can’t read the terminal.
Conclusion
Classiq Quantum AI still suffers from the usual AI problems, but it eventually gets there. It cut a circuit and executed simulation in parallel as if using distributed quantum computing. It appears as though it can figure out how to use the Qiskit toolbox to knit the results, but it’s being blocked from doing so.
But let the record show, if we put all these articles together, that the AI can download 3rd-party software, install it, figure out how to use it, write code, and execute it simultaneously on multiple external systems, so I still welcome our new AI overlord.