You have probably seen some of the things that you can do with OpenAI’s new ChatGPT which was released last week: this unnervingly human-like language model can invent a fictional language, emulate a virtual machine, and debug code for you.

Yesterday I finally spent some time playing around with ChatGPT and I am really impressed by how powerful it is. I decided to try to teach ChatGPT to implement a new computational model from the ground up. The model, based on “mewbits” is equivalent to quantum computing and qubits, but the catch is that I didn’t tell ChatGPT what I was having it do; I just told it the rules of the game for this abstract computational model. Over the course of this transcript, the following happens:

  • ChatGPT is introduced to a new computational model. The basic unit of information of this model is a “mewbit”, which can be in any normalized linear combination of two states: 😾 and 😺.

  • ChatGPT is introduced to how measurement (state collapse) behave for mewbits. I then introduce some basic operations you can do on mewbits, including not (Pauli-X) and H (Hadamard). It understands what these operations do and successfully completes several exercises I give to it.

  • ChatGPT writes a mewbit class in Python, which tracks the internal state of the mewbit and has methods for operations and measurements you can do on a mewbit. It adds comments explaining what is going on.

  • I explain the basics of quantum entanglement to ChatGPT. It constructs a program to emulate a system of multiple mewbits and understands how to generalize single-mewbit operations to this multi-mewbit system.

  • I introduce the concept of a two-mewbit CNOT gate and I explain in plain english how to simulate joint measurements. ChatGPT understands, completes exercises successfully, and updates its Python class to incorporate this complex gate.

  • I ask ChatGPT to write a script which would generate 100 entangled Bell pairs and measure the system states, then plot the results.

  • Finally, I asked ChatGPT if any of this sounded familiar, and it understood that the “mewbits” I described to it were actually qubits!

Teaching ChatGPT about mewbits

How does measurement work on mewbits?

Teaching ChatGPT about quantum gates

At this point, ChatGPT makes the inference that 😾 = 1*😾+0*😺:

Now we introduce a more complicated concept: the Hadamard gate:

ChatGPT formalizes what it has learned into Python code

At this point ChatGPT clearly understands the concepts and answers examples correctly, but I would like to see what it thinks is actually going on under the hood, so I ask it to formalize what it has learned into Python code:

And because all good code should have comments:

Since it seems ChatGPT forgot to add measurement as a thing you could do to a mewbit, I asked it to add that in as well:

Multi-mewbit systems

At this point I made the considerable jump in complexity to explaining how multi-mewbit systems behave, explaining a simplified version of the mechanics of quantum entanglement to ChatGPT. I also explained how you can perform operations on a single mewbit in the system, changing the state of that mewbit but leaving the rest unchanged. Notably, ChatGPT inferred that this updates the state of the multi-mewbit system:

Multi-mewbit gates

This was a complex idea to explain to an AI, and I admit I had to make a few attempts at explaining it properly before ChatGPT was able to understand what I meant. However, as you’ll see, ChatGPT makes an error and then corrects itself:

I now ask it to incorporate its new knowledge into its existing understanding of mewbit systems:

Joint measurement of multiple mewbits

Putting it all together: Bell state preparation

I’d also like for ChatGPT to plot the results:

Unfortunately, ChatGPT doesn’t allow me to run the code and we have some banter about arbitrary code execution:

Does ChatGPT understand what I just taught it about?

At this point I have described in plain English the basic principles of qubits, quantum systems, gates, entanglement, and measurement. ChatGPT has constructed code which would accurately simulate this, but does it understand what I just instructed it to build?

image-29