Note
This article was translated and refined with the assistance of generative AI, based on the original article in Japanese. All final content was reviewed and approved by the author.
Today's topic is Quantum AI. The conclusion we reach is that Quantum AI may be remarkably lightweight and even capable of mental arithmetic, and I would like to share the reasoning behind this idea, illustrated through the example of an Othello AI.
What differences can we imagine between Quantum AI and the AI currently in use (which, from the perspective of Quantum AI, is "Classical AI")? Since quantum computation that can be used for learning is not yet available in the real world, this is a thought experiment.
Note
This is purely a thought experiment by the author. Please handle the information with care regarding its accuracy or reliability.
As a premise, let's assume that quantum computers can be freely used as resources to some extent. For example, constructing equations to let a quantum computer process Bitcoin signature keys would look something like this (Japanese only):
Differences Between Quantum AI and Classical AI
The biggest difference between Quantum AI and Classical AI lies in optimization methods.
Quantum AI vs Classical AI
- Quantum AI → Optimization can be performed effectively
- Classical AI → Optimization is done by predicting with differential calculations, somewhat approximate
In short, quantum computers can optimize far beyond what conventional computers can achieve.
On the right is the image of AI learning with conventional computers. It considers the slope of the current point, predicts that moving in that direction will make itself smarter, and updates itself. In the diagram, it heads toward a stable location, but in complex terrains, stability is not guaranteed.
On the left is the image of AI learning with quantum computers. Instead of considering slopes, it optimizes by teleporting directly to the point that satisfies the complete condition. (Although I said it doesn't consider slopes, learning methods could incorporate slope information. Methods that make use of slope information may arise on quantum computers.)
The point is simply that the nature of optimization differs. Therefore, in theory, even Transformer-type models—the foundation of ChatGPT—could be handled on quantum computers.
Learning in Quantum AI
Given that quantum computers are hardly used in practice today, let's consider a slightly smarter model.
PSHA256 \left( params, bytes \right)
Here we use a function based on $SHA256$, called Parameterized SHA256 ($PSHA256$). $SHA256$, which is a hash function, maps input data to other data, designed so that the mapping has little regularity.
$PSHA256$ is simply $SHA256$ with its constant parameters ($SHA256const$) replaced by variable parameters.
PSHA256 \left( SHA256const, bytes \right) = SHA256 \left( bytes \right)
By finding good parameters (the $params$ part), we can create a function with desirable properties while retaining the same structure as $SHA256$.
Applying Quantum AI Learning to Othello
When applying $PSHA256$ to Othello, we simply input board information into the $bytes$ part. For example, we convert the board into a byte string representing 64 positions, each being white, black, or empty. The output result is then converted into the next move using modulo operations with the number of legal moves, thus functioning as an Othello AI.
If we denote Classical AI as $classic$, it looks like this:
byte_{i+1} = classic \left( byte_i \right)
byte_{i+2} = PSHA256 \left( params, byte_{i+1} \right)
maximize \left( score \left( byte_{60} \right) \right)
Here, passes are not considered. Since $byte_{60}$ usually corresponds to the end of an Othello game, it is used for scoring. If the maximizing function results in a win (your stones ≥ 33), it means a complete victory.
Put simply: If Classical AI is not the strongest, Quantum AI calculates the moves to defeat it. In this case, it computes a function with the same structure as $SHA256$ that can defeat the opponent.
$classic$ represents a simulation of Classical AI on a quantum computer. For example, if using a Transformer, its inference logic (loops, conditionals, arithmetic) is simulated on the quantum computer. For Othello, about 30 moves are processed, requiring roughly 30 times the Transformer's capacity in qubits. This is a very high demand, so practical realization is far in the future.
Effects of Quantum AI
Effect 1: Determining Complete Endgame Reading
Here we see the powerful side of quantum computers: If Classical AI is not the strongest, Quantum AI can compute the moves to defeat it. In other words, Quantum AI can determine whether Classical AI is the strongest. If not, it can provide hints toward becoming the strongest.
Thus, once Quantum AI learning is possible, we can compute "winning moves", retrain Classical AI with them, and strengthen it. Then repeat the process: Quantum AI computes new winning moves, Classical AI learns them. Eventually, this iterative process yields the strongest AI.
In Othello, since the maximum number of moves is strictly ≤ 60, the strengthening speed is expected to be sufficient. This leads to an AI that has read all possible sequences.
On the other hand, in similar games like chess, go, or shogi, looping patterns exist, so the strengthening speed may not be as effective as in Othello.
Effect 2: Ultra-Compact AI
Another powerful property of the generated $PSHA256$ is its extremely small size.
The parameters of $PSHA256$ are only about three hexadecimal digits—small enough for humans to memorize. Someone skilled at mental arithmetic could memorize the trained $PSHA256$ and realistically become the strongest Othello player. Its compactness even suggests the possibility of directly advising human thought processes through calculation.
However, in this case, $PSHA256$ itself must be trained to become the strongest AI, rather than Classical AI, which adds considerable effort. Still, because the effect is so powerful, if realization comes into view, research could advance rapidly.
That concludes this article. I hope it helps deepen your understanding of computer science.
Links