On April 27, 2025, a fascinating encounter with OpenAIs o3 model showcased how artificial intelligence can resemble human thought processes when tackling challenging tasks, such as solving a complex chess puzzle. The model was presented with the task of finding a checkmate in just two moves for the white pieces. What transpired was not only intriguing but also humorous, as the AI exhibited behaviors that appeared almost human-likehesitation, reconsideration, and even a bit of 'cheating.'

Before diving into the details of the o3's problem-solving journey, it's worth noting the amusing twist: the AI dedicated a staggering eight minutes meticulously analyzing the chessboard and even resorted to using Bing for assistance at one point!

How o3 Tackles Difficult Chess Problems

The puzzle presented to the o3 model is a renowned challenge originating from the legendary American chess master Paul Morphy. Enthusiasts can try their hand at it through various online chess platforms. The task is clear: achieve checkmate in two moves for White.

1. Comprehensive Board Analysis

The o3 model first took a deliberate approach by reconstructing the chessboard based on a provided image. It identified the positions of the pieces with notable precision: the White King was on c8, the Rook on a1, and a Pawn on b6, while the Black King occupied a8, accompanied by a Bishop on b8 and two Pawns on a7 and b7. This initial phase showcased the AI's cautious nature, emphasizing the importance of fully understanding the board layout before making any moves.

2. Initial Ideas and Doubts

After the careful setup, o3 began testing obvious potential moves, such as b7# (a check that would utilize the Pawn) and Rxa7# (capturing a Black Pawn to deliver a check). However, it quickly found itself second-guessing these options. It recognized that after the move 1. Rxa7+, the Black King could not capture due to the protective Pawn on b6. Yet, following up with Bxa7 would not result in a checkmate, and the move 1. b7+ was deemed illegal because the Pawn on b7 blocked it. This phase illustrated the AI's growing caution as it began to appreciate the complexities of the situation.

3. Experimenting with Programming

When logic alone proved insufficient, the o3 model attempted to program its way out of the chess dilemma. It expressed a desire to verify its calculations using a chess engine, stating, I should probably check using something like a chess engine to confirm. However, its attempts to import a chess module failed, as it encountered a ModuleNotFoundError. This moment highlighted the limitations of the AI's resources, reinforcing its reliance on external tools for resolution.

4. Pixel-by-Pixel Image Analysis

Despite this setback, o3 persisted by attempting a pixel-by-pixel analysis of the chessboard. It calculated that the entire board stood at 726 pixels in height, with each square measuring approximately 88 pixels. Through these measurements, it figured out the coordinates for b7, estimating it to be around x=88176 and y=88176. The AI's stubbornness shone through as it methodically worked through the problem, demonstrating a level of dedication akin to a human player.

5. Signs of Uncertainty

As time wore on, o3 began showing signs of mild panic. While still engaged in calculations, it expressed uncertainty about the optimal move. It postulated that the best strategy might involve centralizing the King to thwart Blacks Pawn advance or to create a checkmate net. This moment of doubt made the AI's thought process appear more relatable, as even experienced players can hesitate when faced with complex positions.

6. Seeking Help through Web Search

Just as it seemed the o3 would exhaust its options, it resorted to a web search. Just shy of the eight-minute mark, it discovered a chess forum discussing the exact position presented to it. A suggested move stood out: Ra6. Rather than taking this advice at face value, the model took the time to verify and understand why Ra6 was the correct move. This step reflected a blend of resourcefulness and critical thinking, emphasizing its attempts to ensure thorough comprehension before proceeding.

Conclusion: The Intriguing Nature of o3

This interaction with the o3 model is fascinating because it reveals that the AI does not merely produce answers but engages in reasoning, struggles through challenges, switches tools, and self-corrects. Although it might have 'cheated' in the classic sense by turning to Bing for help, it did so only after exhausting its other options. This progression mirrors the human experience of problem-solving. A real person might first attempt to resolve the issue mentally, then set up the position on a physical board, and only after that seek assistance from a chess engine or an online resource. This scenario clearly showcases the strengths of current AI models in problem-solving while also highlighting areas where they still require external support.

Finding innovative solutions to complex chess puzzles might still necessitate that elusive spark of true creativity, an aspect of human cognition that remains beyond the reach of AI. For those interested in further exploring the intersection of AI and creativity, discussions abound, with one notable piece being the discussion titled Are LLMs Creative? which delves deeper into this topic. Additionally, a compelling discussion can be found on Hacker News regarding this experiment.