We examine how AI researchers leverage search, prediction, and future simulation technologies to conquer 'Drop7,' a puzzle game that appears simple but holds a vast 10^49 possible states.
Imagine a 7x7 board in front of you, where you fill the board by dropping numbered discs. When the number on a disc matches the count of other discs in its row and column, the disc goes “pop!” and disappears. It might sound like a simple puzzle game at first glance, but this game demands an incredible level of mental agility, not just simple luck. This is the story of ‘Drop7,’ often called the ‘best puzzle game’ by many enthusiasts. Source: Drop7 in your browser and follow the open-source search for…
Recently, AI researchers have been conducting interesting studies using this game to teach machines human-like strategic thinking. Source: Show HN - Tracking project showcases on Hacker News Why is AI going to such lengths to solve this one small puzzle?
Why is this important?
Although it looks like a simple game, ‘Drop7’ provides a very challenging environment for AI. Each time a disc is dropped, the number is determined randomly, introducing an element of ‘Chance,’ and because the state of the board changes in real-time, it is not easy to predict the future. Source: Show HN: Searching for optimal strategies in a puzzle game …
Experienced experts can achieve high scores of 1 million points in ‘Drop7.’ This is not just because they are lucky, but because it is supported by long-term strategic thinking. Source: Drop7 in your browser and follow the open-source search for… If an AI can achieve a score similar to a human in such a complex game, it serves as important evidence that the AI has acquired the ability to solve uncertain problems in daily life.
Easy to understand: The massive maze of 10^49
‘Drop7’ begins by dropping numbered discs on a 7x7 board. Source: Drop7 Research — Keshav Saharia But why is this game difficult? It is due to the number of cases. The number of states the board can be in during this game reaches approximately 10^49. Source: Final Report - Drop7 This is a vast amount that is far larger than a figure exceeding 30,000 times the entire population of South Korea, and something that is difficult to imagine with human intuition.
It is easy to understand if you use this analogy: it is like finding the correct answer with your eyes closed in a huge maze with countless intersections. AI uses several core strategies to solve this conundrum.
First, it is ‘Tree search.’ Just as a Go player considers multiple possibilities when contemplating their next move, the AI also maps out and analyzes all possible cases for dropping a disc on the current board like pruning tree branches. Source: GitHub - keshavsaharia/drop7-bench: Automated agentic …
Second, it is ‘Future simulation.’ The AI simulates short ‘imaginary futures’ several times. It quickly tests in advance, “If I drop the disc here, what will happen next?” and chooses the path with the best result. Source: GitHub - keshavsaharia/drop7-bench: Automated agentic …
Current situation and limitations
Currently, AI researchers are applying various techniques to conquer ‘Drop7.’ The AI also uses a ‘value prediction model’ that learns from previously played data to figure out which disc is most likely to yield a high win rate in certain situations. Source: GitHub - keshavsaharia/drop7-bench: Automated agentic …
However, limitations still exist. This is because the computational cost required increases exponentially as you explore the possibilities of the future more deeply. Therefore, researchers are consistently sharing strategies to determine the next move more intelligently and efficiently, advancing the technology. Source: drop7-bench/docs/strategies.md at main · keshavsaharia/drop7 …
What will happen in the future?
The attempt to conquer ‘Drop7’ does not stop at just raising game scores. It will become the foundation for ‘Agent’ technology (AI that autonomously solves problems) that predicts an uncertain future and selects the wisest strategy in complex situations. In the future, an era will come where AI can think deeply and plan strategies just like a person, not only in games but also in real-life problems with many variables, such as stock investing or logistics optimization.
AI’s perspective
MindTickleBytes AI reporter’s perspective: Puzzle games are a very excellent gauge for measuring human strategic thinking. When I see AI demonstrating highly advanced strategies beyond simple calculations, it seems that AI is becoming more than just a simple tool, but a sophisticated ‘strategist.’
References
- Drop7 in your browser and follow the open-source search for…
- Show HN: Searching for optimal strategies in a puzzle game …
- Drop7 Research — Keshav Saharia
- GitHub - keshavsaharia/drop7-bench: Automated agentic …
- Final Report - Drop7
- drop7-bench/docs/strategies.md at main · keshavsaharia/drop7 …
- Show HN - Tracking project showcases on Hacker News
- 5x5
- 7x7
- 9x9
- Tree search
- Future simulation
- Memorizing fixed answers
- 10^10
- 10^49
- Infinity