What my dad taught me about AI coding in the 90s

Summer of the late 1990s. Young child, approximately eight years old. Dad, who had previously taught the rules of chess, initiated a unique game. He sat at the table, facing away from the board. As I maneuvered pieces, I verbally relayed my moves. Dad provided his responses, which I replicated on the board. One instance, I inaccurately described a piece's position.

Instantly, Dad corrected me, and I quickly lost the game. However, this experience captivated me. Skilled players don't visualize a board as a photo. Instead, they manage relationships: positions of crucial pieces, attacked or defended squares, open lines, diagonals, pawn structures, tactical relationships like pins, forks, and loose pieces.

These relationships shift as each move occurs. Remarkably, blindfold players often rely on relationships and legal possibilities rather than the literal board. Their chess knowledge automatically reconstructs the position. This skill hinges on spatial working memory, pattern recognition, disciplined calculation, and attention control—not merely photographic memory.

After chess, I pursued computers and programming. With the surge of LLMs in programming, everything is evolving rapidly for software engineers. Everyone is grappling with how to utilize AI effectively. The consensus is that if AI is used sparingly and results aren't diligently reviewed, "vibecoding" occurs, leading to unstable software.

How much coding should AI handle, and how much should humans control and comprehend? This parallels higher-level programming languages. Before, programmers needed deep knowledge of hardware and CPU execution (memory, pointers, threads). Modern AI eliminates the need for such understanding. However, seasoned programmers using AI may still outperform non-technical users.

AI can take control, yet human laziness and difficulty staying focused when AI handles everything remain challenges. Programming with AI resembles blindfold chess: less attention required, less need to remember key elements, less focus on tactical relationships. Yet, the skills defining a strong blindfold chess player align with those of a programmer adept at AI-assisted coding.

For instance, maintaining a deep mental model of code weaving together and simplifying/improving existing implementations through deep thought. It's challenging for me to construct such mental models of software I haven't physically written or read extensively. Balancing detailed discussions and high-level thinking is difficult.

As LLMs and tools advance, those who master AI will move faster and smarter than ever, surpassing the speed of writing/reading code. Unlike chess, where ranking skills is straightforward, measuring programming prowess with AI assistance will prove challenging, as AI can address any quantifiable metric.