My opinion on the current batch of “AI” services based on Large Language Models continues to develop, but I realized today one analogy for how I currently feel.
I’ve begun to feel about LLM-driven AI bots the way I feel about sudoku puzzles.
Most people think of “sudoku” as a 9x9 puzzle with 17 or more given digits arranged so that when a person adds digits to make a finished grid that includes 1-9 in each row, column, and 3x3 box, there is only one possible arrangement. Given that definition, it’s pretty easy to write software that generates such puzzles, and many, many, many people have.
Books and magazines are published filled with computer-generated puzzles, websites generate them on demand or on schedule, and while there is a technical a finite number of possible starting grids (5,472,730,538 without mirrored or rotated grids), the set feels infinite.
While I had discovered a few variants (Killer Sudoku and Sudoku-X) long ago, it was only watching Cracking the Cryptic during the pandemic that I learned there is so much more to sudoku puzzles. Even if one is limited to “classic sudoku” puzzles fitting the definition I outlined above, there is still a vast gulf between a typical computer-generated puzzle and one hand-crafted to lead the solver along a path, finding telegraphed deductions in a certain order the bring joy.
LLM-driven AI chatbots are like computer-generated sudoku puzzles.
They meet the definition of conversational communication, and follow the rules, but they lack the spark and joy of hand-crafted human communication.
I won’t do computer-generated classic sudoku puzzles any more, because once I got into the hand-crafted variety, the generated puzzles seem like a waste of time.
I think reading well-written communication, fiction or non-fiction, provides the same result: once you’ve read a good article or book, the computer-generated stuff just falls flat.
