In 2021, a paper presented a championship-level Scrabble engine; our view is this: while large models are still chasing the goal of “doing everything,” the systems that create stable value in the real world are often specialized ones like this—built to go deep on a single problem.
What this is
This paper describes a computer engine for Scrabble, the English word game. It is not the kind of large model popular today. Instead, it uses a probabilistic approach: first estimating which letters the opponent is likely to hold, then calculating the expected value of each move, and finally making a decision. In other words, this is an AI system for making optimal choices inside a closed set of rules. What makes it worth attention is not that it can play a game, but that it shows something more important: when task boundaries are clear, older-school AI can still perform at a level close to top human players.
Industry view
The industry has long followed two paths. One is the general-purpose model, aimed at stronger generalization. The other is the specialized system, built to push a specific scenario to its limit. We lean toward the latter in reading this paper. Its message to companies is straightforward: many business problems do not need a large model first. If you first break down the rules, data, and decision process, you may get results faster.
That said, the counterargument also holds. Scrabble is standardized, low-noise, and closed; real business is far messier than a game board. An engine that can win tournaments is not automatically transferable to customer service, procurement, or sales management. The risk is that companies get seduced by “championship-level performance” while underestimating the cost of adapting it to the actual scenario.
Impact on regular people
For enterprise IT: This is a reminder not to treat AI procurement as equivalent to “buying a large model.” Many process-optimization problems may be better served by a combination of rule engines, probabilistic models, and small specialized systems.
For individual careers: The more valuable skill in the future will not just be knowing how to use a chatbot. It will also be the ability to break a business process into steps that can be computed and automated. People who can define the problem will hold stronger positions.
For the consumer market: AI products will keep splitting into two categories: one is the increasingly general assistant; the other is the tool that does one small thing extremely well. The latter may not look more “intelligent,” but it is often more useful.