Shakespeare makes you a better engineer.

I made a benchmark for "discernment" several months ago — how well can a model judge. It has been the most accurate predictor of the utility of a model in agents and coding that I have seen to date.

It turns out that discernment is incredibly important when writing software, and models with poor discernment may be good at the syntax but they are terrible at the broader task of engineering.

See, the role of software engineer, when done properly, is not primarily about pounding out syntax. Being good at typing and syntax is just a prerequisite to do the real job: pathfinding in an infinite problem space. The primary skill that all good senior engineers have is an above average ability to make subjective judgment calls at each step of the problem solving process. Given two or three iterations, a good senior engineer's solution approaches the optimal path. Not that dissimilar from a great race car driver finding the optimal racing lines on a given track.

To make coding models faster, more token efficient, and cheaper to operate — they often are trained, and reinforced, with code, code and more code — and if we were training a race car driver, this would work great. However, unlike the racetrack, the solution space of coding is infinitely vast — discerning a path through it efficiently leverages everything we know, everything we are.

In other words: I can't explain why organized sports, reciting Shakespeare, memorizing scripture, or playing the piano makes you a better engineer, but it does. Along with brute experience, it is those seemingly unrelated disciplines that shape and craft your ability to make judgment calls. To discern.

The same is true for models. Stripped down and efficient models can be excellent at various discrete tasks, but finding an elegant solution to a hard problem? Nah, turns out you need to know how to change a diaper at 3am to do that well.