Under the Hood · Part 5 · 5 min read

Why code got solved first

AI got good at code because code keeps score. Most things don't.

By Aikansh ·

An old wooden scoreboard in a dim hall, one column lit.
PLATE 05 — THE SCOREBOARD
The car: Engine·Memory·Persona·Wheels·Gauges·Driver

Try to learn basketball blindfolded, with no one ever telling you whether the ball went in. You could practice for years and never improve. Now put a small beep on every shot that scores. Suddenly you get better fast. That beep — not talent — is most of the story behind why AI can write code but cannot yet run your business.

AI got dramatically good at programming before almost anything else. The easy explanation is "code is structured." The real one is that code keeps score. Run it: it works or it does not. Write a test: it passes or it fails. That is an instant, cheap, objective verdict — a scoreboard. And a scoreboard is exactly what any learning system, human or machine, needs to climb: try, check, keep what scored, repeat, a million times over.

WHERE AI WINS — AND WHERE IT GUESSES

  • Learning needs feedback. No verdict, no improvement. This is as true for a model as for a kid with a basketball.
  • Code hands out feedback for free. Compilers, tests, and types give an automatic, instant, objective answer. So do math and games.
  • That verdict becomes fuel. You can generate an answer, check it automatically, and reward the win — at enormous scale, with no human in the loop.
  • The hard domains have no scoreboard. Markets, medicine, management, taste — not harder thinking, just no clean score. The verdict is delayed, noisy, or a matter of opinion.
FIG 5.1 — AI ADVANCES FASTEST WHERE THE SCOREBOARD IS INSTANT AND OBJECTIVE.

Once you see it

This quietly reorders which work feels the heat. It is not about how clever the task is. A grandmaster-level move in Go is scored in an instant; a brilliant five-year business bet is not scored for five years. Same difficulty of thought, wildly different feedback. So the useful question about any job is not "is this hard?" It is: where is the scoreboard? If you cannot point to one, expect confident guessing, not reliable mastery — at least for now.

AI gets good at
whatever keeps score.

TRY IT — 3 MINUTES

Give an agent — or plain Claude — a task with a scoreboard, then one without, and watch its confidence stay identical while its reliability doesn't:

  • "adjust these expenses so they total exactly $500" — it adds them up, sees it's off, corrects, checks again. A gauge exists, so it converges.
  • "pick the best stock to buy next week" — same confident tone, but nothing checks it, and nothing can for months.

Same model. The only difference is whether a scoreboard exists. That difference is the whole post.

Next: memory, a role, wheels, gauges — and one destination. The last piece is the driver that puts the whole car in motion.

scoreboard

FOR THE CURIOUS — WHAT'S ACTUALLY HAPPENING

Models are first trained to imitate text, then refined with reinforcement learning. The earliest refinement used human preference as the score — people rating which answer was better. The recent jump in reasoning came from a cheaper, harder scoreboard: verifiable rewards.1 On math and code you can check an answer automatically — the proof holds, the tests pass — so you can reward correctness directly, billions of times, without a human grading each one. That is exactly why the newest reasoning models leapt ahead on math and code specifically, and not, say, on management advice.

Which makes the real frontier less glamorous than "a bigger model." It is building trustworthy evaluators — verifiers, simulators, graders — for domains where truth is slow or subjective. Whoever figures out how to put a reliable scoreboard on messy, real-world work unlocks the next set of problems. Until then, treat fluent confidence in a score-free domain as exactly that: fluent, and confident.

FIND THE SCOREBOARD

Point an AI agent at a job and ask one question first: how fast and how cleanly does it learn it won or lost?

  • Coding agent — writes code, runs the tests, sees red or green in seconds. Clean, instant scoreboard. This is the easy case.
  • Sales-email agent — sends, then waits days for a reply that may never come, and can't tell if the email or the timing or the product won the deal. Delayed and noisy.
  • Hiring agent — screens a résumé; whether the hire was good is known months later, if ever. The verdict arrives too late to learn from.
  • Investment agent — recommends an asset; the result is lagged, noisy, and tangled with luck. A win can't be cleanly credited to the decision.
  • Medical-advice agent — “correct” depends on the patient and on expert judgment. No automatic grader exists, so trust has to be earned a slower way.

The opportunity isn't only “where does AI work today.” It's building the missing scoreboard where there isn't one yet.

SOURCES & NOTES

  1. The shift from human-preference reward (RLHF) to verifiable rewards — rewarding provably-correct math and code directly — is the mechanism behind the recent step-change in reasoning models. DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning” (arXiv:2501.12948).