sententia

Code time tokens in 3 acts

plans that either come to naught, or half a page of scribbled lines Post status: useless Saturday’s thoughts bound to be invalidated by some future machines of loving grace.

Pithy answer. only real resource in life is time and time is tokens and tokens are money thus write good code to more quickly adapt to environment shifts.

Q3 is the one with questions.

Q1 and Q2 seem to claim (1) shorter code is easier to maintain (who’d have thought) and (2) by default not produced by more capable models.

Suggested conclusion might be to periodically backtest and maintain peer review hygiene. Feature request triage and adaptation is what will differentiate future complex systems. So: intelligence.

Q1: Does code quality matter?

One-sentence response: as long as human alignment matters, the readability of some form of code also matters (have your pick: prose, Claude, brain-to-brain interfaces, Python, or – God forbid – Java).

Scenius not genius.

More objectively, cf. empirical evidence to demonstrate code quality can reduce overall maintenance cost (huh sounds familiar). Compress!

Q2: Do frontier models care?

Even the birds are chirping on insufficiency of general purpose evals for specific-purpose use; however, some people have at least thought through facets of good code.

FrontierCode considers likelihood of merging in their score, demonstrating more thinking tends to produce larger & unnecessary edits (less compression).

Q3: What is code?

Frontier models are increasingly built by entire companies. Math people seem to think Lean is code. Terry the famous fox makes a point of a big math movement. We can tackle ever larger problems via coordination.

To the degree edge case analysis can be useful. In the limit the entire society builds the artifact.

Pithier answer Only real resource in life is time thus devise robust systems to sustain arbitrary environment shifts.

What do you want to keep legible?