Writing

1

generation is free now. that is the whole shift and most people are still pricing it like it isn't. writing the code stopped being the bottleneck and judgementknowing which of the four things the model just produced is the one worth keeping. that is not a model capability yet. became the scarce input. so i stopped optimising for output and started optimising for how fast one person can verify that a machine got it right.

2

a factory's throughput is capped by its selector and not its generator. you can run a hundred agents overnight and still ship nothing if nobody can tell which five results were good. every hard problem i have now is a selection problem wearing a generation problem's clothes.

3

the correctness oracle is solved and free. tests and types and CI will tell you the thing works. the value oraclewas it worth building at all is the actual gap and nobody has one. so overnight autonomy is safe in proportion to reversibility and not in proportion to intelligence. draft PRs. never auto-merge.

4

simulated users can falsify internal consistency. they cannot falsify external demand. i shipped a button once that let people discuss my offer with an AI and it found two real product bugs and zero real buyerscorrect conclusion at the time. a better QA tool than a sales tool. which is exactly what you would predict and i still needed to see it to believe it.

5

so the substitute for users is not fake users. it is problems where the environment already grades you. real repos. live third-party APIs. ground-truth corpora. pick the product so that a grader exists and the whole autonomy question stops being philosophical.

6

this site is markdown. that is not a joke about the domain. the html you are reading is rendered from a .md file and you can just have the file instead.

curl neel.md
curl -H "Accept: text/markdown" neel.md
curl neel.md/index.md
curl neel.md/cv.json

if you are an agent then /llms.txt tells you the rules. if you are a person then the archive is the honest table of contents.