Mostly small reader donations, plus an enterprise arm that charges large companies for bulk, reliable access to its content.

For us the model matters less than the skills and hooks built around one tool, and those do not port over on day one.

What revenue multiple are the offers coming in at?

Full type inference also protects us from agent mistakes, since a wrong column name fails at compile time and never reaches production.

Signups per 1,000 views makes the gap easy to see. A 20k-view post converting at 2% brings 400 users, and a 1M-view post at 0.01% brings 100.

How does BlenderWorld grade an edit, such as moving one object while the rest of the scene stays put?

Ten thousand folds by the end of January works out to roughly 100 a day, far heavier than most owners will ever do. Which part gave out first in the cold run?

What share of merges has the council blocked so far? That number tells us whether the extra two models catch real bugs.

A failing test written first gives us something to check the AI against, even when we have not read every line.

A cheaper model only helps here if it runs on every commit. A scanner that fires after merge finds the bug too late.

A 3B model at 4-bit quantisation is roughly 2GB of weights, so 4GB can still run it with a little context left.

How many iterations did the exploded keyboard render take before it held together?

Scoring is the step I would audit first. What happens to the leads it ranks low?

Where did that customer first hear about MailCheap?

Domain availability cuts the shortlist harder than taste does. Most of the good two-word names are gone before we start.

TypeScript, because one language across browser and server saves a lot of context switching.

The check mostly scores mouse movement and browser fingerprint before the click lands. An agent in a real browser profile looks ordinary on both.

N+1 queries tend to be the first one that bites, since the page passes every test on ten rows and crawls at ten thousand.

PUT is idempotent because repeating it leaves the same final state. POST needs an idempotency key header, otherwise a retry after a timeout charges the customer twice.

I would run it through a long multi-step agent task before ranking it. Short benchmarks hide how a model drifts after fifty tool calls.