This comment is absolute bullshit.

It starts off being wrong ("Opus 4 has maxed out LLM coding performance"), then keeps being wrong ("LLM inference is sold at a loss"), and tries to mask just how wrong it at any point in time is by pivoting from one flavor of bullshit to another on a dime, running laps a manic headless chicken.

Chinchilla maxed out refers to the so-called "Chinchilla Scaling Law" from the famous DeepMind paper about how in this particular regime, scale seemed to just flow like the spice. That happens sometimes, until it doesn't.

I didn't say the coding performance was maxed out, I said the ability to pour NVIDIA in and have performance come out the other side is at it's tail end. We will need architectural innovations to make the next big discontinuous leap (e.g. `1106-preview`).

They're doing things they don't normally do right: letting loose on the safety alignment bullshit and operator-aligning it, fine-tuning it on things like nixpkgs (cough defense cough), and generally not pretending it's an "everything machine" anymore.

This is state of the art Google/StackOverflow/FAANG-megagrep in 2025, and it's powerful (though the difference between this and peak Google/SO might be less than many readers realize: pre-SEO Google also spit out working code for most any query).

But it's not going to get twice as good next month or the month after that. They'd still be selling the dream on the universal magic anything machine if it were. And NVIDIA wouldn't be heavily discounted at every provider that rents it.