> I asked Claude if there is such a way to basically hard-code branch prediction rules into the machine code, and the answer was that there’s no way to do this on x86, but there is a way on ARM: the BEQP (predict branch taken) and BEQNP (predict branch not taken) instructions.

> Those ARM instructions are just hallucinated, and the reality is actually the other way around: ARM doesn’t have a way of hard-coding ‘predictions’, but x86 does.

This made me chuckle. Thanks.

To be fair, on the x86 side those branch prediction hint prefixes have been functionally ignored by pretty much all cores for about two decades.

If a human wrote that here (on HN) someone would note the error and the poster would reply:

Yes, sorry, you’re correct. I’ve usually had 97 more double ristrettos by this time in the morning.

Some schools of though suggest this has already happened.

I asked Claude once if there was a way to open a tun/tap on Windows without a driver. It hallucinated an entire supposedly undocumented NT kernel API that as far as I can tell has never existed, complete with parts of this hallucinated API being depreciated in favor of newer non-existent parts.

It was so detailed it makes me wonder if maybe it was in the training data somewhere. Maybe it ingested an internal MS doc for a proposed API or something. The case of the missing ARM instructions makes me wonder the same. Maybe someone on a forum proposed them and they were ingested.

I did actually verify on the OS that the calls do not exist in the kernel or any driver or DLL on the system.