Because I'm basing it on where things were 5-10 years ago (almost nothing like it actually working yet), not some hypothetical place where the research could be after 50 years.

GPT-3 is also very impressive to me, even though 30 years ago I thought we'd have Hal by now.

Some problems just turn out to be way harder than anyone anticipated, and so when they make advances I'm impressed.

> We should expect a direct neural interface to far exceed the speed of any manual input device

Counterpoint: If this were the case I would have already heard about techies getting brain implants to optimise their communication.

Since that hasn't happened, the only logical assumption is that available neural interfaces are slower than existing manual input methods.

Because manual input devices aren't available to quadriplegics and earlier solutions (e.g. sip/puff controller used to select single letters from an on-screen keyboard) are relatively very slow.