I’ve been posting comments frequently on Hacker News recently, mostly on threads related to ChatGPT.
I find myself incredibly impressed with ChatGPT, even the 3.5turbo version (aka Mar 14) currently available at chat.openai.com. It is an incredibly useful tool that surprises me regularly with amazing responses. I completely understand why so many people anthropomorphize it, and ascribe reasoning skills to it!
It also completely fails on a near-daily basis in my testing, and in odd ways, reminding me regularly that it’s mathematically-driven spicy autocomplete.
If you’re a skeptic, set your skepticism to the side for long enough to try it out, and see if you aren’t impressed.
If you’re a true believer, step back a bit. I think you’re falling into a very human trap, one which ChatGPT probably won’t, ironically.
If you’re fearful, don’t be! These are impressive tools, but they still need to be used by someone, and you can be that someone. If people end up losing their jobs because of ChatGPT, it will be because oligarchs decided to cut costs, not because spicy autocomplete took your job.
@pwinn Exactly my take on it. Spicy auto complete is really quite impressive, without ascribing mystical qualities to it.
@pseudonym I feel like I’m learning a bit more about what language can and can’t do, and I’m astounded at the extent to which reasoning can be simulated with pattern-matching. Very, very impressive.
@pwinn Yup. You just need a big enough corpus of pre-reasoned text, to extract the relationships from. Semantics, derived from syntax.
I like to noodle on the fact that the existence of LLMs functioning as well as they do, is more an indictment of, or evidence for, the fact that what we do to reason on novel data, is really quite structural and predictable.
If they can parrot us as well as they currently can, how much more is uniquely human?
The "God of the gaps" continues to shrink.