Hacker News new | ask | show | jobs
by jqpabc123 930 days ago
But I'm still amazed how much you can do with this very imperfect, unintelligent "AI".

I'm amazed at how little you can actually do with it.

It's entertaining --- but it really only "works" when you are willing to manually verify the results or just accept the fact that it may be spewing BS. And as you point out, the imperfect results come at a high cost which makes the economics of it questionable for many applications.

Drinking bleach doesn't cure COVID --- but AI may suggest it because it found it on the internet.
2 comments
rowanG077 930 days ago
Verification of a result is often much easier then coming up with the result. Even though chatgpt4 sometimes gives me bs I am much faster in some tasks with it.
link
sorokod 930 days ago
Often implies statistics which are not available.

One counter example is summary verification where you can't do better then summarize yourself.

link
rowanG077 930 days ago
There are definitely things where chatgpt doesn't help much or is worse then doing it yourself. Part of using it effectively is to get a sense when it is effective.
link
jqpabc123 930 days ago
The real question --- Do the benefits really justify the cost?

We don't know the answer to this yet.

link
The_Colonel 929 days ago
> but it really only "works" when you are willing to manually verify the results or just accept the fact that it may be spewing BS

Verifying a solution is often much easier than coming up with a solution. GPT-4 has produced a lot of "hallucinations" when I asked about my programming problems, but they were easy to discount. OTOH it did help me a lot with coming up (or inspiring) solutions which would take me a long time to find otherwise.

Same thing for Dall-E / Midjourney. I'm unable to produce any usable graphics on my own, with these tools I'm able to produce interesting illustrations for my needs, even if I need to discount many results. It literally provides me with a capability I didn't have before. Notice the similar pattern - it's much easier to critique a solution (generated image) than to come up with it.

I believe there are so many applications where this holds true. You can create domain-specific chatbots helping out knowledge workers in almost any discipline.

link