3 ms·

GPT 4 being only 3% does not convince me. If you ask well known stuff yes it will mostly get it right, but as soon as you ask something a little bit less known

by h1fra 3y ago

GPT 4 being only 3% does not convince me. If you ask well known stuff yes it will mostly get it right, but as soon as you ask something a little bit less known it will happily invent anything. About half of my answers from ChatGPT comes with lies or partial info.

troupo 3y ago
It's also extremely bad at anything resembling any math or numbers: calculations, unit conversions, weather etc.
- Kiro 3y ago
  It's bad, but not extremely bad imo. With Code Interpreter I would even say it's decent at conversions (which it defaults to now if you ask it to convert something).
  - d3w4s9 3y ago
    Without concrete numbers to quote, bad == extremely bad
  - troupo 3y ago
    It's extremely bad because it's plausibly bad. Just yesterday I was converting regular cooking units for a friend, and checking against google. It got most of it right.
Kiro 3y ago
I also think 3% sounds low but half of your answers? What kind of questions are you asking?
- pizza234 3y ago
  In my personal experience, the rate of hallucinations is proportional to the technical depth of a question. Questions like "how to build project X" have a high rate of hallucinations, for example.
  - johnisgood 3y ago
    It often tells me to use non-existent libraries, this has been the most common experience for me with regarding to inaccurate information provided by ChatGPT; the next most common is probably using non-existent (not simply deprecated) functions within existing libraries.
- tveita 3y ago
  Things you don't already know? For any topic you already know the basics of but want to explore some particular thing, the first answer will be vague and if you ask followup questions will start making stuff up, IME. For programming the popularity of a language matters a lot - Javascript is probably one of the better ones? I tried asking it completely basic questions about Racket for last year's Advent of Code and it would immediately start making up functions.
sprobertson 3y ago
Your anecdata isn't in the evaluation set - this is summarization
intended 3y ago
Yeah its low, other papers have seen hallucination rates of 30-40% If I understand this model correctly, its doing it over a curated list of content. I would say that for this specific test it achieves 97%, and assume a rate of 70%. Verifying factualness is hard to automate, and all measures are proxies for what an actual human would say.

by h1fra 3y ago

troupo 3y ago

It's also extremely bad at anything resembling any math or numbers: calculations, unit conversions, weather etc.

Kiro 3y ago
It's bad, but not extremely bad imo. With Code Interpreter I would even say it's decent at conversions (which it defaults to now if you ask it to convert something).
- d3w4s9 3y ago
  Without concrete numbers to quote, bad == extremely bad
- troupo 3y ago
  It's extremely bad because it's plausibly bad. Just yesterday I was converting regular cooking units for a friend, and checking against google. It got most of it right.

Kiro 3y ago

I also think 3% sounds low but half of your answers? What kind of questions are you asking?

pizza234 3y ago
In my personal experience, the rate of hallucinations is proportional to the technical depth of a question. Questions like "how to build project X" have a high rate of hallucinations, for example.
- johnisgood 3y ago
  It often tells me to use non-existent libraries, this has been the most common experience for me with regarding to inaccurate information provided by ChatGPT; the next most common is probably using non-existent (not simply deprecated) functions within existing libraries.
tveita 3y ago
Things you don't already know? For any topic you already know the basics of but want to explore some particular thing, the first answer will be vague and if you ask followup questions will start making stuff up, IME. For programming the popularity of a language matters a lot - Javascript is probably one of the better ones? I tried asking it completely basic questions about Racket for last year's Advent of Code and it would immediately start making up functions.

sprobertson 3y ago

Your anecdata isn't in the evaluation set - this is summarization

intended 3y ago

Yeah its low, other papers have seen hallucination rates of 30-40% If I understand this model correctly, its doing it over a curated list of content. I would say that for this specific test it achieves 97%, and assume a rate of 70%. Verifying factualness is hard to automate, and all measures are proxies for what an actual human would say.