In the age of artificial intelligence, super-intelligent machines like GPT have become a reality, leading to the question of how to quantify or measure their intelligence. While IQ tests are widely used to measure human intelligence, could a similar approach be used to assess the intelligence of machines? Is it possible to compare the intelligence of a machine to that of a human in this context, given that machines possess certain cognitive abilities and limitations that differ from those of humans? Additionally, what are some alternative methods for evaluating machine intelligence, and how do they differ from traditional IQ tests used for humans? Finally, what implications might arise from the development of super-intelligent machines, both in terms of how we measure and compare intelligence and in terms of their impact on society as a whole?
-
3We can certainly just give GPT an IQ test, but it might not be a meaningful comparison. We rank humans against each other by assuming all humans are kinda similar (a bad assumption BTW as IQ tests have been found to be biased towards white people) and we assume humans who are better at a set of tasks are probably better at most tasks. Ranking a human against a GPT might not be useful because it doesn't follow this rule - it's good at some tasks and bad at others – user253751 Apr 01 '23 at 09:43
-
@user253751: For me, your comment would also work as an answer. – Broele Apr 01 '23 at 14:01
-
@Broele it's not clear if what it answers is what the question asks – user253751 Apr 01 '23 at 14:01
-
@user253751 Your comment addresses both of my questions. – R1- Apr 02 '23 at 02:23
-
I see four or three questions, if the first and the second are merged: IQ for machines, alternative methods and social implications of super-intelligence. Is that broad scope the reason for the close vote? – Jaume Oliver Lafont Apr 02 '23 at 08:58
-
1In which sense is ChatGPT super-intelligent? If counting the letters in a sentence is a necessary condition for intelligence, then ChatGPT is stupid. ChatGPT is just a stochastic parrot. If you consider that intelligent, ok, but well good luck with your use of it... – nbro Apr 03 '23 at 13:18
-
@nbro you are not fighting me, rather than you fighting yourself, I have a feeling by saying "super-intelligent" is like the old days when Galileo Galilei told that the Earth is not flat. the downvotes have no effect on the question, even voting to close the question is not important, me asking the question is out of curiosity after watching some interviews of Lex Fridman with some guests that the talk was all about GPT4, such as Sam Altman, Ilya Sutskever and Edward Frenkel. I believe you are making the AI.SE worthless with such reactions, we will find out soon whether GPT4 is stupid or not. – R1- Apr 12 '23 at 21:27
-
1At present, ChatGPT can only accept text input. IQ tests involve a lot of images, so that would seem to rule out IQ tests for the time being. – Dennis H Apr 29 '23 at 19:56
1 Answers
tl;dr– Yes, we can give machines IQ-tests. And we basically already do, just often skipping the last step of explicitly recognizing an IQ-score.
IQ-scores can be assigned based on relative performance on some intelligence-test.
Sure, we can give machines IQ-tests.
To be clear, IQ tests are basically just:
Create a test that assigns higher scores to test-takers with higher intelligence.
- For example, if you're primary interested in "intelligence" as in "ability to solve crossword-puzzles quickly", then you can give a test with a bunch of crossword-puzzles, then score test-takers based on stuff like how quickly and completely they do the crossword-puzzles.
Test a bunch of intelligences and score their exams.
Rank the test-takers from lowest-to-highest.
Assign each test-taker a percentile-ranking.
- For example, if a test-taker got a median score, then their percentile-ranking would be 50%.
Calculate IQ by casting the percentile-rankings to a normal-curve with a median of 100 and a standard-deviation of 15.
Equation's $\text{IQ} = \sqrt{2} \operatorname{erf}^{-1}{\left(2p - 1\right)} ~~ p \in \left(0, 1\right) \, ,$ where $p$ is the percentile-rank (e.g., $p=0.5$ for a median-percentile of 50%).
On WolframAlpha: $\texttt{100 + 15 * sqrt(2) * erf^(-1) (2 *}\underbrace{\texttt{ 0.5 }}_{p}\texttt{- 1)}$.
On Microsoft-Excel: $\texttt{=100+15*NORM.S.INV(}\underbrace{\texttt{0.5}}_{p}\texttt{)} \,.$
Folks basically already do this with various AI-models, e.g. when ranking their relative performance on a series of tasks, just don't normally do Steps 4 or 5.
How to compare human-IQ's with machine-IQ's?
Basically just give both humans and machines the same test and rank them, then calculate. That's it.
The main thing would be picking a test that'd test for some notion of "intelligence" that you'd be interested in.
- 142
- 1
- 1
- 8