IQ Tests Results for AI

Re:IQ Tests Results for AI by pama on Sunday August 17, 2025 @10:20AM (#44931750)

IQ of 200 (or higher) do not exist according to the original definitions of this metric. You need a population of 219 billion or higher to have a 95% chance that a sample exists with 6.66 standard deviations away from the mean (assuming mean of 100 and std of 15). Ofc the tests are of limited value and things can be gamed, but it would be silly to try and identify samples that have no chance of existing.

[ Parent | on HN ]

Re:IQ Tests Results for AI by csa on Sunday August 17, 2025 @10:56AM (#44932051)

> IQ of 200 (or higher) do not exist according to the original definitions of this metric

Just to expand in this point…

Most IQ tests for adults lose a lot of precision over 130 (2sd), and they are extremely imprecise over 145 (3sd) — almost to the point that a scores over 145 should simply be labeled 145+.

When I did a deep dive into the IQ test literature 20 years ago, the most reliable correlated predictors for 145+ were standardized tests like the GRE. That said, standardized tests like these have high specificity and (relatively) low sensitivity — that is, very few false positives and many false negatives.

[ Parent | on HN ]

Re:IQ Tests Results for AI by 010101010101 on Sunday August 17, 2025 @10:47AM (#44931955)

“Is unlikely to exist” and “does not exist” are two very different statements, but it really doesn’t matter because GP’s point stands if you replace the exaggerated IQ with something reasonable like 160.

[ Parent | on HN ]

Re:IQ Tests Results for AI by pama on Sunday August 17, 2025 @11:55AM (#44932497)

OK. The more precise statement is that when the test was created you would have had about 99.3% to 99.42% chance that such a sample did not exist if you were to test all the population this test was designed for (depending on population statistics of the time that are a bit unclear).

To be clear, I do not endorse the validity of these tests or their interpretation at any level. Learning to be a lifelong learner can take almost anyone a really long way. The analogy to neural nets is that bigger nets dont always make a better model after a point and every human starts at a very priviledged/huge network capacity.

[ Parent | on HN ]