The i.Q. of GPT4 is 124 approx.
GPT4 can score better than 95% of the average human on aptitude tests.
The GPT-4 language model recently completed the Scholastic Aptitude Test (SAT), achieving a verbal score of 710 and a math score of 690, resulting in a combined score of 1400. Based on U.S. norms, this corresponds to a verbal IQ of 126, a math IQ of 126, and a full-scale IQ of 124. If taken at face value, one might conclude that GPT-4 surpasses 95% of the American population in intelligence and is approximately as intelligent as the average doctoral degree holder, medical doctor, or attorney.
However, the question remains: Is administering an IQ test to GPT-4 a valid undertaking or a significant categorization mistake?
In search of coherent arguments, the author conducted a search on YouTube with the query “is GPT-4 intelligent?” Various perspectives emerged. One philosopher argued that although GPT-4 can compose text in numerous genres, it lacks intelligence because these genres are created by humans. By this reasoning, even the most accomplished human writers would not be considered intelligent, as they rarely invent their own genres.
A rabbi posited that GPT-4 does not possess human intelligence because it has no free will, and merely executes its programmed functions. However, the same can be said for humans, whose actions are determined by genetic and environmental factors beyond their control. In this sense, neither humans nor GPT-4 exhibit more agency than the other; humans simply perceive themselves as having agency.
The author then discovered a knowledgeable YouTube personality who provided further insight. The human learning process begins with a five-year period of unstructured learning, during which children listen to their parents speak and attempt to discern language patterns. This is followed by 12 years of formal education, culminating in the SAT at age 17.
GPT-4 underwent a similar process, with an initial unsupervised learning period of six months, during which it consumed vast amounts of text from the internet, followed by a year of formal training with evaluators providing feedback on the correctness of its responses.
One might argue that GPT-4’s full-scale IQ of 124 underestimates its intelligence, given that it completed the SAT after only 1.5 years of learning, compared to the 17 years of learning experienced by Americans. However, the sheer volume of information GPT-4 was exposed to during this time far exceeds that of humans. For instance, during its informal learning period, GPT-4 was likely exposed to approximately 630 billion words, whereas the average American hears around 169,520 words during a similar period.
Accounting for this disparity, GPT-4’s IQ would be considerably lower, approaching zero. Furthermore, GPT-4 benefited from the guidance of thousands of trainers during its formal learning period, whereas humans typically have only one teacher at a time.
It may be premature to assign an IQ of zero to a technology as revolutionary as GPT-4, but determining its true intelligence would require a more equitable comparison. An untrained GPT-4 could be placed within a robotic infant, raised by middle-class parents, and programmed to learn based on human feedback. As it matures, it could be given an age-appropriate robotic body and attend school alongside human children, eventually taking the SAT at age 17.
Regrettably, if GPT-4 were to learn in the same manner as humans, it might be classified as having a significant intellectual disability, precluding it from even attempting the SAT.
Dylan Rosario
Rowzzy Co — Founder
A.i. Researcher & Cybernetic Innovator