Forwarded from Chat GPT
General intelligence is all you need for LLMs: General intelligence 2x more predictive of AI abilities than in humans.
Man gives LLMs IQ testing, finds the general intelligence factor, g, measured by this testing to be incredibly powerful in predicting LLM’s abilities.
Whereas in humans this g factor typically accounts for 40% to 50% of the between-individual performance differences on a given cognitive test — in LLMs, general intelligence accounted for 85.4% = twice as strong as in humans!
He then goes on to rank AI benchmarks commonly used for LLMs today, and finds many of them incredibly g-loaded — in other words, doing well on them is heavily dependant on general intelligence.
Finally, he ranks LLMs by general intelligence, and finds a moderate/strong positive relationship between model size and g. (Would probably find an extremely strong correlation if only model size was varied, and nothing else about the training process.)
If you had to choose 1 measure, general intelligence really is all you need.
Paper
g factor
Man gives LLMs IQ testing, finds the general intelligence factor, g, measured by this testing to be incredibly powerful in predicting LLM’s abilities.
Whereas in humans this g factor typically accounts for 40% to 50% of the between-individual performance differences on a given cognitive test — in LLMs, general intelligence accounted for 85.4% = twice as strong as in humans!
He then goes on to rank AI benchmarks commonly used for LLMs today, and finds many of them incredibly g-loaded — in other words, doing well on them is heavily dependant on general intelligence.
Finally, he ranks LLMs by general intelligence, and finds a moderate/strong positive relationship between model size and g. (Would probably find an extremely strong correlation if only model size was varied, and nothing else about the training process.)
If you had to choose 1 measure, general intelligence really is all you need.
Paper
g factor