Anthropic just published a graph that’s worth sitting with - “education years needed to understand the human prompt” vs. the share of workers with a Bachelor’s degree (by occupation).
You presume that only those who hold degrees are actually intelligent, but that intelligence is inherent, not on paper. This study is flawed, like much of Anthropic's mess, because they only looked at those who had paper qualifications. Yet there are masses of people whose situation in life didn't enable them to gain a degree yet, if tested, would likely exceed the general intelligence of many of those degree holders. And I've seen many who claim to be degree holders who were abysmal at knowing how to talk to AI. This study is so epically flawed by being entirely biased as to what anthropic thinks intelligence looks like, which explains greatly their studies into what they claim are 'ethics'.
I agree with you that a paper qualification is no guarantee of intelligence (and vice-a-versa) - so I do not presume what you think. It is a flawed methodology, but it is also a "reasonable" starting proxy and does provide some insights. It's not the answer, but it does highlight an interesting relationship. People get out what they put in. RLHF tunes models to respond to match the user. And this is supported by the AI platforms economic driver of "engagement". This all needs much more discussion.
You presume that only those who hold degrees are actually intelligent, but that intelligence is inherent, not on paper. This study is flawed, like much of Anthropic's mess, because they only looked at those who had paper qualifications. Yet there are masses of people whose situation in life didn't enable them to gain a degree yet, if tested, would likely exceed the general intelligence of many of those degree holders. And I've seen many who claim to be degree holders who were abysmal at knowing how to talk to AI. This study is so epically flawed by being entirely biased as to what anthropic thinks intelligence looks like, which explains greatly their studies into what they claim are 'ethics'.
I agree with you that a paper qualification is no guarantee of intelligence (and vice-a-versa) - so I do not presume what you think. It is a flawed methodology, but it is also a "reasonable" starting proxy and does provide some insights. It's not the answer, but it does highlight an interesting relationship. People get out what they put in. RLHF tunes models to respond to match the user. And this is supported by the AI platforms economic driver of "engagement". This all needs much more discussion.