Excellent article.

While I'm big fan of art generation AI tools like Stable Diffusion & Midjourney, I've been wary / suspicious of ChatGPT and struggling to make sense of how a model based on the probability distribution over words could become anything more than a glorified wikipedia / search engine combo with questionable reliability. Your article confirmed my suspicions that ChatGPT & GPT-3 are overhyped and ovepromised.

One minor remark, though. I disagree that GTP-3 is sexist for presuming the woman is a nurse and the man is a doctor. All specialties combined, men are roughly 2x as likely as women to be doctors, according to https://www.aamc.org/data-reports/workforce/interactive-data/active-physicians-sex-and-specialty-2015. And according to https://www.zippia.com/nurse-jobs/demographics/, more than 85% of nurses are women. As such, from a probabilistic point of view it makes perfect sense to presume that the man is a doctor and the woman is a nurse, as both the gender distribution of doctors & that of nurses make this by far the most plausible scenario.