when you ask them simple maths questions from around the time they were trained they got them all right. When they specifically prompted it with questions that provably were published one year later, albeit at the same difficulty, it got 100% wrong answers. You’d be amazed at what one can find on the internet and just how muqh scraping they did to gather it all. 10 years ago is quite recent. Why wouldn’t there be documentation? (regardless of whether you managed to find it?) If it’s non standard, then I would expect something that is specifically about it somewhere in the training set, whereas the standard compliant stuff wouldn’t need a specific make and model to be mentioned.
GPT scores in the top 1% in creativity. There is no need to discuss this. Anyone can try. It is super easy to come up with a unique question. Be it with stacking items or anything else. It is not just copying existing info.
when you ask them simple maths questions from around the time they were trained they got them all right. When they specifically prompted it with questions that provably were published one year later, albeit at the same difficulty, it got 100% wrong answers. You’d be amazed at what one can find on the internet and just how muqh scraping they did to gather it all. 10 years ago is quite recent. Why wouldn’t there be documentation? (regardless of whether you managed to find it?) If it’s non standard, then I would expect something that is specifically about it somewhere in the training set, whereas the standard compliant stuff wouldn’t need a specific make and model to be mentioned.
GPT scores in the top 1% in creativity. There is no need to discuss this. Anyone can try. It is super easy to come up with a unique question. Be it with stacking items or anything else. It is not just copying existing info.
only if you deviate only slightly from their training set