LogFAQs > #978752036

Uh, yes, GPT 4s training data contains good works, but it also contains a lot of bad ones. When theyre pulling from such a wide pool, theyre necessarily capturing a minority of really good things and a large majority of mediocre or bad things. Thats just how human production is. And the primary way OpenAI has advanced GPT iterations has been by simply expanding the data set.

When it comes to inputting prompts that will spit out a school essay in particular, most of the data for GPT to reference will be meh. Because GPT doesnt think, it cant look at a good reference work and a bad reference work and judge them apart. It can only listen to the prompt input and create the output that most closely matches that based on its training data.

I have ChatGPT 4 on my phone right now, I bought a month subscription to try with my job and write some JavaScript, and I found that up to maybe 40% of its lines would contain errors on first pass through. As one example.

i actually thing what text transformers are probably best at is unfortunately that AI girlfriend stuff lol. This is where GPT blows our human minds because you basically can text it and it sounds like a person could be texting right back to you. By expanding context memory in new GPT iterations, it has gotten really good at remembering different traits, facts and history and thus continuing the appearance of a real conversation partner.

But for critical writing, GPT 4 still favors summary and obviously it cant really make new observations about a work. I tried it in a class I took on the 70s film to compare to my own, and personally if I had been a grading assistant I wouldve given the outputs I saw C+ or B- at best. Yes they can be coherent as if a persons writing, which is certainly an achievement on its own. But in my opinion it is not a beast. Granted, my standards arent those of some student trying to scrape by a required class.