Post AWDshk0m6Wie5jKcFs by burnthekernel@emacs.ch
(DIR) More posts by burnthekernel@emacs.ch
(DIR) Post #AWCy2whsP7oS9uSk3E by galdor@emacs.ch
2023-05-31T08:19:20Z
0 likes, 0 repeats
There are complaints that the output quality of GPT-4 is going down. Act surprised when they introduce tiered pricing and you have to shell out a ton of money for the good stuff.Microsoft did not invest 10B in OpenAI because they want to give it away for pennies.
(DIR) Post #AWDd1e9O0UECRFpT6W by timthelion@emacs.ch
2023-05-31T15:58:30Z
0 likes, 0 repeats
@galdor I personally have never found gpt4 to be better than 3.5. I'm not even sure how to objectively judge them. Both produce working code and both occasionally produce wild hallucinations. Maybe it's a placebo effect?
(DIR) Post #AWDebyyY9lI5PEePBo by galdor@emacs.ch
2023-05-31T16:16:17Z
0 likes, 0 repeats
@timthelion It's also possible that the naïve enthusiasm of the beginnings is starting to fade, and all the bad results that used to be ignored are now sticking out like a sore thumb.
(DIR) Post #AWDf4NeZzwfi3Fqp3A by timthelion@emacs.ch
2023-05-31T16:21:25Z
0 likes, 0 repeats
@galdor Probably. Once I got over the shock that the image generators did anything at all, they became quickly boring. But I still use LLMs daily, and I think I'm getting more out of them as time goes on. What I'm seeing a lot of, is that LLMs are very influenced by coding style. My coding style, coming from a functional background, is very good for LLMs because functions are well isolated and need limited context to understand. Trying to do things in OOP style with LLMs gives much worse results. The bigger the relevant context, the worse the results...
(DIR) Post #AWDshk0m6Wie5jKcFs by burnthekernel@emacs.ch
2023-05-31T18:54:10Z
0 likes, 0 repeats
@galdor Almost certainly. OpenAI calls their $20 subscription plan "Plus". Leaving plenty of room for more tiers of service.