ngl code-llama has been better

even claude-2 sometimes

gpt-4 is not as good as it used to be

janitor for /zk and some more stuff here linktr.ee/randomishwalk

Just noticed this too. Weird that it’s getting worse. What would make that happen? I thought it was ChatGPT specifically and they were making adjustments behind the scenes that were degrading it

Mobile and web developer, building re-public.io, journodao.xyz, demo.factland.org. Ex-WIRED. Writer, dadder, musicer. @developer_dao.

RLHF (leading hypothesis i think) and other stuff

haven’t seen benchmarks though. this is just anecdotal on my end