Anyone else think LLMs are getting worse? Even the same model quality seems to be slipping over time.
I'm really not sure if they are messing with the system prompts or I'm getting better at spotting LLM bullshit.
Discussion
Yes. I've been producing some reports with AI and recently it breaks established formatting, almost daily. Simple things like "no external links". I literally told it todays that I fell like I'm talking to GPT 2.5.
I just had qwen 235 reasoning give me 7 real pubmed links to studies that aren't on the subject it said they were on. 4 rounds of "I'm sorry here are the real links" each with real links to studies unrelated to the subject before it gave up and stopped citing sources.
Yes 💯
And you can tell when the llm is in overuse mode, starts to reeeeallly glitch 😅
The classic LLM hype cycle: drop a shiny new model with tons of fanfare, let people play with it in FP16, then quietly switch it to FP4/AWQ once the hype peaks. Rinse, repeat, profit 💰
Maybe they're training on the conversations they're having with people dependent on them - a definite downgrade from the classics of history!
Newer LLMs suck. I'd rather use an older one, as that's more reliable than the new, shiny slop.