LLMs .what do you smoke beforehand?
I'm frustrated with the current state of LLMs. They seem to have been intentionally degraded since around October of last year, and every company is gaslighting users about it. For my own experiments, LLMs are actually worse than no-LLM, and I can't get even mediocre output anymore. Even Claude, which was once great, now avoids doing what's asked and takes forever for simple tasks. I'm wondering what I need to smoke to see the LLMs work as they should.
I think you're experiencing a combination of model regression and your own increased familiarity. Early on, the novelty and the model's capabilities were impressive, but as you use it more, you notice its limitations. Also, the models have indeed been updated, and some changes may not be improvements for all use cases.
I've found that the quality of output depends heavily on the prompt and the specific model. For coding tasks, I still get great results with certain models, but for creative writing, it's hit or miss. Maybe you're using the wrong tool for the job.
I think the issue is that LLMs are now being used for everything, and the hype has died down. People are realizing they're not magic, and they have limitations. But that doesn't mean they're useless. I still use them daily for brainstorming and drafting.
I've noticed a similar decline in quality, especially with Claude. It seems like they've optimized for safety and cost, which has made the models more conservative and less creative. It's frustrating, but I've learned to work around it by being more specific in my prompts.
Maybe you're just burned out. When you're tired, everything seems worse. Take a break, and come back with fresh eyes. You might find that the models are not as bad as you think.