[EveYoung]

In my experience, the threshold to be useful is much lower than GPT-3.5. These smaller models can "easily" be finetuned to achieve a comparable performance on a specific task. For example, I've achieved promising results for data summarisiation and image captioning (BLIP2-based) using Alpaca.

Also, server/hardware costs are still a limiting factor for running and finetuning the larger 33/65B Llama models. Especially, if they can only be used for personal toy projects.

[bugglebeetle]

I don’t use LLMs for anything image related, so I can’t speak to their value there, but almost all simpler NLP tasks are IMO better handled using other techniques that predate them. I’ve yet to see an example where fine-tuning is cheaper/more efficient/better performing than older solutions to these problems.

[EveYoung]

If older techniques work for you, there is of course no reason to switch to LLMs besides general curiousity or to explore what's possible already. That said, in my case I was enable to generate much more engaging text summaries of tabular data using a Llama derivative.