not exactly daily, but I use llama3.2-vision (via ollama) to generate a .txt file alongside every photo I take, containing a description of the photo. then I can just grep for, say, "selfie":

    $ alias omglol='find -name \*txt | xargs -n50 grep -li'
    $ omglol selfie
    ./10/IMG_20241019_204444.txt
    ./09/IMG_20240930_082108.txt
    ./09/IMG_20240930_082118.txt
    ./07/IMG_20240712_154559.txt
    ./07/IMG_20240712_154554.txt

or to do a slide show:

    $ omglol selfie | sed s/txt/jpg/ | xargs feh

Small pretrained models are often too “dumb” to be useful, but, if you have a task in which you can tune a model, their pretraining means you can produce an effective model with a shockingly small corpus. These can be more reliable than off the shelf models in an automated process, because even the large pretrained models have a lot of “behaviors” you can trigger with surprising inputs. Most recently I retrained SmolLM2 to translate intents to SDXL prompts.

I use Deepseek via LLM studio for reading sensitive/non-sensitive docs, contracts, searching bank statements & bills.