Not sure whether the goalpost is set to LLMs or ML in general, but OP clearly is drawing the line at LLMs, so:

Toaster - it can determine the heat and duration needed for the bread/bagel based on the thickness and density, without needing to manually guess the settings

This is already done with traditional models in rice cookers, pressure cookers, and so on. It doesn’t need LLMs.

Fridge - aware of the contents so it can set the thermostat at the ideal temperature range, even better if it’s in control of internal flaps to automatically compartmentalize the contents for optimal storage

Some smart fridges have tracked their contents for years now and sell that data to advertisers.

Toothbrush - able to scan your teeth while you better and inform you if your brushing was adequate or if you missed any plaque, on which teeth, and warn you of developing cavities

I know there’s some crazy toothbrushes out there, but the closest equivalent I know of is my shaver which actually does give recommendations on how to shave better. Or it would, anyway, if I linked it to a mobile app that’s never going to touch my phone. I got this shaver around 5y ago, again predating LLMs.

Bedside light - aware of if you’re in the bed, and whether you’re attempting to go to sleep, or are asleep, based on preferences and rules you give it beforehand, in plain language

The last thing I want is for my bedside lamp to have a camera and perform object detection in my bedroom just to turn off automatically when I get in bed lol. Plus, I usually don’t want the light off as soon as I’m in bed. In any case, traditional models could do this, but multi-modal LLMs tend to perform better.