Yeah, but do either of those match the aims?

If you have a face unlock, you only rarely need to run it. It’s not like you’re constantly doing face recognition on a stream of video. You don’t have the power-consumption problem.

If you have an archive of 10000 photos, you probably don’t need to do the computation on battery power, and you probably don’t mind using a GPU to do it.

I mean, I can definitely imagine systems that constantly run facial recognition, like security cameras set up to surveil and identify large crowds of people in public areas, but:

• I suspect that most of them want access to a big database of face data. I don’t know how many cases you have a disconnected system with a large database of face data.

• I doubt that most of those need to be running on a battery.

The reason I mention speech recognition is because I can legitimately see a laptop user wanting to constantly be processing an incoming audio stream, like to take in voice commands (the “Alexa” model, but not wanting to send recorded snippets elsewhere).

Probably overkill for facial unlock (though it might make it more accurate), but for just facial recognition/processing there are still plenty of use cases, including:

• Security systems: Imagine it doesn’t just recognize faces but being able to go “analyse the face in this time frame and correlate it to anyone in the last two weeks” to determine that the guy who robbed you is actually the same dude who delivered a package or was supposedly doing door-to-door sales a week ago

• Home automation systems, with per resident configurations. Maybe it’s not unlocking your computer but rather your house via a good camera and ZigBee lock.

• Better face mapping: This could be for substituting on faces (deep-fake’ish stuff) but also see-aging, better real-time mapping facial expressions to an online avatars etc

And going beyond faces:

AI mapping could greatly improve stuff like generating 3D models from photogrammetry. Take a video walk around of your room, car, or shed and let AI assist on building the model

VR currently relies on either bouncing fast-moving invisible beams of light off sensors on a headset/controllers, or recognitions of controllers/hands from optics on a headset. These tend to suffer from a lack of identification on other extremities such as legs (hence Facebook’s legless avatars) Kinect worked by identifying humanoid objects with visible light+IR and mapping out a body including leg movement etc, but suffered from lag on fast movement. An AI could probably sort that out faster, reducing lag for smoother full-body movement via camera, and leave the headset to just be a visor or high-res viewing device, maybe with a small eye-tracking camera.

All of these could be done local-only, without needing cloud or LLM’s, enough improves security+privacy and removes dependence on somebody else’s system in the cloud (which may not exist or cost the same in the future)