Just kidding. Honestly, the best ideas out there are low-tech. Anything where as you put it is "crowded" with tech people don't want, but is also of good build.

Or if you must add an LLM and a touch screen to everything, do it in a way that respects user consent and boundaries.

One thing I wanted to get into but can't due to time is sub-150ms responsive LLM for use with drones and bots. Faster response time than a human brain. Multiple LLMs running on dedicated hardware to preempt and predict response for various stimuli.

A drone you can't shoot-down with an arrow, and a robot you can't beat at dodgeball.

Now imagine the implications of that!? But the first problem I'd like to solve is remotely piloted work. if that LLM and response-time combo works, you would be able to safely work alongside a remotely piloted robot in a hazardous construction site better than you would with a human (it could even improve safety for humans). At all times it is getting audio/sonar, lidar, infrared and visual (image processing) input and several of these really fast LLMs (not just, LLMs aren't great at certain tasks, so naturally other types of ML and predefined routines would also be running) will be preparing if/else decision trees in case a potential condition is met. Sees a ladder just leaning there? what if it randomly slips, how would it react? plan is in place under 150ms while other plans are being formulated for other things as well.