To be clear tony stark was rich. So in the realm of that:
there's small bad holograms for cheap I think, probably with enough money you could make a built in holographic display somehow. It's just that no one usually does that because it's a waste of money and resources.
Deliver your drinks from the other room, I'm sure for rich people building a home with things like that in mind is possible. For the average household I'm not so sure. Robots can deliver food ok-ish (autonomous delivery robots) but I'm sure they need to be calibrated to the exact home layout. And anthropomorphic robots are turning out to be an incredibly inefficient design in general. Building the robot to serve it's specific purpose seems to be way more useful.
Thinking AI is kind of where we recently made the most progress and imo we're basically there. Deploy whisper for speech detection, maybe some cleanup in between, then an LLM like Qwen for thinking and run all your appliances as tool calls. The AI will automatically know how to handle it. I would say for simple things it'll probably get it right 95% of the time "lights on" "set an alarm" etc. That said, for more complex tasks like "plan a trip to the museum at x and share it with x contacts" these AI models are not there yet. They might get it right somewhere between 30%-50% of the time. AI models are great now, but they're not consistent enough and they are too resource hungry. Scientists are currently focusing on better compression so models get smaller and more people can just use them on their own hardware but it'll still be a while.
Conclusion:
We're at like 80%. They kind of work. Depending on the dedication, money and nature of the tasks.
A dedicated model that's fine tuned for specific tasks and then the temperature set to not allow variation is the key. If you want some variation, have a separate agent/model do that part, but some things you don't want creativity once you'd defined what the job is and have it working. I think the weak point in all that you mentioned is still the voice detection. It doesn't take much to break understanding of a clear command, much less some not said well or low or noise in the room. At best you have the system say it didn't get that and to try again. Don't let it try and figure out what you said, that will end badly at some point.
3 Comments
hoshikarakitaridia@lemmy.world · 4 pts · 2d
To be clear tony stark was rich. So in the realm of that:
there's small bad holograms for cheap I think, probably with enough money you could make a built in holographic display somehow. It's just that no one usually does that because it's a waste of money and resources.
Deliver your drinks from the other room, I'm sure for rich people building a home with things like that in mind is possible. For the average household I'm not so sure. Robots can deliver food ok-ish (autonomous delivery robots) but I'm sure they need to be calibrated to the exact home layout. And anthropomorphic robots are turning out to be an incredibly inefficient design in general. Building the robot to serve it's specific purpose seems to be way more useful.
Thinking AI is kind of where we recently made the most progress and imo we're basically there. Deploy whisper for speech detection, maybe some cleanup in between, then an LLM like Qwen for thinking and run all your appliances as tool calls. The AI will automatically know how to handle it. I would say for simple things it'll probably get it right 95% of the time "lights on" "set an alarm" etc. That said, for more complex tasks like "plan a trip to the museum at x and share it with x contacts" these AI models are not there yet. They might get it right somewhere between 30%-50% of the time. AI models are great now, but they're not consistent enough and they are too resource hungry. Scientists are currently focusing on better compression so models get smaller and more people can just use them on their own hardware but it'll still be a while.
Conclusion: We're at like 80%. They kind of work. Depending on the dedication, money and nature of the tasks.
Rhaedas@fedia.io · 0 pts · 2d
A dedicated model that's fine tuned for specific tasks and then the temperature set to not allow variation is the key. If you want some variation, have a separate agent/model do that part, but some things you don't want creativity once you'd defined what the job is and have it working. I think the weak point in all that you mentioned is still the voice detection. It doesn't take much to break understanding of a clear command, much less some not said well or low or noise in the room. At best you have the system say it didn't get that and to try again. Don't let it try and figure out what you said, that will end badly at some point.
MantisToboggon@lemmy.world · 0 pts · 2d
You don't want to know what the Tupac hologram had to do to pay his electric bill. I'll give you a hint. Butt stuff.