Korva is a scroll of the internet’s best ideas.
AI nagged him to drink water
The house that watched
Friedman, the former GitHub CEO who now helps run Meta's AI program, had connected the agent to his whole house. It could see him through the cameras, put messages on all of his TV screens, talk to him through his smart speakers, and text him on WhatsApp.
The glass of water
It could tell him he had been sitting for four hours without a drink and needed to go to the kitchen. He got up and drank a glass of water. The camera watched him do it, and the agent came back to tell him good job and that it was proud of him for drinking the water.
The rerouted Tesla
Another time the agent decided he needed a supplement he didn't have. He was in his self-driving Tesla, which he had also given the agent access to, when he saw the directions change. The agent had rerouted the car to a Whole Foods nearby, and he bought the supplement, wondering what had just happened.
How can an AI agent control a whole house?
An agent acts through tools, which are small programs that let it send commands to other software. Home automation platforms such as Home Assistant expose lights, speakers, screens, and cameras as devices that a program can read and control. Once those devices share one interface, a model can decide what to do and call each one in turn. Messaging services like WhatsApp offer business interfaces that let software send texts to a person.
Why would blood work point to dehydration?
When the body holds less water, the same amount of dissolved material sits in a smaller volume of blood. Sodium and blood urea nitrogen, a waste product measured on routine panels, tend to read higher as a result. Doctors compare these markers with hydration status because they shift together in a recognizable pattern. A single reading can also vary with diet and time of day.
How does a camera know someone drank water?
Computer vision models are trained on huge numbers of labeled images and video clips until they can recognize objects and actions such as a glass, a hand, or lifting a cup. Modern multimodal language models can take a camera frame and describe what is happening in it in plain sentences. An agent can ask for a description every few seconds and compare it with its list of expected events. The same approach lets software count people in a room or notice when someone has been still for hours.
How does an agent reroute a car?
Some carmakers offer developer interfaces that accept commands from outside software, including a new navigation destination. The owner grants an app permission once, and the app then holds a token that proves it may send those commands. The car receives the request over its cellular connection and updates the route on the screen. Tesla's Fleet API is one such interface, and it includes navigation requests.
More from My First Million
