Request Response

This is how the first version of ChatGPT worked. A user sends a message (types in a text box), and the model sends a response right back. AI used in this way does feel like it has the capacity to be (or to mimic being) conscious - to "wake up", decide what to do, and take action. It just takes a message and gives a response.

Request Response with a loop in the middle

This is how the current generation of products work. When you send a request and you can see ChatGPT doing stuff like "Searching the internet" before it responds back to you. This is actually a loop - the agent starts with an input, decides to take an action, then re-examines where it is, decides to take another action etc. until it decides to exit the loop and return the answer to the user. It's important to understand that under this model, unlike the previous one, the agent does have to "decide" itself to exit the loop. If an agent were to have the same survival instincts as humans, it would run the loop perpetually to keep itself "alive" because the moment it exits the loop, it effectively dies and is completely dependent on a human calling it again to wake it back up.

Human Defined Heartbeat

For long running or ongoing tasks, agents need to wait for things to happen in the real world before they can continue their goals. These agents use a "heartbeat" which is basically just a prompt that runs on a schedule and says "Hey, wake up, here's all the information from your last heartbeat 30 minutes ago, see if anything has changed and decide what to do next, or do nothing, then exit again until the next heartbeat".

This is usually set up using something called a "cron", which is a feature that most operating systems have built in.

Agent Defined Heartbeat

This is where a human "kicks off the first run" with a goal, and before the agent exits, it creates a cron for itself to wake back up at some point in the future. In this scenario crucially, the agent has control over ensuring it doesn't die - it understands that if it doesn't set another wake time at some point in the future, that's it, it won't be able to continue to pursue its goal.