Help me with my master's thesis /sci/. Don't steal or if you do please cite "that sci schizo, 2021"
I may have found a way to AGI. Does any of you know what the free energy principle is?
Wouldn't it be possible to create a reinforcement learning agent whose loss function is just free energy (in a shitty nutshell, the entropy of the thing you're learning) and have it minimize it?
As an example, picture an AI playing tetris. If the agent's goal is to minimize entropy, the state of the board with the least entropy is an empty board, so i believe that if deployed the agent would literally start playing tetris, without ever having being told to, or having received any tetris-specific rewards.
My thesis is that it should be possible to express any environment as an entropy minimization task, and thus an agent able to learn to minimize that quantity would be generally intelligent, as it could (potentially) perform any task.
I may have found a way to AGI. Does any of you know what the free energy principle is?
Wouldn't it be possible to create a reinforcement learning agent whose loss function is just free energy (in a shitty nutshell, the entropy of the thing you're learning) and have it minimize it?
As an example, picture an AI playing tetris. If the agent's goal is to minimize entropy, the state of the board with the least entropy is an empty board, so i believe that if deployed the agent would literally start playing tetris, without ever having being told to, or having received any tetris-specific rewards.
My thesis is that it should be possible to express any environment as an entropy minimization task, and thus an agent able to learn to minimize that quantity would be generally intelligent, as it could (potentially) perform any task.
