Gwern — AI Safety↗

Autonomous AI systems (Agent AIs) trained using reinforcement learning can do harm when they take wrong actions, especially superintelligent Agent AIs.

A time dilation tool from an anime is discussed for its practical use on Earth; there seem surprisingly few uses and none that will change the world, due to the severe penalties humans would incur...

Or to put it perhaps more clearly, for a fixed amount of computation, at each greater level of intelligence, a smaller increase in intelligence can be realized with that amount of computation.

Pain is a curious thing. Why do we have painful pain instead of just a more neutral painless pain, when it can backfire so easily as chronic pain, among other problems?

GPT-3, announced by OpenAI in May 2020, is the largest neural network ever trained, by over an order of magnitude.

By this point in the run, it’s 3AM Pacific Time and no one is watching the TensorBoard logs when HQU suddenly groks a set of tasks (despite having zero training loss on them), undergoing a phase...

Heather Murphy, “Why Stanford Researchers Tried to Create a ‘Gaydar’ Machine” (NYT), 2017-10-09: So What Did the Machines See? Dr. Kosinski and Mr.