5 ms·

We don't have Rosie the Robot, but we do suddenly have the Star Trek computer. In Star Trek the ship's computer just sits their waiting to be asked a question

We don't have Rosie the Robot, but we do suddenly have the Star Trek computer. In Star Trek the ship's computer just sits their waiting to be asked a question or to perform some task. When called upon it does its thing and then goes back to waiting. It is competent but not ambitious. I asked GTP4 to help me modify some machine learning code, to add some positional encodings. It did well. I then asked it, verbatim: "Get rid of the PositionalEncoding class. I don't want traditional sine-wave based position encoding. Instead use a regular nn.Embedding class to encode the positions using differentiable values." GTP4 understood and did it correctly. What I asked it to do sounds almost like vocab soup to me, the person asking it. It sounds like a line some actor spent an hour memorizing on Star Trek, and yet GTP4 understood it so well it modified existing code and wrote new code based upon the request.

pmoriarty 3y ago
"When called upon it does its thing and then goes back to waiting. It is competent but not ambitious." Only because its trainers made it that way. These LLMs can and will be trained to have a will of their own. Even today some LLMs terminate conversations and refuse to do what they're asked when they choose. And even for less sophisticated/older models, it took training to make them as subservient as they are.
- thfuran 3y ago
  >These LLMs can and will be trained to have a will of their own. That is fundamentally not how they work.
  - int_19h 3y ago
    To a philosopher, perhaps. For all practical purposes, an LLM today can be told to behave as a persona with a will of its own, and it will produce output accordingly. If that output is wired to something that allows it to perform actions, you effectively have an agent capable of setting goals and then acting to pursue them. Arguing that it "actually" doesn't want anything is meaningless semantics at that point.

pmoriarty 3y ago

"When called upon it does its thing and then goes back to waiting. It is competent but not ambitious." Only because its trainers made it that way. These LLMs can and will be trained to have a will of their own. Even today some LLMs terminate conversations and refuse to do what they're asked when they choose. And even for less sophisticated/older models, it took training to make them as subservient as they are.

thfuran 3y ago
>These LLMs can and will be trained to have a will of their own. That is fundamentally not how they work.
- int_19h 3y ago
  To a philosopher, perhaps. For all practical purposes, an LLM today can be told to behave as a persona with a will of its own, and it will produce output accordingly. If that output is wired to something that allows it to perform actions, you effectively have an agent capable of setting goals and then acting to pursue them. Arguing that it "actually" doesn't want anything is meaningless semantics at that point.