AI understands or just follows the rules?

Does an AI really understand your job, or does it just follow a rulebook?

In 1980, John Searle imagined a man in a room who knows no Chinese. He gets Chinese symbols, follows a rulebook, and sends back perfect answers. People outside think he understands. In reality He doesn’t.

I used to think that’s all AI is. Then I heard Geoffrey Hinton argue that modern AI isn’t just guessing the next word. In building that guess, it forms a model of the world, and he says that is what understanding is. He argued that the current AI understands like a human and is closer to us then machine code.

So I’ve been wondering about ICM (organising a job into folders, stages, and rules, judgement calls, opinions written down on how the work should be done). Are we just handing it a better rulebook? Or are we giving the AI a real picture of the job, with a human checking each stage?

I lean toward the second. What's your opinion?

Just a rulebook. The AI follows steps without understanding.

Description of the job. The AI works out the how, like a person would.

Both. The rules keep it safe, the description helps it think and understand, and the human reviews.

11 votes

20 comments

AI understands or just follows the rules?