Hacker News Reader

Comment by abduhl | Hacker News Reader

And you think that a driver commands their Tesla to run into an object when they enable autopilot or FSD?

One is a failure to act as an agent - to control the car and make decisions. Another is a failure to even be a reliable tool - to do what the operator commands. Very different.

Creating Agents is a lot harder.

Coming up with new words and terms in order to escape the comparison smacks of This Time It’s Different. Therac-25 was an Agent for performing radiation therapy - it controlled the radiation machine and made decisions. Autopilot/FSD is supposed to be a tool for the driver and it fails to be a reliable tool by driving into things.

This Time It’s Not Different At All

Autopilot is NOT advertised as being able to drive perfectly on its own, and your disingenuous analogies won't change that.

Agent - Something that makes decisions and actions that are not pre-programmed or predictable. Humans are definitely agents. LLM (eg Chat-GPT) are not quite there yet (they don't take action on their own).

Tool - Pretty much everything humanity has made to date. It does a preprogrammed thing or provides mechanical advantage. But it does not act on novel decisions it is making.

The terms come from an AI safety expert that Lex Friedman interviewed recently. [1]

Tesla FSD is almost an agent, though it asks a human to be ready to take over at any time.

[1] https://www.youtube.com/watch?v=NNr6gPelJ3E

Reply on news.ycombinator.com