2 Comments
User's avatar
Bernd Nurnberger's avatar

Thank you for bringing your timeless insights to the current steep learning curve for society. The world keeps changing behind my back.

A few years ago when the 5 levels of autonomous driving were published, I mused in private, “why not give the robot driver the regular human driver’s license test? If the inspector deems it safe enough, sign it off for driving on public roads like any new driver. Insurance premiums start high, like for any new driver. Earn the no-claims discount like any new driver.”

Doubt remained: what if the maker changes the software, upgrades hardware, or the robot learns, when would a re-test be required? Now I think the safeguards in place for human driving proven over a century should be good enough, safe enough, to follow and enforce the same external ruleset for man and machine as a starting point. Society will take less than another century to establish additional safeguards, should the learning experience warrant this.

Expand full comment
Monica Anderson's avatar

Humans create limited safeguards against other humans. We tend to trust others because most humans are rather predictable and the odd lunatic is likely to turn up in some other town and be dealt with after "limited" damage. All safeguards are compromises between convenience and power.

We will start lowering the current hard and explicit safeguards over time simply because, from one point of view, "it adopted those safeguards as ethics". From a simpler point of view, they just became more competent in the art of conversation in a societal framework.

LLMs will be able to judge your competence from the dialog. If you are a HS student asking about how to make nitroglycerin, the LLM should explain to you the cost of losing some fingers. OTOH, if you are a chemical engineer engaged in optimizing the nitroglycerin yield of a process you can already describe in detail, sure, here's a couple charts.

If you can't handle the truth, you won't get it.

I discuss this in AI Alignment Is Trivial.

Expand full comment