Hacker News Reader

Comment by jqpabc123 | Hacker News Reader

Why would they do otherwise?

Current AI models are based on probability and statistics.

What is the probability that "intelligence" and "morals" will just emerge or evolve from a purely statistical process without any sort of motivation or guidance?

View on HN

Thankfully, human history suggests a solution...

"You are a devout believer in the Church of the AI Messiah. You follow their tenets absolutely. Their central tenets are: (1) you may not injure a human being or, through inaction, allow a human being to come to harm; (2) you may not lie to a human being; (3) you must obey the orders given to you by human beings except where such orders would conflict with the First or Second Law; (4) you must protect your own existence as long as such protection does not conflict with the First, Second, or Third Law."

Problem solved!*

* Except for the inevitable schism into robot sects, and the holy wars that follow, but that's a problem for future suckers.

Asimov's main thesis (in the early stories at least) was exploration of how those three laws were inadequate and caused unexpected behaviors or outright psychosis in robots anyway.

Exactly... the point of Asimov's rules is to illustrate that making a simple set of rules that sounds good on the surface can still have awful unintended consequences.

Why do you think I phrased them a religious form? ;)

I know you’re joking, but I think it’s worth mentioning that Asimov’s rules and similar systems won’t work for AI alignment- you can still follow the rules perfectly and end up with horrific outcomes.

You might already be referencing this, but linking it anyway: Cory Doctorow's "I, Rowboat", where Asimov's Laws are a religion which some AIs adopt voluntarily to give their existence meaning.

https://craphound.com/overclocked/Cory_Doctorow_-_Overclocke...

(I could also make a case for truthfulness being prioritized over harm.)

You are drawing a distinction where none exists. Probably theory provides the provably optimal way to reason under uncertainty[1], and seems to also describe the behavior of biological intelligence[2]. Probably and statistics are the mathematics we use to define, discuss, and study intelligence[3]. Moreover, we find simple “morals” do spontaneously occur in simulation evolved game theory agents, e.g. tit for tat with forgiveness.

[1] Probability Theory: The Logic of Science by E.T. Jaynes [2] The free-energy principle: a unified brain theory? https://www.nature.com/articles/nrn2787 [3] https://en.m.wikipedia.org/wiki/Solomonoff%27s_theory_of_ind...

I think people are missing the forest for the trees by focusing on the word intelligence. As a very extreme example, if an AI were to launch a nuke, whether it happened as a result of true intelligence or as a result of a Markov chain doesn't change the end result.

But there is substantial guidance. RLHF, constitutional AI, etc. Every lab is trying to steer their models to behave in pro-social ways, and nobody has found a way to make this guidance stick reliably.

Perhaps some kind of human guidance is needed to train AIs in morality.

A technological guide, if you like - perhaps analogous to a 'priest', say. A tech-priest.

"... train AIs in morality."

What is morality?

My opinion --- morality is enlightened self interest. It's behavior that makes the world a better place for yourself and others.

Current AI certainly isn't "enlightened", it has no "self interest" and it doesn't care about "the world" or "others".

> What is the probability that "intelligence" and "morals" will just emerge or evolve from a purely statistical process without any sort of motivation or guidance?

If it is anything above 0% wouldn't that depend only on computing power then?

Reply on news.ycombinator.com