Comment by upwardbound | Hacker News Reader

Comment by upwardbound | Hacker News Reader

For those new to the field of AGI safety: this is an implementation of Paul Christiano's alignment procedure proposal called Iterated Amplification from 6 years ago. https://www.alignmentforum.org/s/EmDuGeRw749sD3GKd

According to his website he previously ran the language model alignment team at OpenAI. https://paulfchristiano.com/

It's wonderful to see his idea coming to fruition! I'm honestly a bit skeptical of the idea myself (it's like proposing to stabilize the stack of "turtles all the way down" by adding more turtles - as is insightfully pointed out in this other comment https://news.ycombinator.com/item?id=40817017) but every innovative idea is worth a try, in a field as time-critical and urgent as AGI safety.

For a good summary of technical approaches to AGI safety, start with the Future of Life Institute AI Alignment Podcast, especially these two episodes which serve as an overview of the field:

- https://futureoflife.org/podcast/an-overview-of-technical-ai...

- https://futureoflife.org/podcast/an-overview-of-technical-ai...

In both of those episodes, Cristiano's publication series on Iterated Amplification is link #3 in the list of recommended reading.

> AGI safety

I genuinely laughed. Oh no somebody please save me from a chatbot that's hallucinating half the time!

Joke aside, of course OpenAI is gonna play up how "intelligent" its models are. But it's evident that there's only so much data and compute that you can throw at a machine to make it smart.

Thanks, added to my collection of AGI-pessimistic comments that I encounter here, and that I aim to revisit in, say, 20 years. I'm not sure I will be able to say: "you were wrong!". But I do expect so.

You make a good point so I should clarify that Iterated Amplification is an idea that was first proposed as a technique for AGI safety, but happens to also be applicable to LLM safety. I study AGI safety which is why I recognized the technique.

Covid isn't what most people would call "high intelligence", yet it's a danger because it's heavily optimised for goals that are not our own.

Other people using half-baked AI can still kill you, and that doesn't have to be a chatbot as we have current examples from self-driving cars that drive themselves dangerously, and historical examples of the NATO early warning radars giving a false alarm from the moon and the soviet early warning satellites giving false alarms from reflected sunlight, but it can also be a chatbot — there are many ways that this can be deadly if you don't know better: https://news.ycombinator.com/item?id=40724283

Every software bug is an example of a computer doing exactly what it was told to do, instead of what we meant.

AI safety is about bridging the gap between optimising for what we said vs. what we meant, in a less risky manner than if covid — and while I think it doesn't matter much if covid did or didn't come from a lab leak (the potential that it did means there's an opportunity to improve bio safety there as well as in wet markets), every AI you can use is essentially a continuous supply of the mystery magic box before we know what the word "safe" even means in this context.

> Every software bug is an example of a computer doing exactly what it was told to do, instead of what we meant.

That is only as long as the person describing the behaviour as a bug is aligned with the programmer. Most of the time this is the case, but not always. For example a malicious programmer intentionally inserting a bug does in fact mean for the program to have that behavior.

Sure, but I don't think that matters as the ability to know what we are even really asking for is not as well understood for AI as for formal languages — AI can be used for subterfuge etc., but right now it's still somewhat like this old comic from 2017: https://xkcd.com/1838/

I am not sure why you didnt see the citations to 3 different papers from Cristiano. A simple search in the linked PDF suffices: citations 12, 19, and 31.

Thank you for providing the reference numbers, this is helpful! I'll update my GP comment.

> it's like proposing to stabilize the stack of "turtles all the way down" by adding more turtles

That's completely fine. Say each layer uses the same amount of turtles, but half the size of the layer above. Even allowing for arbitrarily small turtles, the total height of the infinite stack will be just 2x the height of the first layer.

Point being, some series converge to a finite result, including some defined recursively. And in practice, we can usually cut the recursion after first couple steps, as the infinitely long remainder has negligible impact on the final result.

You seem to have interpreted the analogy as meaning "you might run out of turtles" instead of something like: "stacks of turtles aren't stable without stable beneath them, no matter how many turtles you use."

Space them out. I meant that infinite stack of turtles can be stable, of finite height, and for practical purposes, cut off after few layers without noticeable impact on stability.

Reply on news.ycombinator.com