Agreed point 2 seems to be the crux.

An important concept behind this is Omohundro's Basic Drives. Any maximizing agent with a goal will try to acquire more resources, resist being shut off, resist having its goal changed, create copies of itself, and improve its algorithm. If it is possible to maximize its goal in a way that will not guarantee humanity's flourishing, then we all die, guaranteed.

If you want something more spelled out, I'll refer to Tim Urban's explanation. It's quite long, but is about as detailed as you can ask for. https://waitbutwhy.com/2015/01/artificial-intelligence-revol...

Yes point 2 is the weak one.

Aren't humans/biological life a counterexample? Simple bacteria are clearly maximizing agents, and cyanobacteria did infact almost destroy all life on earth by filling the entire planet with toxic oxygen.

We've known how to improve ourselves via selective breeding yet vanishingly few humans are proponents. We and and other intelligent animals have a wide variety of goals, and even animals sometimes share food across species.

The evidence just doesn't seem to support this concept of Basic Drives. If anything, the evidence (and common sense) seems to suggest that the more intelligent the organism, the more easily and more often it ignores its basic drives.

Natural selection and biology have different constraints that limit things (and even then often selection can cause extinction)

Intellectual capability and alignment are orthogonal - the former doesn’t get you the latter for free.

Most people new to this issue don’t intuitively grasp this at first.

What evidence do you have that intellectual ability and alignment are orthogonal?

Remember that ML models are also subject to selection. If they're not useful/beneficial to us, we stop running them.

For clarity I think EY has made that case well, I think it's the heuristic Cohen and Pinker are relying on that's right most of the time but can still be catastrophically wrong (everything so far has not lead to human extinction, but that was also the case for all animals that went extinct prior to their extinction event).

Some others in this thread have linked to stuff (it's variations/examples on the paperclip maximizer argument).

I'd be curious why John Carmack thinks these risks are unlikely (he thinks fast take off is not something to worry about) - is that because he thinks we'll get some sort of trainable AGI first or something else? There are also some other substantive disagreements here: https://www.lesswrong.com/posts/wAczufCpMdaamF9fy/my-objecti...

John Carmack mentions this in the Lex Fridman Podcast, basically the argument is that AGI performance characteristics will be similar to that of LLM (which is another way of saying that AGI will NOT be hyper efficient, P != NP), and that the performance characteristics pose a problem for fast takeoff. The bottleneck to performance in training these models is GPU memory bandwidth when the entire model fits inside VRAM, which on modern cards is 1TB/s, when the model cannot fit in GPU memory, the performance now instead scales along PCIE, which is currently 32GB/s. AGI (or an LLM) attempting to replicate across mobile devices, or desktops, will be severely hamstrung by the network connection. So using all of planet Earth's computing resources is not necessarily better than say a single data center in this respect. The second piece of the argument here is essentially saying that a data center is also not enough, or rather, not enough in a detectable timeframe. Could AGI hack the entire data center for an entire month to perform it's training (and then execute it's strategy enough to gain say nuclear codes?). Unlikely. Is say 8 hours of training using an entire data center enough to go from intelligence to super intelligence? Intuition says no. I expanded a bit upon what I believe his argument to be, I definitely recommend watching that part of the podcast episode.

Here's my take on the implications of not having fast takeoff, a secretly antagonistic AGI will be constrained to cooperation, and slowly leeching compute resources towards its goal until it is able to aquire enough resources to confidently sprint towards the inflection point. This could mean teaching us how to build better semiconductor foundries, how to create nuclear fusion energy sources, how to educate our youth to better fit into those jobs, slightly alter our culture's value systems over time via astroturfing to be more sympathetic, cooperative, trusting, and less vigilant towards these AI systems, how to build more robust computing systems so we stop needing to detect when something goes wrong (because it already has 99.9999% uptime).

I believe slow takeoff is actually worse in some sense, because it is a hell of a lot harder to detect, and humans tend to get complacent and apathetic when something "just works" for decades, even if it has been plotting since the beginning.

Thanks - I saw he was on there, but I find lex insufferable as an interviewer so tend to avoid the pods (despite the great guests)

Who said it followed? That's a strawman. To dumb it down even further for you, I guess...

There are brains, therefore there can be created an artificial super-brain. Artificial super-brain might have goals which don't align with human brain and we will have no way to understand or control the situation.

Two unrelated facts, which together mean that we should be careful with experimenting with the science working towards super-brains. Just a single super-brain could end us.