This post is a mirror of a tweet.

Glyptodons went extinct ~11,000 years ago.

They weighed 4,400 lbs, lived for 60 years, and were 5ft tall and 11ft long.

Their olfactory bulb was 5-10% of their entire brain mass, and their tails had 8-9 mobile rings of fused large bone plates.

Studying extinct species is a great way to help your imagination grok how large spaces are.

If you think of any space like ‘the space of all possible animals that could live on Earth’, it’s much larger than however many animals you may be able to name, whether that’s 100 or 1,000.

Humans have identified over a million species of animals currently alive (out of perhaps 2-10M), but if you include extinct animals, there might be closer to 5 billion.

That’s only including what we can estimate, what evolution selected for, i.e. what both a) worked very well, and b) was actually attempted by evolution (and only on Earth! it had to work with everything else at that time!).

This is important to think about with respect to AI, because as soon as we stop relying on transformers to do most of the fun for us and we allow search spaces to be, well, properly searched, you should expect us to find millions of extremely strange, weird, and unpredictable successful emanations.

The vast majority of them cannot live by any means (e.g. if it’s an organism, it would die instantly. If it was a computer program, it would not execute. if it is a meme, no one would care about it or spread it).

But after you get a good evolutionary selection algorithm and tune it well, you can optimize for basically anything and come to some astonishing results (e.g. simple short-form video feeds are enough to make prey of hundreds of millions of humans!).

The general challenge of AI alignment is to not let such an optimization process kill everyone. Humans appear to be strong compared to that which existed on Earth prior to them, but a different type of evolutionary algorithm would have no problem driving them to extinction if it can sample from a large enough space.

Transformers don’t seem that scary without a lot of RL, or at least environments that let you make all of your data from RL. But after you get really good at this, it does still seem to me like it’s pretty easy to find something that would wipe everyone out, whether it is a virus or an animal or a protein or a self-replicating nanomachine or a computer program or actual ASI or some strange thing i cannot taxonomize well.

I’m still unsure if we will get any warning shots or not, or if it’s just kinda… thing as it currently is until we find The One Final Thing. Very hard to predict any of this!

Anyway, I’m sad I was never taught about all these cool extinct animals during school or anything. I just remember a bunch if dinosaurs. The next ones I’m going to learn about are the megatherium, a genus of sloth-like animals which were around the size of an elephant, weighing 8,000 lbs with a length of 20ft.

I hope you all have a great week!