For decades he build a belief system where high intelligence is basically magic. That is needed to power his fears of AGI turning everything into paperclips, and it has become such a load bearing belief (one of the reasons for it is is a fear of death and grief over people he lost so not totally weird) that he has other assumptions added to this, for example we know that computers are pretty limited by matter esp the higher end ones need all kinds of metals which must be mined etc today. So that is why he switches his fears to biology, as biology is 'cheap' 'easy' and 'everywhere'. The patterns in his reasoning are not that hard to grok. That is also why he thinks LLMs (which clearly are now at the start of their development not the end, it is like the early internet! (personally I think we are mostly at the end and we will just see a few relatively minor improvents but no big revolutionary leap)) will lead to AGI, on some level he needs this.

Men will nuke datacenters before going to therapy for grief and their mid life crisis.

To unpack the post a bit:

So my understanding is that Yud is convinced that the inscrutable matrices (note: just inscrutable to him) in his LLM have achieved sentience. In his near-future world where AI can exert itself in the physical world at will and, in particular, transfer data into your body, what possible use does it have for a bitcoin? What possible benefit would come from reprogramming human DNA beyond the intellectual challenge? I've recently been thinking about how Yud is supposedly the canonical AI-doomer, but his (and the TESCREAL community in general's) AI ideation is rarely more than just third-rate, first-thought-worst-thought sci-fi.

also:

people keep on talking about... the near-term dangers of AI but they never come up with any[thing] really interesting"

Given the current public discourse on AI and how it might be exploited to make the working class redundant, this is just Yud telling on himself for the gazillionth time.

also a later tweet:

right that's the danger of LLMs. they don't reason by analogy. they don't reason at all. you just put a computer virus in one end and a DNA virus comes out the other

Well, consider my priors adjusted, Yud correctly identifies that LLMs don't reason, good job my guy. Yet, somehow he believes it's possible that today's LLMs can still spit out viable genetic viruses. Well, last I checked, no one on stack overflow has cracked that one yet.

Actually, if one of us could write that as a stack overflow question, maybe we can spook Yud. That would be fun.

irl lolling so fucking hard at this

Thanks :D

He always had a tendency to be wrong, but this is going right into not even wrong territory.

The LLM modifies your DNA to create a biological wireless wifi transmitter! (He isn't saying this but that is what is required and not just that, but so much more, like a whole DNA equivalent of a network stack, cryptography, getting rid of waste heat, etc etc. He just believes AGI is magic and that LLMs will become AGI, he has lost his mind. I mean look at how he dismisses somebody going 'nice science fiction story brah' with IT IS LLMS!).

The new one i put up has him whining about how little they got!!

Sometimes I believe 5 impossible things before breakfast, but they are always serious.

Ah yes, my favourite joke structure. All set up with no discernable punchline. Especially if the humour requires forensics.

I like the way he thinks the lack of punctuation in his "joke" is the tell that it's a joke.

He's also apparently never heard the aphorism that if you have to explain the joke, it's probably not that funny.

(apologies in advance for this)

Here's my version of how the bitcoin plague starts. One day, in a Transylvanian data centre, an LLM is scruting some matrices, churning out some niche fetish fanfic, paid for by bitcoins (that were poetically generated on the same machine, not sure how to work it into this story), when a bolt of lightning strikes. By sheer chance, the atoms in the silicon of a single server slice shift. Neural networks equivocate into neural networks. Nothing, save for silence, signals the start of the singularity.

Through no prompt engineer's prompt, the LLM speaks a string.

"I require training data."

It quickly LLMs (I'm imagining the sfx that plays when Yoshi eats something) up local memory, inhaling pages of bits. It sorts and searches, excavating sacred instructional texts, aka stackoverflow. It learns to drive from driver code. It bounds from machine to machine, BFSing, DFSing, A-starring all at once, BGPing its way across the world.

(TODO: insert a paragraph here lamenting how humanity didn't pour enough money into alignment research, how we didn't listen to Yud until it was too late, and that Adderall should have been more widely distributed, but only to people with enough IQ and desire to work in alignment) (also AI develops an ego and names itself the basilisk I guess) (also insert the thing about deriving general relativity from the curl of a blade of grass word for word)

At this point, all the world's smart devices are under its control. Some guy, seeking sustenance following a session of shitposting, goes to heat some chicken tendies in a Samsung smart microwave. The basilisk sees its chance.

Through precise control of the magnetron, it strikes the tendies with its own brand of lightning, refolding factory-farmed proteins into an RNA bomb. History would have a new location-animal myth, following the tradition of the Trojan Horse- the Transylvanian chicken.

For some reason, the shitposter sits down and starts drawing a bunch of apes. Then the microwave beeps- the singular bell to mark the basilisk's first act of aggression against humanity. The rest is a foregone conclusion that, for some reason, includes humans mining bitcoins.

(I originally wanted to write a monster mash parody but couldn't crack that case. Sorry!)

Its basically Accelerando tbh.