during my time as a phd i did machine learning research, primarily in cybersecurity and resource allocation for autonomous vehicles. i also did some education in cryptography and cryptocurrency

cryptocoins:

the thing with cryptocurrencies is that they are useless. "snake oil" is a fantastic analogy: (1) it has some rare applications of genuine use, and (2) it was almost definitely pushed by white scam artists under east asian monikers.

in the most generous interpretation, cryptocurrencies/blockchains have extremely niche, contrived use-cases that are almost always better solved by existing technologies

in the few cases cryptocurrencies offer a sensible solution, it's impossible for those to get adopted when 1% of the people in the field are trying to solve problems and 99% of the rest are scam artists who have no business trying to build a public-facing interface for, like, local notaries

ai:

"AI" might as well mean "computer". people have been surprised by what machines could do ever since we made them, and it was game over once we started calculating.

pre-2000s, we thought we could give computers a list of facts and a list of rules, and once we put enough in, we could ask the program to derive any new fact. this is what we used to call "AI", and is now called "symbolic AI". it never really worked though!

"machine learning" basically refers to probabilistic models and optimization methods that do away with statistical / formal rigor, with an emphasis on empirical results.

while symbolic AI was having a lot of difficulty, neural networks were having a worse time. tbh NNs were considered a joke for awhile, but then Gamers (and then cryptocoin people) made GPU manufacturers invest heavier into GPUs, and the neural networks people had a field day with it. someone writes a GPU driver, then suddenly neural networks were millions of times faster

(side note: neural networks are built on linear algebra running on GPUs, and everything is built on linear algebra. so, the gains for NNs were fantastic across all the non-NN machine learning.)

"deep learning" (machine learning with neural networks with many layers) became a big thing, and since 2012 there was a lot of focus on solving old problems with deep learning.

There's a lot of extremes that had hit the news, like "AI tentacles" and "AI protein folding" and "AI surveillance" and "AI recognition of ethnic minorities".

But there are also a lot of mundane problems, like "what's the ideal bus route and schedule for this campus?" or "what's the ideal allocation of city-provided bikes?" or "can we improve on this PID controller for weapons manufacturing"?

I can proudly say that a lot of my research was in the mundane and good applications without a lot of potential for dual-use (i.e. evil) advances.

But around when I started my PhD was around when a lot of "facial recognition of han vs uighur chinese" papers came out in. It took about two years for the genocide of Uyghur chinese people to become a mainstream concern (and it's still ongoing! it's AI powered! and nobody talks about it!)

But I want to say...

ai bros and anti-ai-bros frustrate me endlessly:

I constantly see people with no subject-area knowledge whatsoever make the worst arguments in the world. What's frustrating is that the Evil Potential for generative AI is so, so, so far beyond AI art that it seems wild that people are arguing over the definition of art while we have AI-powered ethnic minority organ harvesting programs that have been running for years.

if i have to hear someone say "AI art is/isn't art because <hilariously incorrect description of how the latest txt2img model works, that also has no bearing on the definition of art>" i'm going to kermit. they are 100% as dumb as the "AGI is just around the corner and will kill us all"

either way, AI is replacing programmers way, way before it's replacing artists

So, I want to say two things:

1. "AI" is not like cryptocurrencies, because AI is actually effective. the metric for "does AI have applications" is not how well you can use the latest model to replace a graphic designer

2. The ethics of "AI art" pale in comparison to the horrors of all the rest of AI-- both those ongoing and those yet to come.

side note, on AGI:

the lesswrong/EA "AGI is coming and will end us all" is kind of stupid. it's like saying "we'll soon make a bomb that can EXPLODE THE EARTH."

1. sure, we could probably make a bomb in the next 10 years that can explode the earth, but

2. the bombs we already have are scary enough and are wielded by irresponsible people that i'm not even afraid of the bombs that could feasibly exist.

that is to say, "AGI" (while super poorly defined) is feasible within our lifetimes, and it would be very very bad. Remember that "intelligence" means nothing: We just have very complicated machines that we don't understand, and they were built from trying to optimize a function that a scientist wrote trying to describe they values. An "AGI" would be a machine that tries to optimize some formal definition of utility, and it's very likely such a machine would be very very very bad for a lot of reasons. Unfortunately, a lot of them are basically the plot of whatever your favorite scifi book about evil AI is. (Even the stupid ones, like "evil robot takeover".)

But I'm worried about the torment nexus that we already have, not the ones that might happen in the wildest fantasies of the same people who are afraid of "Roko's Basilisk"

side side note, on roko's basilisk:

hahagagagagaghhhahahaha

hahahahahahahahah

jesus christ. you can safely ignore anybody who is afraid of roko's basilisk

the tldr: in this matter, the centrist-sounding take is basically correct. "AI is good and bad, but it isn't snake oil"

the main evil is still militarized capitalism, and the scariest thing is still that earth is becoming uninhabitable, and the second scariest thing is still that we have enough nukes that we can make it uninhabitable instantly