⁠It Begins: AI Is Now Improving Itself

In one year, the smartest AI jumped

from 96 to 136 IQ on the Mensa Norway test In other words AI went from the intelligence of an average person to nearly the IQ of a genius This has AI scientists terrified And the godfather of AI and

Nobel laureate But why exactly are scientists so worried about AI getting smarter? So to understand how close

we really are to that outcome we're breaking down Leopold

Aschenbrenner's famous report situational awareness, which is

circulating among top US government officials We would rapidly go from human level

to vastly superhuman AI systems Most people think the atomic

bomb is the most gameboard-flipping invention in the history of war But arguably One atomic bomb could wipe out a city,

but one hydrogen bomb could wipe out a country A single bomb with more explosive power than

all the bombs dropped in the entirety of World War II combined That's the scale of the difference we're talking

about between AGI or artificial general intelligence and superintelligence But what does this difference mean in

real life, and how dangerous for humans could this really be? Imagine this Each vastly more intelligent than many humans, capable of novel,

creative, and incomprehensibly complex behavior Their power wouldn't stop at machine learning Superintelligence would ignite explosive

progress across every field Robotics, biotech, weapon systems,

material science I mean, what took humans decades,

they'd solve in days An industrial explosion would follow, then a military one It would offer godlike power both

to create and to destroy And once it begins, we will be living

through the most volatile moment in the history of our species So how close are we to this future? Well, Aschenbrenner lays out four

critical steps to get to superintelligence The first is getting to AGI, which we previously covered The next steps are all about using AI

to automate AI research itself And here's the problem Former Google CEO Eric Schmidt Warned Isn't just a theory It's already happening A scientist at Google DeepMind reveals But first things first, we have to reach

AGI and then we can worry about all the other steps, right? The problem is AGI is much closer than we think The three major heads of AI labs all

think this is going to happen in the next two to five years And everyone's timelines for AGI

are shrinking, even famous AI Skeptics like Yann LeCun Besides, automating AI research

isn't actually as hard as it seems To get there, you don't need the

AI to become genius level in every field Imagine trying to create an expert biologist AI

that can research new cures for diseases, you don't start

by training it on biology You only need to automate AI research so that the AI itself can

think faster and become smarter Then once it is much smarter it could easily and quickly learn

advanced biology, or any topic for that matter. And we've already seen leaps like this happen before No education of like the history of

human chess playing 10 hours a day for 30, 40 years, being the

greatest human in the world at it And then something can come in and Kasparov, after the move C4 has resigned Right now, this is happening at an insane

pace in the field of robotics A year ago, humanoid robots weren't

advanced enough to even walk smoothly But a tactic called reinforcement learning changed everything Robots can now learn not just to walk, but to do side flips, kip-ups, and even kung fu Completely on their own Boston Dynamics Atlas' learned to run and break

dance by simulating running Think of implications That's how Neo in the Matrix was able to learn kung fu in the blink of an eye I know Kung Fu Do you realize just how suddenly our little chatbots might go from blinking cursors to very real things? Some people assume that we'll see

robots in the real world slowly getting better over years or decades But if they can learn 10,000 times faster via

simulation, they could blow our expectations out of the water So there's a very real chance we'll see

humanoid robots casually walking around pretty much everywhere

in the next five years And there are many more examples

of these "fast takeoffs" using reinforcement learning Minecraft, Shogi, Go, chess, etc But even if we don't get a fast takeoff We're in for a wild ride Remember, Aschenbrenner argues

that we don't need to automate robotics research to get to superintelligence implement experiments to test those ideas,

and interpret the results and repeat Where simple extrapolations of current AI capabilities could easily take us to or That's why AI being able to research

better algorithms on its own is such a big deal Sit with that for a second That's like if a $50,000 Tesla

dropped to just $50 and ran better And that's not just about saving money When we have more compute, we can

use it to make the models smarter and cheaper and the cycle repeats But once we do reach AGI, we won't just have one of its kind Aschenbrenner and

other researchers used Perhaps 100 million A single H100, one of today's top AI chips, has about the same raw computing power as a human brain And we're on track to have 100 million H100 equivalents worldwide in just a few years That's potentially 100 million AI workers right there But the thing is Gemini 1.5 Flash is 10 times faster than the originally released GPT-4 merely while providing similar performance The automated AI researchers will be able to find similar wins very quickly They'll each be able to do a year's worth of work in just a few days This is why Geoffrey Hinton is terrified The final step to reach superintelligence

will happen when AI becomes superhuman not only in coding, but in They'll become qualitatively better than

the best human researchers because Learn in parallel from each of their copies and rapidly accumulate

the equivalent of millennia of experience Geoffrey Hinton is also worried that And here's why this is just so

dangerous for the human race Humans didn't out-compete the

Neanderthals because we were smarter We were just better at sharing knowledge

culturally and learning from the mistakes of other humans Now imagine millions of AIs, all thinking

100 times faster than humans, and instantly sharing new

learnings with each other They'll speed run a thousand years of

their own cultural evolution in what feels to us like a single year Okay, but is this really the future that's waiting for us? Or just a cool sci-fi scenario? Let's take a look at the limitations

that could slow down the intelligence explosion So this is a series of questions

every major AI lab has been asking about how likely And how the majority are interpreting the evidence To make algorithmic progress, AI researchers have to run experiments and that takes serious computing power You simply can't run machine learning

experiments if you don't have enough chips to run them on which is why the CEO of Deepseek said that But here's the thing AI researchers working at 100 times human

speed are far more likely to design more efficient experiments in the same amount of time compared to human researchers So they won't need to waste as much

compute as the inefficient humans And as algorithmic progress continues those same experiments also get

cheaper and more effective Remember Maybe the law of diminishing returns applies here And there will be a point where developing more intelligence will eventually lead to smaller and smaller increases is inefficiency until we hit a wall But that seems pretty unlikely because

AI is still a young field and we haven't thrown nearly as much

of humanity's brain power at it as we have at other fields, like math or physics For almost all of human history, progress

was flat then suddenly it explodes What does it feel like to stand here? And the intelligence explosion won't

just stay in the lab Once we have superintelligence, it will spill out into everything This would massively accelerate scientific progress Imagine going from the Wright brothers to

fighter jets to the moon and to ICBMs in just a few years That's just how Aschenbrenner sees the 2030s going an industrial and economic explosion

driven by an alien intelligence Physical labor could be automated with

robot factories that cover the entire Nevada desert And all mental work could be automated by

data centers filled with AI, all thinking faster than any human ever could just by using standard economists' models of economic growth Yes, existing future red tape

would slow things down For example, governments could force lawyers

and doctors to be human by law So, the intelligence explosion will likely be very uneven Some fields would grow exponentially,

while others, like the long shoremen port workers where automation

can literally be made illegal, would say the same But none of these changes comes close to what superintelligence means for military and governments Imagine mosquito-sized drone swarms,

novel WMDs, things we can't yet fathom It'll be like 21st century militaries

fighting 19th century brigades of horses and bayonets Even without robots, the small civilization of superintelligences would be able to hack any undefended military, election, television, etc and

also cunningly persuade generals. They could design new synthetic bio-weapons and then

pay a human in Bitcoin to synthesize them and so on Conquered the Aztec* Empire of 10 million They didn't have godlike power, but the

old world's technological edge and an advantage in strategic and diplomatic cunning

led it to an utterly decisive advantage Superintelligence is the ultimate gameboard flipper Once it arrives, It becomes almost

impossible to predict what will happen next The shift will be fast. Overwhelmingly fast The world won't have time to process

what's happening before everything starts to go insane Think about all the crazy things

that happened in the 20th century Now imagine all that compressed into just a few years Maybe this sounds like sci-fi to you Super Intelligence is the stated goal of the leading scientists at the AI labs And the heads of the top three

AI labs all think this is likely coming in the next two to five years Thousands of academics with

no financial stake in the AI industry signed a letter stating that AI

poses a risk of human extinction Geoffrey Hinton, who quit his

high-paying job at Google to warn the world, put our chances of going extinct at 50% The intelligence explosion is an

imminent possibility, and we need to prepare for now If you want to see a detailed

concrete scenario of how this superintelligence

might take over, check out this video next This is a small channel, and this video

took a long time to make If you made it this far, thanks for

watching I appreciate your support