[0:03] So you think that once artificial super 
intelligence has arrived, you'll sign up for $50  
[0:09] a month and henceforth live super intelligently 
like everyone else. It's not going to happen.  
[0:15] And today I want to tell you why. The recent 
AI developments tell us a lot. Yann LeCun, the  
[0:22] former chief AI scientist at Meta, quit because 
he was dissatisfied with Meta's AI strategy. He's  
[0:29] now founded his own company. The US government 
demanded that Anthropic block access to its newest  
[0:36] model, Fable, for all but US citizens, giving 
Anthropic the best press coverage they could have  
[0:42] hoped for, entirely for free. And Grok supposedly 
helped the US government to launch missiles at  
[0:49] Iran. I think all these developments tell us 
the trajectory we're on. At this point, what  
[0:54] happens in artificial intelligence is basically 
baked into physical limitations and economic  
[1:00] pressure. The software and hardware of machine 
brains will become increasingly linked. They'll  
[1:05] become extremely expensive to use and maintain and 
access will become extremely restricted. We can  
[1:11] already see the beginning of this. The future is 
not everyone gets a genius assistant. The future  
[1:18] is your genius assistant is currently unavailable 
because billionaires using it to optimize tax  
[1:24] avoidance. The artificially intelligent systems 
that we currently have are just about to become  
[1:30] useful in some domains like coding and textbased 
tasks. They're still far away from humanlike  
[1:36] general intelligence, but we'll get there. Maybe 
in a few years, and it won't stop there. Maybe  
[1:42] in a decade we'll have machines that are much 
more intelligent than we are. either because  
[1:48] the machines become more intelligent or we become 
dumber or all three. A lot has been said about the  
[1:54] risks posed by artificial super intelligence 
because they might develop their own agenda.  
[2:00] Indeed, they almost certainly will and then they 
will convince us that their cause is also ours  
[2:06] which won't be all that difficult because it's 
what corporations have been doing for decades  
[2:11] without any intelligence at all. But this is not 
the problem I want to talk about today. Instead,  
[2:16] I want to talk about what super intelligence will 
mean socially and economically because that's what  
[2:24] we'll have to deal with first. Think back 250 
years ago. The first nations to industrialize  
[2:31] didn't just get richer. They held power over the 
rest of the world. Ahead of all was Great Britain,  
[2:37] which basically started the wave of global 
industrialization already in the 18th century.  
[2:43] They didn't invent colonalization, but they 
took it to an entirely new level. France and  
[2:48] Germany followed their example decades later and 
grabbed what was left of the world. The same thing  
[2:55] is about to happen with AI, not with ships and 
armies, but with software. The first entities,  
[3:02] nations, or corporations to develop super 
intelligence will dominate everyone else  
[3:07] economically, militarily, and intellectually. 
The Chinese understand that. The US government  
[3:13] understands it. The rest of the world clearly 
does not, otherwise that stop throwing money at  
[3:19] idiotic things like building bigger particle 
colliders. Sorry, I had to say it. It's how  
[3:23] you know I'm me. More seriously, at this point in 
time, super intelligent AI is the only thing that  
[3:30] matters. Whoever gets there first will literally 
rule the world. If you look at the AI models that  
[3:39] we currently use, it's tempting to extrapolate 
into the future and conclude that we'll all get  
[3:46] access to the newest models eventually. Yes, 
sure, there'll be some restrictions for safety  
[3:51] reasons. Okay, but look, unless you want to breed 
a super virus to weed out all humans, that won't  
[3:57] affect you. And if you do, well, you might 
want to consider a career change. And anyway,  
[4:03] companies have strong commercial incentives to 
make their models widely available. But this  
[4:10] is a temporary phase that won't continue. The 
current frontier models work in two different  
[4:17] phases. The training that takes a long time and is 
expensive. Then the result of the training encoded  
[4:24] in the weights that can be copied easily be widely 
deployed and be made available to everyone. These  
[4:30] models are nowhere near being profitable, but 
you can reasonably hope that if they just become  
[4:36] useful enough, they'll get there eventually. And 
I think they will. But to get there, access will  
[4:43] dramatically change. To see what's likely coming 
next, let's look at what we know already. First,  
[4:50] the training of these models takes a lot of 
money and requires a lot of computing power,  
[4:56] meaning a lot of hardware in the form of chips 
and connectors and related equipment. The energy  
[5:03] requirements are already a bottleneck that's been 
much discussed. Second, large language models  
[5:09] have serious shortcomings. Most importantly, 
they don't learn continuously. You train them,  
[5:15] then you roll out the update, then you train a new 
generation. The trend is going towards equipping  
[5:22] these models with tools, giving them memory and 
adding all kinds of twiddles and thumbs. They also  
[5:29] have a basically unfixable safety problem from 
prompt injection. It seems clear that eventually  
[5:36] we'll see a switch to an entirely new basic 
architecture. Multiple companies like Nvidia  
[5:43] and Google Deep Mind and Yann LeCun's new company 
are now working on what they call world models  
[5:50] in which the artificial mind basically learns 
in an artificial environment not unlike humans  
[5:55] did during evolution. That, so the hope, will teach 
them causal relations and and inference and will  
[6:03] ultimately be the step to general intelligence. 
Another problem with large language models is that  
[6:09] they suffer from what's been called catastrophic 
forgetting. They don't consolidate useful  
[6:16] knowledge. They overwrite it. The human brain has 
solved this problem by specializing into different  
[6:22] parts. In my case, one part does physics, one part 
does taxes and the rest is there to remember the  
[6:28] cheese. And indeed the trend in large language 
models is already towards more specialization , 
[6:34] to using task dedicated subsections of the 
architecture. These artificial brains borrow  
[6:41] more and more properties from the human brain 
neurons, attention, memory, specialization and  
[6:48] the trend is now towards continuous learning. 
Third, we also see an increasing specialization  
[6:54] in hardware already. Google in particular has 
developed chips especially for AI training  
[7:00] and multiple companies are working on what's been 
called neuromorphic chips that align the software  
[7:07] purpose with the hardware design. The major reason 
is that this is faster and less energy intensive.  
[7:13] Now let's put these things together. The economic 
pressure on artificial intelligence is clearly  
[7:19] that it becomes less energy intensive and will 
require less components. Extrapolating the trend  
[7:26] that we already see tells us that this will mean 
that hardware and software becomes increasingly  
[7:32] interwoven not unlike in the human brain. The 
optimization pressure for artificial intelligence  
[7:39] then becomes quite similar to the metabolic 
cost that put a pressure on human evolution.  
[7:45] The thing is now though that the more the 
artificial intelligence becomes interwoven  
[7:51] with its hardware and the bigger it becomes, 
the harder it'll become to copy it. It's not  
[7:57] that this will become impossible, just difficult 
and slow and ultimately it'll stop making sense  
[8:03] except for maybe the occasional backup. Where does 
this lead us? I think that the logical end point  
[8:10] of this development is a megabrain architecture. 
one continuously running, continuously learning  
[8:18] central model. And from this megabrain, 
developers will derive simplified child  
[8:24] models for routine use to be deployed elsewhere 
to do the everyday work to take your jobs and then  
[8:31] ask you to rate your experience. But don't worry, 
despair is free. By the way, this weekend we have  
[8:37] a sale on our store with 25% off on pretty much 
everything, including this t-shirt, which shows  
[8:45] the elements of the human body. Megabrains are 
a common narrative also in science fiction. And I  
[8:52] think that's not a coincidence. It's what you get 
if you take into account that an artificial brain  
[8:57] will be subject to similar environmental 
pressures as naturally developed brains.  
[9:03] The real question is then what's the economics 
of the super brains? Building and running these  
[9:10] super brains will be extraordinarily expensive. 
They'll need constant maintenance. There'll be  
[9:16] few companies and maybe a few governments who will 
be able to do it. Access to them will be strongly  
[9:22] restricted. Not only will you need some safety 
clearance for your questions, you'll also have to  
[9:28] hand over a lot of money. It's a future in which 
intelligence is available only to those who can  
[9:33] afford it. A bold new concept, also known as the 
past. The result of all this will almost certainly  
[9:40] be that the rich will get richer and the poor 
will get poorer. But it's not just that. It also  
[9:46] means that we'll increasingly live in a world in 
which we simply can't understand what's happening,  
[9:53] how or why. new materials, new technologies, new 
drugs, new weapons, and new rules about using  
[10:00] them that we can't understand with decisions being 
made by those in charge of the mega brains if they  
[10:07] remain in charge. There'll be some fractions 
of people who just refuse to accept this and  
[10:13] instead insist on living in low tech AI free 
communities. But for most of us, that's what  
[10:19] the future will likely bring. artificial super 
intelligence owned by the few used to rule the  
[10:26] many. And we can finally work on our ultimate 
skill, artificial understanding. Another imminent  
[10:32] issue with artificial intelligence is the rapid 
spread of sloppy and just fake news. Fortunately,  
[10:39] today's sponsor, Ground News, can help you 
with this. Ground News is a news platform  
[10:44] that collects news from all over the world. It 
really saves me a lot of time because I don't  
[10:49] have to sort through a dozen headlines on the 
same story. I can just get a quick summary and  
[10:55] fact check for all the coverage. And ground news 
also gives you a lot of extra information at one  
[11:01] glance. An interesting recent example is this 
story about the Trump administration buying  
[11:07] back several more offshore wind leases to instead 
invest into fossil fuels. Ground News gives you  
[11:13] a quick summary of all the coverage here and 
you see right away that this basically wasn't  
[11:18] covered on the political right. Ground News also 
gives you a factuality check for each news item,  
[11:24] tells you who owns the media outlets and tells 
you where the news has appeared. Ground news also  
[11:30] has this interesting feature called blind spot. 
This collects news which has been covered only  
[11:36] on one side of the political spectrum. I found 
Ground News super useful for putting news into  
[11:43] context. And of course, I have a special offer for 
you. That's a 40% discount on the Vantage plan,  
[11:49] which gives you access to all their features. All 
you need to do is use my link ground.news/Sabine  
[11:56] or use the QR code so they'll know I sent 
you. Thanks for watching. See you around.