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