You can buy Greg Johnson’s The Philosopher Is In here.
89 words / 1:54:20
Greg Johnson and David Zsutty on the AI Economic Apocalypse, Palantir suppressing the far Right, the Remigration Summit, Jared Taylor’s harassment in France, the slip and fall fraudulent lawsuit against Return to the Land, how White Nationalists should talk about Trump, Henry Nowak’s murder and the reaction, exiting the Iran Crisis, and more.
Now for your streaming or downloading pleasure. To listen in a player, click here or below. To download, right-click the link and click “save as.”
2 comments
“If it’s not radically new, why are they branding it as artificial intelligence”
When electronic computers first started becoming available in the 40s and 50s, they were manually wired and required large, bulky electronic components. Think cathode tubes and long cables that ran across rooms in bus-bars. This made resources scarce (computers were the size of rooms) and they were wired for bespoke tasks to conserve tasks. They were not general purpose, they were highly optimized, with the memory cache and connection buses wired in the right quantities, the ratios of memory to algorithmic operations kept within a certain threshold, specifically to do the task in a way that conserved resources the most. A computer that specialized in calculating artillery trajectories was wired different to a computer that computed financial information for banks. If they used the same structure for both, they’d pour in significantly more copper wiring and cathode ray tubes.
In the 1960s, the Silicon chip was invented, this allowed a lot of transistors to fit on a tiny wafer of silicon. This was a massive jump in computing. It was a general purpose invention, like the steam engine. Then Moore’s law happened, the amount of transistors on a single wafer that went into chip production doubled every 3 or so years. Basically, computers got about twice as good every three years, and ten times better every ten years. Because there was such exponential gains, bespoke and specific computers got replaced by general computers that did every task inefficiently. They just churned out chips, shoved them into computers, programmed them for a task using abstract and interchangeable programming languages (This is the reason the C programming language proliferated around this time, you started writing general programs that could be read by any general purpose, single core machine, and have an interchangeable compiler to reduce it to assembly/binary). Even though there were a lot of efficiency losses through lost specialization, it didn’t matter, things were getting better so fast that it made no sense to specialize. Electrical Engineering began to be replaced by Computer Science, you stopped needing to know the specific architecture. This went on through the 70s, 80s, 90s….There’s an Onion News skit about 90s computers and how by the time you buy it, it’d become obsolete, and the consumer flips out at this and kills the Bill Gates stand-in enraged. This exemplifies it, exponential gains were being filtered through inefficient general purpose machines, but because the gains were exponential, it didn’t matter.
Around 2005 or so companies started to hit a brick wall, first with clock cycles (basically the computer heartbeat, I’ll spare the technical details). If they tried to shrink down the transistors and run a single CPU core, the clock cycle limitation would mean general use computers would start to overheat. If you remember laptops and game consoles developed in this era, you’ll remember they suffered from overheating issues pretty regularly. They solved this through multi-core processing. This was advertised as a new technology…But it was really the first regression back to specialization over generalization. Early 1960s servers had separate CPU cabinets that prioritized and balanced instructions. this is why computers from back then had multiple front facing terminals that meant many different people could use the computer all at once. Servers that needed two CPU cabinets would have 2 (two cores basically), those that needed 4 would have four. Of course because new CPU centres were located specifically within the chip, they pretended this was a new innovation, even though the topographical layout is little different and concept is functionally the same. This will be the start of a long trend of explaining away the declining generalization towards specialization.
Since 2005 you’ve seen an explosion of specialization within computers and a pivot away from generalization. The Graphics Processing Unit being one such device. A Graphics Processing Unit is basically a Central Processing Unit that specializes in calculating along thousands of simplistic cores. Think of a GPU like a hundred bicyclists carrying one brick each slowly while the CPU carries a pallet of bricks on a truck. It was invented earlier for computer graphics, but it got leaned into more once single core CPUs faced limitations. This was advertised as a new thing, but really it’s again, a pivot back towards the pre-silicon bespoke era. Transistors were still getting clustered closer on silicon wafers, and there were still improvements in general resources, but after 2005 the gains were being laundered more and more through chip architecture (how the chip is structured) rather than pure chip transistor space shrinkage. Moore’s law of doubling resources was being laundered through specialization. Think of it like shooting a cannon ball 2 miles to 4 miles. The 1980s computer/cannon achieved that gain almost purely through making a larger boom, while the 2005 computer achieved 70% of it through a larger boom and the rest through adjusting the barrel length, adjusting the angle, changing the angle of fire, etc.
This brings us back to AI, the same technology that was utilized in neural networks was invented back in the 1950s and still experimented with in the 1960s, in the bespoke era. You could be experimental when you were specializing. It got shelved because of the pivot to generalized architecture was single core CPUs, unsuited to parallel operations, which made it too resource heavy to bother. It’s only when chip architecture became more diverse and the pivot towards GPUs happened did people take it off the shelf, and begin experimenting with it again.
https://en.wikipedia.org/wiki/Perceptron
Now in 2026 we’ve basically reached the end of transistor shrinkage. Go look at the Wikipedia page of 2mn die node spacing and 3nm die node spacing, and you’ll see them plainly admit it’s a marketing gimmick and branding at this point. The spacing for large segments of the chip are still around where they were in 2020 or so, it’s now almost entirely specialization driving the difference. The problem with specializing architecture is that you can min-max it 100% for certain tasks and can’t squeeze out anymore gains. We’re arguably starting to hit that point. That’s why they’re fragmenting further into things like “Neural Processing Units” (again, another specialized part on the chip even more tailored towards AI calculations than GPUs, which were originally for graphics). But fragmenting like this is frighteningly expensive and requires more fabrication plants (which cost billions) and more resources.
The problem here is that it’s questionable as to whether they can scale up or sustain the production of modern chips, if it’s a bubble feeding the specialization, then the bubble bursting unravels it. The reason the chip market at the top end has condensed towards a cartel-like monopoly (they don’t really compete, they bail each other out all the time and share the same suppliers), with a supply chain that’s like a straight line, is because it’s an absolute technological feat to achieve and that damn hard. The money that’s feeding into the AI boom is going to GPUs and these advanced chips, and filtering down to companies like Nvidia. An AI bubble bursting could fracture and shatter the entire top end market, and rupture this delicate inverted pyramid. And to cut costs they have to unravel specialization, cut the least productive specialized fab factories, and narrow production, and if it’s specialization driving the gains then that means a several year technological regression if those specialized fab plants close. The foundational layer before this that can handle organic, national supply chains and has a healthy market is around 2012 chip technology, about where China specializes. I used the analogy of the Concorde in my last post on the issue, that’s exactly what I mean.
China isn’t actually trying to scramble for the top end of chip performance or really compete with AI. They listened to the electrical engineers in their country who told them Moore’s law was over and the next gains require reaching precariously from a step ladder. They instead structured themselves horizontally and safely in the legacy chip era, are hyper-competitive at that level, and rest on a comfortable economic bedrock.
“I want 40% of the FBI hunting down antifa”
Bad News.
40% of the FBI is antifa.
If you have a Subscriber access,
simply login first to see your comment auto-approved.
Note on comments privacy & moderation
Your email is never published nor shared.
Comments are moderated. If you don't see your comment, please be patient. If approved, it will appear here soon. Do not post your comment a second time.