We continue from DeepSeek and the AI Bubble; Napkin Calculation Part 2.
The bubble continues to grow, but is near popping. We live in an age of fabulous thinking, where false hopes are inspired by technology so alien to human experience, it becomes impossible for most of us to think critically about the subject. We substitute judgments based on credibility, social standing, slick presentations, and the metrics of investment, which appear to generate new wealth based on market capitalization. How can this substitute for understanding the subject?
It can’t. A market system that works for pricing soybeans can’t deal with AI on a rational basis. Everybody can understand beans; nobody understands AI, even those who work in the field. Bear with me as I try to back fill the gap. If I partly succeed, you will have new doubts, valuable in this fabulous age.
There is a thread going way back that bears on this. It began with Cantor’s diagonal slash of 1891. While lying in a field in 1936, Alan Turing used Cantor’s trick to extend the results of Kurt Gödel. He invented the mostly-hypothetical Turing machine to prove that there are some problems which cannot be decided by computer programs. As part of the proof, Turing invented the Universal Turing Machine, a hypothetical programmable computer. Around 1938, John von Neumann, considered by many to have had the most extraordinary mind of any human, saw Turing’s proof as the basis of what would become the modern computer, which you are using right now. The von Neumann machine can do anything a computer can do, though it might take a long time to do it.
This is the paradigm of modern computing, with roots in the year 1891, built by giants standing on the shoulders of giants, where a computer program is equivalent to sequential execution of instructions, one step at a time. You don’t have to know the math to realize that when a technical paradigm becomes embedded in society to this extent, it becomes a social truth, like the law of gravity. It remained inviolate, except in very minor ways (multiprocessing), until the advent of high resolution displays.
All this theory, this precise understanding, preceded the construction of the first modern computer. The AI bubble reverses this; huge data centers are under construction to execute AI paradigms, while the theory is weak or absent.
A 4K display, including subpixels, contains 24,883,200 addressable locations. A computer that executes one instruction at a time on each individual pixel would be insufferably slow. Every chip maker had to address this. Intel’s solutions were simple, AMD’s were wonky, while Nvidia’s sophistication was capitalized by the game market. We owe a lot to Grand Theft Auto for where we are today.
Nvidia et al. tinkered with the von Neumann machine. In place of a single computer that can do everything, they devised tiny computers, specialized to running game graphics, that could be placed in the thousands on a single chip. With each tiny “core” responsible for a manageable number of pixels, computer displays became dynamic, full of fun and cash flow. In the early 2000’s, Nvidia realized that massive arrays of these simple, graphics-oriented cores, enhanced for the purpose, could take on some of the responsibilities of von Neumann machines. The massively parallel desktop computer became a reality with the Nvidia Tesla (not to be confused with the car maker) architecture.
AI based on evolved graphics processors has an upside and a downside. Nvidia has made possible the closest thing to AI that is currently practical. But as a descendant of the von Neumann architecture, albeit with many creative twists, it is a dead end,.
In the next half decade, that billions of infrastructure will go out in the trash, destined for metal recyclers. All that investment will be lost, because it is premature.
This is dense, so I’ll continue shortly.
***Fate of Nvidia data centers***