No. of Recommendations: 3
A disclosure before plowing into this: Softbank was one of the first stocks I bought on the Tokyo exchange - I forget how long ago it was, but it's treated me very well (up nearly 600%). When investing in technology, they always seem to be at the right place at the right time.
Maybe I'm antiquated, but I don't separate the world into CPU's and GPU's. I started with microcomputers when they were still based on glorified machine-control chips. Over the years, I saw arithmetic processing chips sit next to Intel CPU's until they were incorporated. I saw RISC (reduced instruction set chips) chips steal the high road for a while. Then the ability of graphics cards (such as those made by Nvidia) to do math faster than the Intel and AMD designed CPU's and, all of a sudden, it seems like they were a different beast entirely.
There seem to be only a handful of ways to make a faster (non-quantum) computer.
Add more cores to an existing design
Add more chips in a larger multiprocessing matrix - and create efficient software to control each CPU doing is own thing and then re-assembling
Add auxiliary chips which offload specific functions and handle them quicker than the "main" CPU
Modify the complexity of the operating system/instruction set to perform the same tasks more efficiently
While you may think that the same functions could have been done "cheaper" with Nvidia chips, first of all, the toys they used are significantly less expensive than most CPU's let along GPU's, and more importantly, they are likely beyond the reach of the US ability to sanction them.
The Chinese trick was, not only to create a matrix of an apparently vast number of "mini" CPU's, but equally importantly, to create a software environment to control this multi-processor behemoth.
Wendy mentions: "However, if your goal is to design a breakthrough jet engine, predict a hurricane’s path, or simulate a nuclear reaction, generative AI is useless because it “hallucinates” and guesses numbers. For those tasks, the brute-force, high-precision mathematical certainty of a CPU-heavy supercomputer is the only tool that works".
I have done some pretty diverse projects recently using three different AI's - ChatGPT, Gemini and Claude. While they each have different strengths - and weaknesses, I find the sfest way to use them is to have one critique the work done by another.
For assembling an essay using content found on the Web, Gemini clearly has an advantage. For doing technical scientific/engineering/programing projects, Claude brute forces to a reasonable solution. While trying to get ChatGPT to actually "do" anything technical is one of life's great frustrations, it excels on finding ways to tell Claude how to improve its results. If I was a verbose document created (whether an eight page love poem or a legal protest), ChatGPT seems to have the edge.
While blatant hallucinations are not as frequent a problem as they were in the past, EV EVRYTHING created by an AI (especially ChatGPT) should be read (and reread) before sending out - not only to trap hallucinations, but to "correct" verbiage which which it has decided is appropriate but sounds too weird for one reason or another.
I think the difference between how AI is used in China compared to the US is that they are rapidly applying it to their social structure and our companies are competing to be king of the mountain without regard for the money others are pouring into the process. Think of the amount of money lost building the original attempts at the Panama Canal, the Suez Canal and the world's railroad infrastructure (not to mention the internet). A few companies drifted to the top and became wildly profitable, while the majority snk into oblivion and most investors got fleeced.
Jeff
