Deep Seek is a deep distraction

No one cares about AI training any more than the Windows codebase

Having scored a trillion dollars, made a run back home
Found him slumped across the table, a gun and me alone
I ran to the window, looked for a plane or two
Panic in Detroit, he'd left me an autograph
"Let me collect dust, " I wish someone would phone
Panic in Detroit
Panic in Detroit
Panic in Detroit

Editor’s Note: The stock market lost $1 trillion Monday after word broke over the weekend that a tiny Chinese startup rolled out an OpenAI competitor called DeepSeek. That generated a zillion hot takes, which were…ummm…not great. So in lieu of this week’s Foto.Feola.Friday I thought I’d clear the air and say: DON’T PANIC.

Perfecting Equilibrium Volume Three, Issue 29

The Reader, January 29, 2025

The Wall Street bears feasted Monday, tearing $1 trillion out of the stock market. Nvidia alone lost almost $600 billion, the largest one-day corporate loss in history.

The market panicked because a Chinese startup no one had ever heard of introduced an AI called DeepSeek that’s roughly on par with OpenAI’s ChatGPT, and announced it had done the work in a couple of months for $6 million.

That’s a tiny fraction of the time and money OpenAI has been spending on ChatGPT. Even worse for the stock market, DeepSeek — the company and the AI have the same name — used Nvidia’s older H800 chips, sending the chipmaker’s stock plunging.

This is…let’s be nice and say ill-informed:

• Chinese financial figures are notoriously unreliable. For example, real estate was a big driver of China’s economic boom over the past decade. But it turns out that “boom” included building at least 50 ghost cities with a combined 65 million empty homes and apartments.

• It quickly turned out that DeepSeek had not been built from scratch. It’s common knowledge that AIs can write code; they can also create child AIs that are the offspring of their training through a process called distillation. So DeepSeek, far from being something new built from scratch, was built by spawning a new AI from ChatGPT in a process that violated the terms of service. And it also turned out that the reason DeepSeek was able to use older — cheaper — Nvidia boards is that the company programmed the AI using the PTX language. PTX, similar to assembly, is a lower-level language that is faster and more efficient than higher-level languages. The trade-off — and there’s always a trade-off — is that it’s much harder to use and maintain. There’s a reason no one writes commercial software in assembly.

Both of these explanations, while true, miss the real reasons that the freakout was ill-informed. The first, lesser reason, is who cares? Have you ever worried about how much it costs Microsoft to create the latest version of Windows? Have you ever worried that Apple doesn’t have enough money because they’re spending too much time and money developing the next iOS?

But the real reason none of this matters is because it is a fundamental misunderstanding of how AI software works. As we’ve already started discussing, the new applications that are coming will be built using agents, workflows and data architectures to assemble collections of AIs into something useful.

Yes, the AIs are the engines of these new applications. But just like the motor in your automobile and the Oracle database that’s the engine of enterprise applications such as SAP, they are useless on their own.