|
|
Reprinted from The Deep View: https://www.thedeepview.co/p/ringing-the-deepseek-alarm-bells-what-we-re-missing
Tech stock selloffs have hit your retirement account. Trade tensions have increased the prices you pay for housing and food. Fears abound that your private data will be leaked to bad actors overseas who will use it to impersonate you. Unless you’ve been living under a rock, you’ve likely heard about DeepSeek, the Chinese startup entangled in a chain of events contributing to our increasingly volatile everyday lives.
In January, DeepSeek released an AI application called R1 that was reported to have achieved a level of performance equal to that of industry leaders like Open AI’s ChatGPT and xAI’s Grok. Quickly demo R1, which is available freely online, and you will find that its performance really does feel just as good as, if not better than, a ChatGPT, a Claude, or a Grok. That’s a problem so big that federal bills have been proposed to make using DeepSeek illegal.
A Wakeup Call for the West
The present debate in the popular media focuses on competition between the United States and China: perhaps the United States doesn’t have the technological advantage we thought it had.
- The fact that a start-up out of China was able to create something that performs at state-of-the-art levels is a wake-up call for folks who thought Western bloc technologies would remain dominant for the foreseeable future. That raises the same questions that the United States started asking when the Soviet Union sent Sputnik on its maiden orbit: Do they know something that we don’t? If they possess this advantage, what other advantages do they possess that we don’t know about? The obvious take-home here is that the United States isn’t necessarily winning the analogous “space race.”
- But that take-home obscures the flaws in our fundamental assumptions about what will make the United States competitive in artificial intelligence. The blockbuster revelation from DeepSeek’s release of R1 was this: it only took DeepSeek’s developers a couple million dollars to train an industry-leading model.
In other words, a little startup — which was late to the game — was able to release a multi-billion-dollar product for less money than these big tech firms spend on just one or two senior engineers.
Punching Holes in Big Tech’s Approach
The big tech firms have been raising billions of dollars of capital (see, for example, Stargate) by arguing that the most effective way to assert dominance in AI is to build bigger, more complex models — complete with bigger datasets, more advanced chips, more electricity, greater real estate for data centers and top engineering talent.
The DeepSeek revelation throws a wrench in the story leading firms have been telling: that success in LLMs is capital intensive, that bigger is better and that early winners will dominate. And they’ve told that story at the expense of supply-constrained resources — like chips, talent and energy.
Speed v. Efficiency
If a model like DeepSeek can use clever mathematical optimizations — the bulk of which were readily accessible in the public scientific literature — to run large language models at less than 98% of the GPU utilization required by other leading models, then why have we been trying to invent GPU chips that do hard work faster, rather than just inventing new statistical methods that make the work itself easier?
It reminds me of the old tale of the consultant who realizes they make more money when they fail to solve the client’s problem quickly, because they’re charging by the hour. Our best-capitalized firms may lack incentives to create more efficient, competitive models. Perhaps leaders committed to scaling faster because they believed it would produce a first-mover’s advantage. But that, too, has been revealed as false — the upstart firm was late to the game.
Shifting Priorities
The energy required to run these GPUs at large data centers is massive, and that energy has to come from somewhere. As states add new energy-generating capacity to support growing demand, tax- and rate-payers risk getting stuck holding the bag if we overbuild capacity for a technology that just became staggeringly more efficient.
- Moreover, the environmental impact of such energy use is significant: studies suggest that training a large language model can emit as much CO2 into the atmosphere as a few thousand airplane trips — a non-trivial amount of pollution.
- Why have we prioritized greater, more pollutive industrial scale, before solving thelow-hanging mathematical efficiency problems that would reduce energy risk and environmental impact?
A Better Approach to Dominance in AI
The good news is that there is a path forward. I believe DeepSeek could prompt the industry to refocus its efforts on investing in intellectual property, true methodological innovation, unique protected assets like curated data and domain-specific applications.
The wake-up call from DeepSeek’s emergence isn’t just about global competition — it’s about the need for a fundamental shift in how we approach artificial intelligence. If the U.S. wants to maintain its leadership, it must move beyond brute-force scalingand capital-intensive expansion. The focus should be on smarter, more efficient innovation: refining mathematical methods, optimizing resource use and fostering an environment that rewards true breakthroughs over sheer size.
The next era of AI dominance won’t be won by those who spend the most — it will be won by those who think the smartest. DeepSeek has shown us that agility and ingenuity can outpace deep pockets. Now, it’s time for the U.S. to prove that it can still lead by doing what it does best — innovating with purpose.
