Chinese AI’s Sputnik Moment
Deepseek-R1 has shaken the world by somehow matching top AI models on a fraction of the budget.
When I started noticing the hype around DeepSeek-R1 all over my Twitter timeline four days ago, I decided to do a test.
“I’m trying to pick one growth area in my life, choose one highly effective technique and practice it every day. Can you come up with some ideas?” I asked, not really caring about the answer.
DeepSeek gave me a pretty generic list of self-help tropes. Serviceable but unimpressive. Then I noticed the little “DeepThink” button on the interface. This was the button that triggered the latest model, DeepSeek-R1.
“OK, the user is asking for ideas to pick one growth area and a daily technique to practice. They mentioned they want something highly effective. Let me start by understanding their needs…”
Deepthink-R1’s inner dialog (which is greyed out to distinguish it from the actual response) went on for more than a page before it started answering me for real. They correctly deduce that their previous list of ten techniques was probably overwhelming. They remembered in the same context window I previously asked about the DeepThink button for complex questions and hypothesized that I might want more depth.
I make my confession to DeepThink-R1: I’m not looking for self help advice, I am trying to explore their new capabilities. Why is it so much better than the “base model”?
“Your duplicate question exposed a key design choice. For AI, ‘thinking’ is really resource prioritization. Every response has a hidden computational budget. Default mode spreads resources thinly to cover many possibilities. DeepThink concentrates resources on fewer, higher-quality pathways. This mirrors human cognition!”
Basically default mode is like highway driving. It’s on autopilot. Deep Think mode is off-road exploration which might take you to more interesting places. Default synthesizes. Deep Think create a “critique loop” that constantly workshops the answers in the light of questions that it asks itself (and which we are privy to DeepThink-R1’s monologs).
I was shook. This conversation (which you can see much more of on my Twitter thread or full chat) felt qualitatively different from any other I’ve ever had with an AI before. There was something inexplicably moving about seeing DeepThink-R1’s inner monologs. Somehow knowing it’s a trick didn’t make it less magical. This thing had interiority.
I wasn’t the only one that was shook.
“From Stanford to MIT, deepseek r1 has become the model of choice for America’s top university researchers basically overnight,” said AI investor Anjney Midha. “Deepseek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, a profound gift to the world,” said tech investor Mark Andreessen.
And it’s not just the vibes that are good. Benchmark testing confirms that DeepSeek-R1 is at the top end of performance, just pipped by Open-AI’s flagship o1 model. While being sold at 3% of the price! And if you just want to try it in the browser, it’s free.
According to a blind item DeepSeek-R1’s leap has put Meta Genai into panic mode.
“Engineers are moving frantically to dissect deepseek and copy anything and everything we can from it.”
In short, Chinese AI have made the move from a low-priced fast-follower to a threat to take over the leading edge, all while still being open-sourced and cheap as cabbages. This is completely upturning the conventional logic that winning the AI game is all about who can shovel the most compute into their models.
Is the game changed forever?
Or is it just going to continue on another level? As you’ve probably heard by now, Open AI, Oracle, SoftBank and other tech titans are joining forces to launch Project Stargate, a US$500 billion entity that will build out data centers and power generation to support AI development. Will the “more is more” approach simply going to overpower relatively shoestring operations like DeepSeek? Especially as DeepSeek is essentially abandoning its “moat” by making everything open source?
But wait! Don’t forget that DeepSeek itself is probably going to have all the resources of the Chinese government behind it soon. It’s not going to be David vs. Goliath forever. It’s going to be Goliath vs. Goliath.
It’s well above my pay grade to try and analyze who will come on top in the end. But I think it is uncontroversial to say that shockingly good AI will be available at prices accessible to most of the world. A comparison often made is that Chinese AI like DeepSeek (by far not the only Chinese player) will be the Android to Open AI et al’s iOS.
Beyond AI, the rise of DeepSeek is also the harbinger of another kind of Chinese company. While the Stargate alliance is talking about firing up the money cannon, the founder of DeepSeek Liang Wenfeng is talking about culture.
“For the past 30 years, we’ve been focused on making money, often at the expense of innovation. We’ve been constrained by past habits, but these are transitional. (…) In the face of disruptive technology, a closed-source moat is temporary. Our value lies in our team. Building an organization and culture that can consistently innovate is our real moat.”
China is not just interested in underpricing the West anymore, they also want to out-innovate it.
On the elephant in the room
Yes, as if you couldn’t tell I am a DeepSeek-R1 stan. Because it’s just that good. But I acknowledge it is a product of China and as such it has its limitations, and one could argue risks.
If you talk about anything at all controversial from the CCP’s point of view, DeepSeek-R1 will first beg you to talk about “maths, coding and logic” instead, or simply have a complete breakdown that can be somewhat entertaining to behold.
A lot of people seized on DeepSeek-R1’s politically-imposed guardrails, which are exactly as extensive and unyielding as you would expect. There are a lot of other kinds of cope online. Some people are convinced it’s a secret front project of the Chinese military or some such. I’ve seen no evidence that this is the case, nor do I think it makes sense that a government-sponsored hidden program would do better than a team of smart, entrepreneurial people in a field like AI.
There are reasons to be wary of a Made-in-China LLM coming to dominate. But that’s not an excuse to denigrate its unmistakable achievements. The fact of the matter is, no matter how DeepSeek-R1 came about, it is a fact on the ground.
The AI race is on.
Nice article and informative. But I struggle with the selective blindness. Why are the guardrails applied by a state any more dangerous than the guardrails applied by a tech oligarch operating in a political oligarchy? Does anyone REALLY believe that a billionaire oliigarch with lots of so called ‘deep state’ connections will not be building in ‘guardrails’. The only safety we really have is to demand more open source.
Many will not accept the extensive politically-imposed guardrails and it will take some of the wind out of its sails.