Posted by 
Last week I told you about the Chinese AI company DeepSeek’s recent model releases and why they’re such a technical achievement. The DeepSeek team seems to have gotten great mileage out of teaching their model to figure out quickly what answer it would have given with lots of time to think, a key step in previous machine learning breakthroughs that allows for rapid and cheap improvements.
This week I want to jump to a related question: Why are we all talking about DeepSeek? It’s been called America’s AI Sputnik moment. It’s at the top of the iPhone App Store, displacing OpenAI’s ChatGPT. The CEOs of major AI companies are defensively posting on X about it. People who usually ignore AI are saying to me, hey, have you seen DeepSeek?
I have, and don’t get me wrong, it’s a good model. But so are OpenAI’s most advanced models o1 and o3, and the current best-performing LLM on the chatbot arena leaderboard is actually Google’s Gemini (DeepSeek R1 is fourth).
All of which raises a question: What makes some AI developments break through to the general public, while other, equally impressive ones are only noticed by insiders?
The lesson of ChatGPT
Several months before the launch of ChatGPT in late 2022, OpenAI released the model — GPT 3.5 — which would later be the one underlying ChatGPT. Anyone could access GPT 3.5 for free by going to OpenAI’s sandbox, a website for experimenting with their latest LLMs.
GPT 3.5 was a big step forward for large language models; I explored what it could do and was impressed. So were many other people who closely followed AI advances. And yet, virtually no one else heard about it or discussed it.
When OpenAI launched ChatGPT, it reached 100 million users within just two months, a record. ChatGPT was the exact same model as the GPT 3.5 whose release had gone largely unremarked on. The difference was that, instead of a “sandbox” with technical phrases and settings (like, what “temperature” do you want the AI to be?), it was a back-and-forth chatbot, with an interface familiar to anyone who had ever typed text into a box on a computer.
It wasn’t the technology that drove the rapid adoption of ChatGPT — it was the format it was presented in. And I think that’s the same phenomenon driving our present DeepSeek fervor.
DeepSeek R1 isn’t the best AI out there. As a largely open model, unlike those from OpenAI or Anthropic, it’s a huge deal for the open source community, and it’s a huge deal in terms of its geopolitical implications as clear evidence that China is more than keeping up with AI development. But none of that is an explanation for DeepSeek being at the top of the app store, or for the enthusiasm that people seem to have for it.
I suspect that what drove its widespread adoption is the way it does visible reasoning to arrive at its answer. It’s the first to have visible chain of thought packaged into a friendly chatbot user interface. People love seeing DeepSeek think out loud. They talk about how witnessing it “thinking” helps them trust it more and learn how to prompt it better. (Ironically, it also makes the model’s Chinese government-driven censorship more visible — don’t ask it about Taiwan — but I think that’s ultimately a good thing, compared to doing similar ideological censorship more subtly.)
OpenAI, by contrast, made the decision when releasing o1 (which does similar thinking and reasoning before producing an answer) not to make the “thought process” public and visible. This is probably for several reasons — it’s a trade secret, for one, and the model is much likelier to “slip up” and break safety rules mid-reasoning than it is to do so in its final answer. (Indeed, there have been lots of videos of DeepSeek R1 saying things critical of China before it notices the problem and backtracks.)
But I think that the thought process does something similar for typical users to what the chat interface did. It makes the AI more immediate, more accessible, more interactive, and less confusing. It’s not a major difference in the underlying product, but it’s a huge difference in how inclined people are to use the product.
“Seeing the reasoning (even how earnest it is about what it knows and what it might not know) increases user trust by quite a lot,” Y Combinator chair Garry Tan wrote.
AI has improved since you last checked in
Let’s quickly respond to a few of the most prominent DeepSeek misconceptions: No, it doesn’t mean that all of the money US companies are putting in has been wasted. DeepSeek demonstrated (if we take their process claims at face value) that you can do more than people thought with fewer resources, but you can still do more than that with more resources.
DeepSeek might be an existential challenge to Meta, which was trying to carve out the cheap open source models niche, and it might threaten OpenAI’s short-term business model. But the long-term business model of AI has always been automating all work done on a computer, and DeepSeek is not a reason to think that will be more difficult or less commercially valuable.
Another thing that is driving the DeepSeek frenzy is straightforward — most people aren’t AI power users and haven’t witnessed the two years of advances since ChatGPT first launched. But during those two years, AI has improved dramatically along almost every measurable metric, especially for the frontier models that may be too expensive for the average user.
So if you’re checking in for the first time because you heard there was a new AI people are talking about, and the last model you used was ChatGPT’s free version — yes, DeepSeek R1 is going to blow you away. And while it’s a very good model, a big part of the story is simply that all models have gotten much much better over the last two years.
I wrote at the start of the year that, whether or not you like paying attention to AI, it’s moving very fast and poised to change our world a lot — and ignoring it won’t change that fact.
That’s why it’s a good thing whenever any new viral AI app convinces people to take another look at the technology. To decide what policy approach we want to take to AI, we can’t be reasoning from impressions of its strengths and limitations that are two years out of date — not with a technology that moves this quickly. Inasmuch as DeepSeek has inspired policymakers to stop and take notice of how the 2025 world is different from the 2023 world, that’s great news.
Inasmuch as DeepSeek inspires a generalized panic about China, however, I think that’s less great news. The Chinese Communist Party is an authoritarian entity that systematically wrongs both its own citizens and the rest of the world; I don’t want it to gain more geopolitical power, either from AI or from cruel wars of conquest in Taiwan or from the US abdicating all our global alliances. But the AI race is not like the nuclear weapons race, because there was never any risk that the nuclear weapons would decide to take matters into their own hands.
AI, experts warn quite emphatically, might quite literally take control of the world from humanity if we do a bad job of designing billions of super-smart, super-powerful AI agents that act independently in the world. (Would we be that careless? Yes, absolutely — we are hard at work on it!)
A lot of people, nervous about this situation, have taken to morbid humor. “Call me a nationalist or whatever,” one popular X post reads. “But I hope that the AI that turns me into a paperclip is American-made.” But let’s get serious here. China doesn’t want to destroy the world. There are indications they’re imitating most of the safety measures recommended by US institutions and taken by US labs. Those measures are totally inadequate right now — but if we adopted adequate measures, I think they might well copy those too, and we should work for that to happen.
We are in a real geopolitical competition with real and enormous stakes, but we cannot afford to lose sight of where there’s common ground, and not creating a powerful new geopolitical entity that will gladly seize control from us and the CCP alike is a place where there’s common ground.
A version of this story originally appeared in the Future Perfect newsletter. Sign up here!
