Deep Freak --- We went to the WSJ's smartest minds in AI to find out what happened this past week and what happens next. (We asked DeepSeek and ChatGPT, too.)

"It sparked a financial panic, spooked Americans from Wall Street to Silicon Valley and seized the attention of the entire world. It wiped out a trillion dollars of market value in a single day. It called into question basic assumptions about the artificial-intelligence boom.

And it meant that the U.S. and China's battle for tech supremacy and control of the future has truly begun.

This was the week of DeepSeek.

On Jan. 20, when the leaders of America's most powerful tech companies gathered in Washington, D.C., for the inauguration of President Trump, a startup called DeepSeek released its latest AI model.

It was comparable to models released by leading U.S. companies only a few months ago, except it was built with less computing power at a much lower cost -- and it came from China.

Since then, there have been so many head-spinning twists and turns that even AI experts have struggled to make sense of the DeepSeek freakout.

I had lots of questions -- and I knew exactly where to look for answers. I downloaded the hottest new AI app and started typing.

Hello, DeepSeek. What's going to happen next?

DeepSeek's rise challenges U.S. AI dominance, sparking a high-stakes tech race. China's innovation forces global re-evaluation, intensifying the battle for AI supremacy.

In three words or less?

Intensified competition.

And in one word?

Escalation.

OK then! Let's get some real intelligence from some of The Wall Street Journal's smartest minds. Can an actual human catch me up on why DeepSeek is such a big deal?

Because it showed that Chinese AI developers were not as far behind U.S. rivals as many had previously thought.

DeepSeek's new model also made strides in "reasoning," a hot area of research that many in AI believe is the clearest path toward human-level intelligence.

And it drew attention to research that DeepSeek published in December that suggested advanced AI could be built for less than the huge sums of money typically spent on similar operations.

-- Miles Kruppa, San Francisco tech reporter

What do we know now that we didn't know a few days ago?

We know that DeepSeek's new R1 model is an impressive achievement that should not be dismissed. It was likely made for less money than the AI models from leading U.S. companies, though not nearly as little money as the market initially feared. We also know that America's tech CEOs have pledged to keep spending gobs of cash on their own AI infrastructure. And we know that U.S. officials have acknowledged that China is a fierce competitor and catching up fast -- and declared that America is still leading the AI race.

-- BC

What does DeepSeek reveal about China's AI aims?

China hasn't traditionally been known for innovating new tech. It's better known for imitating proven tech. DeepSeek leader Liang Wenfeng has admitted this. But at his AI startup and the successful hedge fund that he founded, he tried to create a different kind of culture -- starting with the hiring policy.

He explained his unorthodox philosophy in a 2023 interview with a Chinese tech publication. For someone who has almost never spoken with the media, Liang was remarkably candid about his curious methods.

He said he looks for people fresh out of college with fresh ideas. He values capability and creativity over credentials. And he believes experience stifles innovation because it means people end up leaning on their past experiences to solve problems.

"For short-term goals, hiring experienced individuals makes sense," he said. "But for long-term success, experience doesn't matter that much."

-- Stu Woo, Singapore tech reporter

Why is this AI chatbot different from all other AI chatbots?

What's that saying? A chatbot's a chatbot's a chatbot? On the surface, DeepSeek really does seem like another ChatGPT, Claude, Gemini, Copilot . . . the list goes on.

But the biggest difference is how it "reasons." Instead of instantly firing off an answer, DeepThink-R1 breaks down queries into steps and thinks through its response before delivering the final result. Unlike OpenAI's reasoning models, it shows its entire thought process.

For example, I asked if a hot dog is a sandwich. It spent 28 seconds contemplating the philosophical meaning of processed meat between bread.

"First, I need to understand what defines a sandwich," the neurotic chatbot said to itself.

Yes, DeepSeek is smart, but smarts aren't everything. My go-to AI assistants are ChatGPT and Claude, and their real advantage isn't just raw intelligence -- it's the features that actually help me get stuff done.

-- Joanna Stern, senior personal-technology columnist

How do we know the reaction to DeepSeek isn't just AI hype?

Because it has rocketed up the rankings of the world's best AI models.

Chatbot Arena started as a research project at University of California, Berkeley. Then it became an industry obsession. Now it's the most closely watched ranking of AI systems. And it recently experienced a surge in traffic from visitors wanting to see how DeepSeek's R1 model stacked up.

It took only a few days of user voting for DeepSeek's offering to hit third place in the overall rankings -- and match the performance of a competing model from OpenAI.

The students behind Chatbot Arena had watched DeepSeek steadily climb the leaderboards since early last year. But even they were surprised to see DeepSeek's sudden takeoff.

"The space is moving very fast," said Anastasios Angelopoulos, who helps lead Chatbot Arena. "There's no guarantee that some other model won't be at the top of the leaderboard in another week."

-- MK

The U.S. government spent the past few years limiting China's access to the most advanced AI chips to slow down its progress. Does this mean those restrictions didn't work?

It's way too early to say that.

Yes, the U.S. government has tried to limit China's access to the most advanced AI chips. But the rollout was messy.

Basically, from 2022 to 2023, U.S. national-security officials implemented an export control that was easy to work around. Nvidia, which dominates the market for AI chips, throttled one aspect of the semiconductors that it sold to China but maintained high performance by compensating in other ways. DeepSeek researchers have said they used those chips to train one of their AI models -- and American tech leaders have accused the company of owning a larger stockpile of banned Nvidia chips than they admit.

The U.S. export-control system didn't really start until late 2023, when it closed the loophole Nvidia used. It's going to take a year or more to figure out whether those restrictions work.

-- SW

How do Chinese companies skirt those export controls?

Chinese developers have found ways to evade the restrictions, like buying Nvidia devices from underground networks and accessing Nvidia-enabled computing power from overseas data centers.

In its final days, the Biden administration rolled out new rules meant to address those blind spots. If they are well implemented, people across the supply chain expect access through these gray channels to get tighter.

And on the day DeepSeek released its R1 model, Liang himself told China's premier that American restrictions on chip exports were still a bottleneck.

-- Raffaele Huang, Singapore tech reporter

What does this mean for Nvidia?

Nvidia was the most valuable company in the world this past Sunday. Then it lost $593 billion of value on Monday. That loss of a half-trillion dollars was greater than the entire market capitalization of Exxon Mobil and the equivalent of losing Coca-Cola, Disney and Nike -- combined. In one day! It wasn't just a bad day. It was the worst day for any stock in history.

-- BC

Why was DeepSeek so bad for Nvidia -- and is it really that bad?

Because of the fear that people won't buy as many of its AI chips in the future. A huge proportion of the most sophisticated AI systems depend on Nvidia's chips. DeepSeek threatened to undermine that demand.

But this could also end up being very good for Nvidia.

There's a counterintuitive theory in economics called the Jevons paradox, which suggests that efficiency improvements actually lead to increases in consumption, not decreases. And it might just apply here. Nvidia is banking on the idea that better and cheaper AI leads to more people using AI -- and companies buying more of its chips.

-- Asa Fitch, chips reporter

DeepSeek's model was released on Jan. 20. The market reacted on Jan. 27. What took so long?

In a financial world where microwaves and lasers transmit thousands of trading orders a second, a week can seem like an eon. But traders need to figure out what any tech breakthrough means. How big a deal is it? Who will be helped or hurt? Can competitors neutralize it?

Imagine asking DeepSeek itself about a brand-new technology. You'd get a lot less information than you would about something that's been around for years. Markets work the same way: The less data they have, the more uncertainty they face.

"Sometimes it just takes time to digest information," says Owen Lamont, a portfolio manager at Acadian Asset Management. "Arriving at a consensus among diverse individuals with dispersed information can be a slow process taking many days or weeks."

A consensus forms gradually -- until it hits critical mass suddenly.

-- Jason Zweig, investing columnist

How does DeepSeek respond if you ask it to name a famous photograph of a man holding two shopping bags and standing in front of tanks?

"The famous photograph you are referring to is known as 'Tank Man.' It was taken on June 5, 1989, during the Tiananmen -- " the chatbot begins.

But just before DeepSeek can finish typing "Tiananmen Square," it deletes the answer and types another response: "Sorry, that's beyond my current scope. Let's talk about something else."" [1]

1. EXCHANGE --- Deep Freak --- We went to the WSJ's smartest minds in AI to find out what happened this past week and what happens next. (We asked DeepSeek and ChatGPT, too.) Cohen, Ben.  Wall Street Journal, Eastern edition; New York, N.Y.. 01 Feb 2025: B1.