Is This New Chinese AI a 'Sputnik Moment'?
DeepSeek has released a cheap, open-source artificial intelligence. Does it challenge American AI supremacy?

The Chinese startup DeepSeek has released the R1 advanced reasoning model, a cheap open-source artificial intelligence, prompting Silicon Valley investor Marc Andreessen to declare that we have reached "AI's sputnik moment." The analogy implies that the product threatens American AI dominance in a way comparable to the space-race challenge posed by the Soviet Union's first orbital satellite in 1957.
The U.S. has imposed export controls on the most advanced computer chips, forcing the Chinese developers to optimize their new model using much less capable chips. This spooked the markets, and the stocks of U.S. chip makers like Nvidia and AI developers like Microsoft and Meta fell this morning.
DeepSeek R1 performs comparably to the top releases from American AI labs, although Google's Gemini Flash-Thinking still outperforms it in the Chatbot Arena. But DeepSeek's chief immediate threat to U.S. models is that it claims to be much cheaper than many of its American competitors. One AI insider has argued to me that the circumstantial evidence around the DeepSeek models suggests that part of the data to train them came directly from other models in ways that go beyond contamination from web scraping. Nevertheless, DeepSeek's optimizations and compute efficiencies will undoubtedly spur other AI developers improve their models similarly. We could land in a Jevons paradox situation where such improvements in efficiency actually increase the demand for high-end computer chips.
Not everyone is rushing to declare the end of American AI dominance. Investment analysts at the Wedbush financial services firm argue that though "the model is impressive, and it will have a ripple impact, the reality is that [Magnificent] 7 and US tech is focused on the [artificial general intelligence] endgame with all the infrastructure and ecosystem that China and especially DeepSeek cannot come close to in our view."
The AI insider who I spoke with agrees, noting that projects like Stargate aim to go far beyond the chatbot level to deriving new insights from first principles in biology, physics, chemistry, climatology, etc. He analogized DeepSeek and current chatbots as the equivalent of running one experiment at time; the American AI projects currently in the works, he argues, will be comparable to running millions simultaneously.
Writing on X, the former OpenAI staffer Andrew Mayne rejected the "Sputnik moment" comparison. Instead he called this "a Buran moment," invoking the Soviet effort to copy the American space shuttle.
Sputnik or Buran? We'll know well before the end of this decade.
Disclosure: I have small holdings in both Microsoft and Nvidia.
Editor's Note: As of February 29, 2024, commenting privileges on reason.com posts are limited to Reason Plus subscribers. Past commenters are grandfathered in for a temporary period. Subscribe here to preserve your ability to comment. Your Reason Plus subscription also gives you an ad-free version of reason.com, along with full access to the digital edition and archives of Reason magazine. We request that comments be civil and on-topic. We do not moderate or assume any responsibility for comments, which are owned by the readers who post them. Comments do not represent the views of reason.com or Reason Foundation. We reserve the right to delete any comment and ban commenters for any reason at any time. Comments may only be edited within 5 minutes of posting. Report abuses.
Please
to post comments
Boldly going where others have already gone.
Watching this story has been amusing. The CIO for the Chinese company even admits this is a smaller, more focused AI, not thr broad attempt with openAI.
"DeepSeek did something really impressive with far fewer employees and funding than openAI did, and in less time," said Mark Malek, CIO at Siebert. Malek noted that DeepSeek, "DOESN'T COMPETE WITH OPENAI," and went to explain some of the differences between DeepSek and more well-known AI apps.
"OpenAI (and Google's Gemini) is a broad, general-purpose tool, based on a vast corpus of information. It can be used to create more narrowly focused applications as well. OpenAI has vast capabilities in natural language processing, while DeepSeek is created to be task-specific," Malek said. "Now, that doesn't mean that DeepSeek is not good. What is important to understand is that it is not the same thing as OpenAI, so it would logically take fewer resources."
Yet news and media are freaking out.
It’s the new Mastodon!
I would just like to remind everyone that Bailey's 'science' record during the pandemic was fucking atrocious.
Take everything Bailey writes with mega-rocks of salt.
"DeepSeek has released a cheap, open-source artificial intelligence."
Says what reliable source?
The CCP.
Anyone who assigns any credibility to anything any government says is a Damned Fool.
Anyone who assigns any credibility to anything any communist government says is a Damneder Fool.
Anyone who assigns any credibility to anything the CCP says is the Damnest Fool.
How's that Wuhan Lab working out, Bailey? Still believe them?
The millions of DDoS attacks that have been preventing anyone from registering to test drive it for many hours now. The entities behind those attacks are clearly confident that there is enough value in the product that denying its existence is worth spending effort on to undermine - for at least one news cycle.
Those attacks succeeding kind of prove that they were focused on one thing - and not the other thing - which is evidence of cheap.
Oh - and open source means that its test results and setup and everything else are available online at github.
Granted that will take around $6 million to replicate to see if its all true. But compared to the closed source Big Tech shit that hypes the need to spend hundreds of billions (or just sign up with our product) - well it won't take long for 'reliable sources' to create product based on it.
How does one DDoS something behind 'the great firewall' one wonders.
As far as I can see the only source for that claim is the company itself, and it's not outside the realm of possibility the claim of a DDoS is to shield them from anyone actually verifying their product claims.
They've put everything on github. It's truly open source. Obviously the vast vast vast majority of people (basically everyone except Big Tech) want to see a working model not specs for how to build/train/test/deploy a working model.
They need the same stuff that chatGPT needs to get hyped. An ability to register so you can start asking questions and testing prompts. And in fact they are able to accept registration via Google. Which I hate - and which does bypass the DDoS attacks on DeepSeek itself. But hey - it's a way to play with it unless/until you want to serious deploy it beyond having Google as a useless data-stealing intermediary.
Nothing you've written here explains how it could be that a DDoS attack is occurring in China from international sources.
Perhaps you are saying that GitHub itself is being attacked, but as others have pointed out that could simply be normal traffic on an unprepared web service.
Deepseek-V3-FW
Deepseek-v3-T
And BTW - here's a partial example of an identical prompt that I've also set up through chatGPT:
Prompt: You are a hematologist specializing in uuuu. The patient (xxxx) has been diagnosed with yyy, is being treated zzz. Please interpret the following current comprehensive blood count (CBC) metrics and confirm/adjust the current treatment regimen
ChatGPT: Based on the provided CBC and the patient's clinical history of yyy, the current treatment regimen appears to be appropriate, with some potential considerations.....long details ....The current treatment is aligned with evidence-based guidelines for yyy management, but close monitoring and patient-specific adjustments remain crucial.
DeepSeek: This patient with yyy is showing some improvement in response to the current treatment regimen, but further adjustments may be necessary based on the CBC trends and clinical goals. Below is an interpretation of the CBC metrics and recommendations for the treatment plan....long details....Summary The patient is showing a partial response to A and B, but the F remains above target. Continue current therapies with close monitoring, and consider increasing the B dose if the F does not improve further. Address potential G and maintain vigilant follow-up to reduce the risk of H complications.
The long details of both are quite similar and in line with what is actually online and clinical practice. Not black female Vikings writing the Constitution. DeepSeek is imo much better at the intro and summary paragraphs. Those two are far more cohesive and in synch with the details. ChatGPT is just generic.
As a simple intro, this easily passes into the same bucket as the other LLM's. As open source, this opens the world up to all sorts of specialized deployments where I pay a simple fee to DeepSeek rather than rentier royalties to BigTech that rob me blinder and blinder because they overspent.
Pay royalties? DeepSeek source code is released under the MIT license, which explicitly grants royalty-free usage rights.
Yes I know. It's Big Tech that will be trying to extract rentier royalties from those who use their proprietary systems
I'm guessing not an attack. They just didn't plan for the desired usage from all the buzz. This is a distributed denial of service via normal user behavior. Scale better guys.
C'mon man! [LatinX hip swiveling] Their Github page has 3 authors, 4 contributors, and had its first commit last week! People owning things is, like, an abstract social construct!
Knees jerking all over Wall Street. Oceans of red ink flooding millions of computer screens. I feel like I've been here before.
1. If true, it illustrates that sanctions have unintended consequences, in particular strong incentives to work around them.
2. If even partially true, it's a step forward for creative destruction, the result of which is to advance civilization writ large. The establishment is on notice that it can't rest on its laurels.
If complete and total communist propaganda ?
Can you say "economic meltdown"?
As in the lies about a certain "pandemic" that made sure our economy took a hit we will be decades in recovering from?
A pandemic that was really just another flu?
In the words of Jurassic Park - Life finds a way
In the words of Baltasar Gracian - Never compete with someone who has nothing to lose.
Regardless - the techbros deserve a giant sledgehammer to their arrogant face. If DeepSeek is able to solve today's problem - the millions of DDoS attacks (presumably orchestrated by Silicon Valley) that are preventing anyone from registering and taking it for a test ride - there is an entirely set of new AI customers who really despise Big Tech and its VC rentiers. THAT is what will drive the AI revolution. Small specialized decentralized AI agents (human scale) - built on an open source LLM that isn't trying to steal everything those AI agents do.
That cool AI assistant in your phone that Reason writers are snapping their fingers and swiveling their hips about? Yeah, it just made your end-to-end encryption useless.
AI is going to peak soon anyway because we're running out of information to feed it.
Not a chance, sarcasmic.
Wait. Is it as good as Tecmo Bowl?
It's as good as techno bowl without Bo jackson
In 1992 I saw Tecmo Bowl AI perform some of the most amazing feats of athleticism the game of football has ever witnessed.
No, like everything else from China, it's a cheap, stolen copy of other people's work.
All your base are belong to us
Somebody set up us the bomb
DeepSeek models suggests that part of the data to train them came directly from other models in ways that go beyond contamination from web scraping.
Our theft of IP is ok, China theft is bad.
Just laughing at the AI creators now worried about IP theft when arguably that is central to the AI business model. China probably has just outright stolen the code (or faking results, but I've not seen reporting on what testing was done to set off these alarm bells yet).
Yeah, it's actually pretty amusing to me that now they might care about IP law.
Feels and awful lot like the "The borders should be completely open (but only in one direction... or when I say so... and only to the people and things I say can come across... and only the one border)." social construction/re-engineering going on.
Regardless, does anyone remember the results of the first "Sputnik moment"?
Yes, the US went on to put people on the moon.
Is This New Chinese AI a 'Sputnik Moment'?
It's actually a Hayek moment. Until now, Big Tech's model for AI has been:
OK. Yeah we've stolen everything that you all put on the Internet. More fool you. You shouldn't have trusted us as far as you could spit. Now - if you want it back, we're spending a whole shit ton of tens of billions of dollars so that you can access it via all the ways we choose to let you access it from inside a black box that works by magic. And everything that you access from that is also going to be stolen by us.
Obviously that leads to the tech dream of AGI and Skynet and lost jobs by the billions and serfdom for everyone but the few who will fulfill their dreams of boldly go where no man has gone before (or Mars) with virtual supermodels in a metaverse wearing elf ears and saying Nanoo nanoo.
The DeepSeek model is (or can be):
OK. Yeah. We stole what they stole earlier. But we spent a lot less money and just proved the aphorism - yesterday's breakfast isn't worth a rental stream on all the food you will ever eat in the future. It's worth leftovers. So, here's a way you can tap into all that cheaply. And here's another way you can tap into both it - and knowledge that you will create in future that is both valuable and won't be stolen this time because you are going to protect it this time right?
It changes the incentive to create and use knowledge - in future.
the knowledge of the circumstances of which we must make use never exists in concentrated or integrated form but solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess.
More accurately - the personal knowledge (which is valuable) can rest on the shoulders of the common knowledge (which is worth yesterdays stolen leftovers).
Deepseek-V3-FW
Deepseek-v3-T
I'm frankly amazed more people aren't with me on the whole "nuke Beijing and Shanghai" plan.
This will likely reveal the holes in the chip embargo because it's very likely this model was built on the same Nvidia chips OpenAI uses and China isn't supposed to have. It will be reviewed that China obtained the chips in some backdoor manner.