What's DeepSeek, and why should you care?
How a Chinese company's open-source approach is making advanced AI more accessible, challenging US tech hegemony, and why it matters.
A few weeks ago, while big tech leaders were busy bending a knee to the president at his inauguration, a Chinese AI upstart called DeepSeek released some powerful models into the wild. The move disrupted the markets and possibly changed the whole AI game.
DeepSeek's new AI models can solve complex problems just as well as OpenAI's GPT-4, but they use 90% less energy for each task and need only one-tenth of the computing power to train. They achieved this by creating a system called "Mixture of Experts," which only uses the parts of the AI needed for specific tasks, instead of using the entire system every time.
"This is a wake-up call," declared President Trump, and Wall Street reacted swiftly. US tech stocks lost billions in market value as investors grappled with a future where American dominance in AI is no longer guaranteed.
DeepSeek challenges big tech AI dominance
DeepSeek’s emergence challenges the status quo in several ways:
Power shift: While the US still holds significant advantages in AI development, DeepSeek’s success signals that China can compete at the highest levels of innovation.
Democratization: Open-sourcing under MIT (with paid tiers) lets smaller players build tailored AI tools, threatening Big Tech’s paywall models.
Economic ripples: The market reaction underscores the potential disruption posed by cost-efficient competitors like DeepSeek.
Efficiency: DeepSeek’s advanced AI models with lower computational power enable high performance with fewer resources, potentially driving a shift towards more sustainable and cost-effective AI development.
The empire strikes back
This news surfaced days after the US government announced a program called Stargate. The $500 billion plan aims to keep America ahead in AI by building new data centres and funding research. (But, only $100 billion of the money has been officially set aside, and analysts aren't sure if the plan will work.)
OpenAI's CEO Sam Altman admitted that DeepSeek's technology is impressive but promised that future versions of ChatGPT would be even better (cough, cough). Meanwhile, US lawmakers want stricter rules about selling advanced computer chips to Chinese companies.
How Chinese AI is reshaping global innovation
What makes DeepSeek's achievement more impressive is that they did it despite US restrictions on selling advanced computer chips to China, which started in October 2022.
"These Chinese models were processor-constrained, so it led to some creative techniques in training, and the DeepSeek model has come out with better-than-expected performance given the processors that it's been trained on," a tech analyst for Morningstar told Business Insider.
But concerns persist about governance: operating within China’s regulatory framework raises questions about censorship and control over how the tools might be used globally. For instance, when asked about Tiananmen Square, DeepSeek replied: “Sorry, that’s beyond my scope,” a hallmark of China’s censorship regime.
Also, under Chinese law, DeepSeek must share user data with the government if asked. (Read: if the Chinese government requests your data, DeepSeek will hand it over.)
Either way, Chinese tech companies are showing that they are competitive in the space, which could democratize the race to AI leadership… and even AGI, in my opinion.
Why it matters
DeepSeek is sharing their basic technology freely (while charging for advanced features), which means smaller companies can build their own AI tools without relying on big tech companies. This is different from how US companies typically keep their technology private and charge for access.
As AI becomes more important in healthcare, jobs, and education, DeepSeek’s rise gives us a window into a fractured AI future where AI technology could originate from many different places and cultures, rather than mainly from Silicon Valley companies. While this could lead to more diverse and locally-adapted AI tools, it also raises questions about how to maintain safety standards and security across different countries.
To me, the good news is that DeepSeek's move suggests a future where AI solutions will be as varied as the communities they serve, rather than being filtered through Silicon Valley's worldview. The emerging multipolar world of AI is one, I hope, where power and innovation flow from multiple sources.
Biweekly disruptions
AI systems with ‘unacceptable risk’ are now banned in the EU (TechCrunch) Regulators to ban AI systems deemed an “unacceptable risk,” such as social scoring, emotion detection in schools, and real-time biometric surveillance. Companies violating these rules face fines of up to €35 million or 7% of their revenue, though enforcement won’t begin until August.
Why we need more women in the AI revolution (Time) “Artificial intelligence could become the most pivotal tool yet for advancing equality—or it could set us back decades. The choice depends on whether women (as well as anyone from marginalized backgrounds) are able to fully embrace and help shape the technology that is rapidly transforming our world.”
OpenAI to release new artificial intelligence model for free (The Guardian) OpenAI is releasing o3-mini, a free AI model, to counter competition from China’s DeepSeek, whose R1 model recently rattled tech investors. The move underscores the intensifying global AI race and questions about long-term profitability in the sector.
Strengthening America’s AI leadership with the U.S. National Laboratories (OpenAI press release) OpenAI is partnering with U.S. National Laboratories to provide 15,000 government scientists access to its AI models for research on nuclear security, cybersecurity, and disease treatment. OpenAI researchers with security clearances will consult on nuclear security projects focused on weapons safety and nuclear war risk reduction. The move shows the growing role of AI in US national security and signals its growing importance as a global government priority.
Help train this newsletter's neural networks with caffeine! ⚡️ Buy me a coffee ☕️