In this week's newsletter

What we’re talking about: How DeepSeek's AI model is reshaping our understanding of AI development, challenging assumptions about compute requirements, and sparking geopolitical tensions.

How it’s relevant: With a claimed training cost of just $5M (compared to the usual $100M+) and performance rivaling industry leaders, DeepSeek demonstrates that sophisticated AI development might be more accessible than previously thought – this challenges an industry dominated by tech giants.

Why it matters: This requires a reassessment of fundamental assumptions about AI's future: the necessity of massive computing resources, the effectiveness of export controls, and the balance between open-source and proprietary models. The implications extend far beyond technology into geopolitics, market valuations, and global innovation patterns.

Big tech news of the week…

⚖️ The first part of the EU AI Act came into effect on February 2, including AI literacy requirements and the prohibition of certain AI systems.

🖥️ Google has removed their pledge to not use AI for weapons or surveillance from the website.

📱 Does AI affect our memory? It’s a valid question to ask, and Nature Magazine is exploring the answers.

1. The Geopolitical Dimension

While DeepSeek operates independently without apparent state backing, its success has significant geopolitical implications. It has sparked fears in the U.S. national security community that the United States’ most advanced AI products may no longer be able to compete against cheaper Chinese alternatives, along with other questions relating to the U.S./China rivalry.

The model's responses on sensitive topics reflect China's information control policies, raising questions about its global adoption. However, the bigger story might be how DeepSeek challenges assumptions about technological innovation – demonstrating that breakthrough AI development can happen anywhere, potentially reshaping the global AI landscape.

2. The Technological Leap: Is AI reaching a plateau?

Just when discussions about AI reaching a plateau were gaining momentum (we wrote about this in our December newsletter), DeepSeek demonstrates that innovation might come from unexpected directions. Rather than following the traditional path of throwing more computing power at the problem, DeepSeek's team found creative ways to achieve similar benchmark results as the main players with fewer resources: 

Using innovative techniques like Mixture-of-Experts (MoE) architecture and reinforcement learning, they've shown that AI development isn't plateauing – it's evolving, with efficiency innovations potentially becoming as important as raw computing power. Want to know more? This article is a great place to start.

3. Markets shaking: Beyond the Numbers 

The market's dramatic reaction to DeepSeek goes deeper than this one company: It’s a fundamental reassessment of the assumptions of the AI industry as a whole. When tech stocks tumbled following DeepSeek's announcement, it wasn't merely market volatility; it was investors grappling with three realizations:

• The moats protecting major AI companies might be shallower than thought

• The value of massive computing infrastructure might be overestimated

• The path to advanced AI might be more accessible than previously assumed

At the same time, we’ve never heard “Jevons Paradox” (the idea that efficiency leads to more use of a resource, not less) being mentioned so many times over such a short period of time: This means that improved efficiency in AI may boost its demand and make it a commodity.

4. The Data Privacy Concerns and OpenAI’s Copyright Claims 

Suddenly, OpenAI seems to care about copyright. That’s right. The very same company that has been at the center of intellectual property disputes, is now making allegations about DeepSeek using OpenAI's data for training (which also raises the question on whether OpenAI actually owns their data, but we’re saving that discussion for another time). This claim, as all IP right claims related to AI, highlights the complex nature of AI development, and how outdated legal systems are not designed for these technological and societal developments. 

Moreover, the claims also reflect a common misunderstanding of how large language models work. The reality is more nuanced: Similar outputs don't necessarily indicate data theft, as models can generate similar responses through different training approaches.

In addition, researchers can download DeepSeek to their own servers and run and build on it, also for free, which is a plus for privacy. This controversy underscores the need for clearer standards in AI development and training data transparency.

5. “Is DeepSeek better than OpenAI?”: What the research tells us

While DeepSeek's raw numbers are impressive, matching o1’s performance in preliminary tests on data-driven scientific tasks – scoring 97.3% on the MATH 500 benchmark and ranking in the 96.3rd percentile on Codeforces – the full picture reveals important nuances. Recent security research exposed significant gaps in safety measures, with the model failing to block harmful content that other leading AI systems routinely filter. The model excels at mathematical and logical reasoning but shows slightly lower scores on general knowledge tests like MMLU (90.8% compared to OpenAI's 91.8%). This suggests that while DeepSeek has made remarkable strides in efficient AI development, particularly in reasoning tasks, it still faces challenges in creating a comprehensive and safe AI system that matches industry standards.

Additionally, let’s remember that the model shows clear signs of Chinese censorship, refusing to address topics like the Tiananmen Square events or providing state-aligned responses on sensitive issues like Taiwan. It's unclear whether this is built into the model or is applied to its interface (probably a combination of both). The model also fails on some relatively simple tasks, such as counting the number of US state names that contain the letter W. This is not unique to DeepSeek: Examples like these are known to also trip up other large-language models.