This is educational material and does not constitute legal advice nor is any attorney/client relationship created with this article, hence you should contact and engage an attorney if you have any legal questions. No warranties, express or implied, are made with respect to its accuracy. Information contained herein, or information relied upon, is subject to change without notice.

What is Deepseek AI?

Deepseek is an AI model trained on 67 billion parameters, placing it amongst the best, and achieving a 79.8% score on advanced math tests. It matches or even exceeds the performance of GPT-4 in several benchmarks. It’s currently (as of January 27, 2025) topping the charts on the iOS App Store.

Built for (Super) Cheap

One of the most striking features of Deepseek is its cost-effectiveness. The model was developed for just $5.6M in computing power, significantly less than the typical $100M+ for similar models. This efficiency extends to its usage, with API costs up to 90% less per million tokens than leading providers, making AI implementation more accessible for businesses and potentially for legal practices as well. Practically speaking, early tests are suggesting a 25x cost reduction versus ChatGPT-o1.

How Was it Built?

Deepseek’s innovative approach uses pure reinforcement learning, reducing the need for massive labeled datasets, and they’ve managed with fewer GPU hours (2.78M vs Meta’s Llama, for instance, that required some 30M hours). They’ve also found creative ways to navigate around China’s chip restrictions which has allowed them to keep costs low through less compute and data usage.

When comparing Deepseek-R1 with other models like GPT-4, Claude 3, and Gemini, Deepseek-R1 stands out with a 90.8% MMLU score (Massive Multitask Language Understanding) a context window of 192K tokens (enabling very long conversations in the same chat session), and incredibly low costs at $0 for input and $0.28 for output per 1M tokens. In contrast, ChatGPT costs $60.00 per 1M output tokens for GPT-4 and $30.00 per 1M output tokens for GPT-3.5 Turbo, making Deepseek-R1 significantly more cost-effective. Its key strengths lie in math and reasoning tasks, making it potentially useful for legal analytics and research.

Can it be Trusted?

There’s definitely reason for caution. Deepseek’s terms grant the company broad rights over content users submit, including the ability to modify, publish, and sublicense it, which raises privacy and intellectual property concerns. Deepseek also claims ownership of AI-generated outputs, limiting users’ control over their content.

For a deep dive privacy analysis into Deepseek, visit the AI Privacy Guide.

Global Impact and Concerns

The rise of Deepseek seems to have caused a significant sell-off in AI-linked stocks as investors reassess the landscape. Industry leaders have likened Deepseek’s emergence to AI’s “Sputnik moment,” signaling a potential shift in the power dynamics of AI technology. Moreover, shortly after its launch, Deepseek faced a “large-scale malicious attack,” prompting temporary restrictions to only users with Chinese phone numbers, highlighting cybersecurity concerns. Additionally, the use of older Nvidia chips for training, which are less subject to export controls, underscores the complex interplay between technology, international trade policies, and AI development.

Can I Test It?

Yes, it’s easily accessible at deepseek.com and on mobile. However, users should be aware of its rules regarding content restrictions. Despite some content limitations, many find the platform fascinating and worth exploring for its potential.

If you’d like to discuss, please contact us here.