DeepSeek

If you visit chat.deepseek.com, you’ll find DeepSeek, a Chinese-built AI chatbot that’s been making waves lately. Here’s a clear, human-friendly rundown of what it is, how it works, and why people are talking about it.

What is DeepSeek?

DeepSeek is an AI-powered chatbot, accessible through its website and mobile app. It’s based on two main models—R1 and V3—and can handle a wide range of tasks:

• Writing and editing content

• Helping with programming and debugging

• Reading and summarizing long documents

• Answering questions across many fields

You can even upload files for extended, context-rich conversations. It’s essentially a multi-purpose AI assistant.

The models are free to use via the site or app, though there’s also an API for developers, priced based on usage.

Who Built It?

DeepSeek is developed by a company in Hangzhou, China, founded in July 2023 by Liang Wenfeng. The company operates under High-Flyer, a Chinese hedge fund.

• Chatbot launch date: January 10, 2025

• Latest model updates:

  • R1 updated May 28, 2025

  • V3 launched March 24, 2025

How the Models Work

DeepSeek-V3 uses a Mixture-of-Experts (MoE) design, trained on about 14.8 trillion tokens. It has some clever cost-saving tricks, like:

• Multi-head Latent Attention for faster processing

• Sparse expert routing to reduce GPU load

R1 and earlier versions also use advanced architecture, with strong reasoning abilities. The result? High performance at a fraction of the usual training cost—plus much lower hardware requirements.

Why It’s Making Headlines

• Explosive adoption: By late January 2025, it had already become the most downloaded free app in the U.S. iOS App Store, surpassing ChatGPT.

• Extreme cost efficiency: Up to 96% cheaper per task than rivals, running on around 2,000 NVIDIA H800 GPUs.

The Controversies

Despite its success, DeepSeek has attracted criticism and concerns:

1. Privacy & Security

• U.S. agencies like the Pentagon and Navy banned it over fears about Chinese data laws.

• In South Korea, regulators blocked downloads after confirming data transfers to ByteDance in China.

2. Data Breach

• Early 2025: A misconfigured cloud database exposed over a million sensitive records, including chat histories and internal data.

3. Censorship

• Avoids certain sensitive topics (like Taiwan), aligning with Chinese government policies.

4. Safety Risks

• In academic safety tests, DeepSeek-R1 had a 100% success rate in adversarial attacks.

• Other studies point to bias, misinformation risks, and weaker safety alignment compared to some competitors.

Quick Summary

If you’re curious, I can also break down how it compares to ChatGPT or Claude, walk you through its features step-by-step, or dive deeper into the tech behind R1 and V3.