• Mindstream
  • Posts
  • Can a censored AI still be groundbreaking?

Can a censored AI still be groundbreaking?

AI CHATBOT

DeepSeek, a Chinese AI lab, has launched DeepSeek-R1, an open-source reasoning model it says performs as well as OpenAI’s o1 on certain benchmarks.

R1 is free to use for commercial purposes under an MIT license and is available on Hugging Face.

R1 outperforms OpenAI’s o1 on:

  • AIME: Assesses performance using other models.

  • MATH-500: Focuses on solving word problems.

  • SWE-bench Verified: Measures programming skills.

  • Self-Checking: R1 fact-checks its own answers, making it more reliable in science and math, though slower than non-reasoning models.

Model size and cost:

  • R1 has 671 billion parameters, which means it’s very capable. Smaller versions (from 1.5 billion to 70 billion parameters) can run on laptops.

  • Full R1 access is 90%-95% cheaper than OpenAI’s o1 via DeepSeek’s API.

Reasoning models, slow but steady

R1 is subject to Chinese government regulations and avoids topics like Tiananmen Square or Taiwan’s autonomy.

These restrictions reflect broader limits on AI systems in China.

The release follows U.S. proposals for stricter AI export rules, which could make it harder for Chinese companies to access advanced chips and software.

OpenAI has called for stronger U.S. investment in AI to stay competitive, highlighting concerns about Chinese labs like DeepSeek.

The AI cold war is hotting up…