DeepSeek R1 broke the Internet just 5 days ago. People can't believe how powerful and "intelligent" it is. Unlocking new possibilities. 10 wild examples:

Many people have a filter regarding companies like OpenAI and Anthropic, believing that they define the direction and path of AGI development. DeepSeek proves that without supervised learning, process rewards, or MCTS, long CoT processes can naturally emerge through RL training!

One point that many people overlook is that Deepseek's greatest strength lies not in its capabilities rivaling OpenAI's O1, but in its price being 96.4% cheaper than ChatGPT, thanks to the MoE architecture and MLA mechanism that DeepSeek has consistently adhered to.

The power of simple RL.

Excellent research! It also demonstrates the effectiveness of Deepseek R1 naturally "growing" longer outputs.

It's hilarious how some folks with the processing power of a pocket calculator still insist DeepSeek is lying about their training costs.