Changelog Master Feed Deep-dive into DeepSeek (Practical AI #302)
11 snips
Jan 31, 2025 The podcast dives into the buzz around DeepSeek's new AI model, DeepSeek R1, unpacking its rising popularity amid privacy and geopolitical concerns. It addresses how this model contrasts with open science principles and sheds light on evolving perceptions of AI accessibility and security. The hosts explore the implications of AI model biases and prompt injection attacks, while also examining the training architecture of DeepSeek. Lastly, they discuss the shifting landscape of enterprise budgets and the necessity for businesses to adapt their AI investments.
AI Snips
Chapters
Transcript
Episode notes
DeepSeek's Disruptive Model
- DeepSeek R1, a Chinese LLM, rivals OpenAI's models in performance.
- It was trained at a significantly lower reported cost, surprising many.
Tiananmen Square Test
- Chris Benson avoided using Chinese tech due to his job but tested DeepSeek R1.
- When asked about Tiananmen Square, it replied with a canned, harmless response.
DeepSeek's Open-Source History
- DeepSeek, while a smaller company, has a history of open-source work.
- They've released models on Hugging Face, unlike OpenAI, promoting accessibility.