AI Safety Newsletter AISN #47: Reasoning Models
7 snips
Feb 6, 2025 A new frontier reasoning model, DeepSeek-R1, is making waves in AI with its impressive capabilities in mathematics, coding, and scientific reasoning. Meanwhile, state-sponsored AI cyberattacks pose significant challenges, as over 20 countries leverage advanced technology for cyber warfare. The conversation also highlights fresh developments in AI safety and regulation, featuring groundbreaking frameworks and government initiatives. It's a thought-provoking exploration of how AI is reshaping our world.
AI Snips
Chapters
Transcript
Episode notes
DeepSeek-R1's Impact
- DeepSeek-R1, a new reasoning model, has shown impressive capabilities, impacting NVIDIA's stock.
- It performed well on benchmarks and was developed with relatively low compute costs.
DeepSeek's Cost Efficiency
- DeepSeek claims to have trained V3, the predecessor to R1, for only $6 million in compute costs.
- However, this excludes full development costs and potential use of OpenAI's models for distillation.
OpenAI's Response and Deep Research
- OpenAI responded to DeepSeek-R1 by releasing O3-Mini with varying reasoning efforts and Deep Research with tool use.
- Deep Research, combining O3 with online research tools, achieved higher scores but lacks a public system card, seemingly violating OpenAI's safety commitment.
