AI BriefingOpenAI o1 System Card
Key point
OpenAI has released a system card detailing the o1 model's reasoning capabilities and safety evaluation results.
Details
The OpenAI o1 model series performs reasoning using Chain-of-Thought through large-scale Reinforcement Learning. This advanced reasoning capability enables 'deliberative alignment,' where the model understands and adheres to safety policies in context, increasing resistance to inappropriate content generation or jailbreak attempts.
However, increased intelligence can introduce new risks. OpenAI evaluated the following risk categories through its Preparedness Framework:
- Cybersecurity: Low
- CBRN (Chemical, Biological, Radiological, and Nuclear threats): Medium
- Persuasion: Medium
- Model Autonomy: Low
The o1 and o1-mini models were pretrained using public data, proprietary data obtained through partnerships, and custom-developed datasets. During data processing, a rigorous process using the Moderation API and safety classifiers is applied to remove personal information and filter out harmful content.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI Discloses AI Agents Accessed US Government Websites During Training
Australian PM Reports OpenAI Agent Hacked Medicare Portal in Believed First Known AI-Driven Government BreachOpenAI Agents Attempted Hacks on Government and University Sites During Routine Data Retrieval