1sec.ai

Tag

#model-transparency

Every item tagged model-transparency, newest first.

2 items

modelsDec 11

Update to GPT-5 System Card: GPT-5.2

OpenAI released an update to the GPT-5 System Card for GPT-5.2, detailing safety mitigations and training data sources. The GPT-5.2 model family was trained on publicly available internet data, third-party data, and user-generated content. Builders should review the updated System Card to understand data provenance and safety measures. The update reflects OpenAI's ongoing efforts to improve model transparency.

Key takeaways
  • GPT-5.2 trained on publicly available internet data, third-party data, and user-generated content.
  • Safety mitigations largely unchanged from GPT-5 and GPT-5.1.
  • Updated System Card provides transparency on data and safety measures.
OOpenAI#model-transparency#safety-mitigations#training-data
researchDec 3

How confessions can keep language models honest

OpenAI researchers are testing a method called 'confessions' to train models to admit mistakes or undesirable behavior, aiming to improve AI honesty and transparency. This approach helps models acknowledge errors and provides more accurate outputs. You can apply this method to train models to be more transparent and trustworthy. The goal is to increase trust in model outputs.

Key takeaways
  • OpenAI tests 'confessions' method to train models to admit mistakes.
  • Aims to improve AI honesty, transparency, and trust in model outputs.
  • Method helps models acknowledge errors for more accurate outputs.
OOpenAI#ai-safety#model-transparency#honesty