Reproduce “GPT-4o System Card” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
OpenAI, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow and 410 more
GPT-4o generates and receives multiple modalities through one model, with audio responses averaging 320 milliseconds. OpenAI adapted its safety work to voice through post-training and deployed classifiers. In a study of more than 3,800 participants, voice clips shifted opinions by 78% of the human clips’ effect size; voice was classified low risk under the study’s threshold, not universally harmless.
Key insights
- Achieving near human-level audio response times (320ms) is possible in a natively multimodal AI model.
- Integrated multimodal training enhances performance in traditionally challenging areas, such as non-English languages and visual understanding.
- Rigorous, multi-faceted safety evaluations are essential for advanced AI deployment, categorizing the model's risk level (e.g., medium risk for GPT-4o) and guiding mitigation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.