GPT-4o mini surpasses GPT-3.5 Turbo on LLM benchmarks with enhanced resistance to jailbreaks using an instruction hierarchy method, offering improved robustness and defense against system prompt extraction.
The model supports multiple languages and modalities similar to GPT-4o, with upcoming audio and video input/output. It maintains safety features, a 128k token context window, and training knowledge till October 2023.
OpenAI aims to reduce costs, enhance model capabilities, and integrate AI models universally, leading to more accessible and reliable AI applications. GPT-4o mini facilitates efficient and affordable development of powerful AI apps.
While technical details are limited, OpenAI introduced an instruction hierarchy method to train models, aiming to improve their resilience to attacks exploiting the equal priority given to system prompts and untrusted user text.
[
Collection
]
[
|
...
]