In a blog post on Tuesday, OpenAI wrote, "We have created GPT-4, the next milestone in our attempts to expand deep learning.
"We spent 6 months iteratively aligning GPT-4 using lessons from our adversarial test software, as well as ChatGPT, resulting in the best results with respect to factuality, steerability, and refusal to break the barrier."
Compared to GPT-3.5, the new AI model is more reliable, more creative, and capable of handling complex instructions.
GPT-4 outperforms existing extensive language models (LLMs), including more modern models (SOTA), which may include additional standards-specific constructs or training methods.
The company claimed that GPT-4 outperformed GPT-3.5 English and other LLMs (Chinchilla, PaLM) in 24 out of the 26 examined languages, including low-resource languages like Latvian, Welsh, and Swahili.
The company is also using this new model internally, with a significant impact on functions like support, sales, content modification, and scheduling.
Unlike a text-only setup, this form can accept a claim with text and images, allowing users to define any language or vision task.
The GPT-4 base model, like previous GPT models, was taught to predict the next word in a document. It was trained using authoritative and publicly available data.
ChatGPT Plus subscribers will get access to GPT-4 at chat.openai.com with a usage limit, while developers can subscribe to the GPT-4 API queue.
"We expect GPT-4 to become a valuable tool in improving people's lives by powering a multitude of applications," the company said.
Also Read: Koo launches new feature to enable creators to compose posts using ChatGPT