GPT-4, the fourth iteration of OpenAI’s Generative text system, has a fresh, multimodal language model. It was officially launched on March 14, 2023, and made accessible to ChatGPT Plus subscribers. Prior to its public release, Microsoft had already integrated GPT-4 into certain versions of Bing. As a model, GPT-4 underwent pre-training using a combination of public and licensed data, followed by fine-tuning through reinforcement learning guided by human feedback (this would be a horrible job).
OpenAI has refrained from disclosing the model size due to the competitive nature of the AI industry and potential safety concerns related to large-scale models. There were speculations that GPT-4’s parameters would increase dramatically from GPT-3’s 175 billion to 100 trillion, a claim that OpenAI CEO Sam Altman dismissed as false.
GPT-4 outperforms GPT-3.5 in terms of reliability, and the ability to process complex instructions. GPT-4 can work with up to 25,000 words of text, which marks a considerable advancement over previous versions. GPT-4 demonstrates significant accuracy improvements compared to GPT-3.5, developing the capacity to summarise and interpret images, and condense complex texts. However, it still occasionally generates bullshit!