OpenAI proudly announced the arrival of GPT-4, a cutting-edge multimodal model designed to handle both text and image inputs, while generating text outputs. Although GPT-4 falls short of human capabilities in various real-world scenarios, it has achieved human-level performance on multiple professional and academic benchmarks. Remarkably, it even managed to surpass the performance of most test takers by securing a score in the top 10% on a simulated bar exam.
GPT-4 operates on a Transformer-based model and is pre-trained to predict the next token in a document. To enhance its reliability and adherence to desired behavior, a post-training alignment process was employed, resulting in improved measures of factuality. Notably, the development of GPT-4 required the creation of infrastructure and optimization methods capable of delivering consistent performance across diverse scales. Intriguingly, certain aspects of GPT-4's performance could be accurately predicted using models trained with only a minuscule fraction of its compute power (approximately 1/1,000th).
With GPT-4, OpenAI showcases a truly impressive achievement that pushes the boundaries of AI capabilities, making us question just how much work we'll have left for ourselves in the future.
Source: Open AI URL: https://cdn.openai.com/papers/gpt-4.pdf