The Generative Pre-trained Transformer 4 (GPT-4) is the latest addition to OpenAI’s GPT series, released on March 14, 2023. This multimodal large language model is a significant upgrade from its predecessor, GPT-3.5, with improved reliability, creativity, and the ability to handle more nuanced instructions. In this blog post, we’ll explore GPT-4’s capabilities, applications, and some controversies surrounding its release.
Training and Capabilities
GPT-4 offers two versions with context windows of 8192 and 32768 tokens, outperforming GPT-3.5 and GPT-3, which were limited to 4096 and 2049 tokens, respectively. Unlike its predecessors, GPT-4 can process both text and images, enabling it to describe humor in unusual images, summarize screenshotted text, and answer exam questions containing diagrams. However, GPT-4 still shows a tendency to hallucinate answers.
The model has demonstrated aptitude on standardized tests, scoring in the top percentiles on the SAT, LSAT, and Uniform Bar Exam. To enhance control over GPT-4, OpenAI introduced “system messages” that can direct the model’s tone and tasks.
GPT-4’s Technical Details and Security Controls
OpenAI took a closed approach to GPT-4’s technical details, not specifying the model size, architecture, or hardware used during training or inference. The decision to limit this information was influenced by the competitive landscape and safety implications of large-scale models. Sam Altman, CEO of OpenAI, visited Congress in January 2023 to demonstrate GPT-4 and its improved security controls compared to other AI models.
Reception and Criticisms
The release of GPT-4 has been monitored closely by EU institutions as they propose the introduction of the Artificial Intelligence Act. Some AI researchers have criticized OpenAI’s decision to reveal fewer technical details, arguing that it hinders open research into GPT-4’s biases and safety.
Capabilities and Usage
GPT-4 has been tested on medical problems, exceeding the passing score on the USMLE and outperforming earlier general-purpose models and those specifically fine-tuned on medical knowledge. Researchers have suggested that GPT-4 could be viewed as an early version of artificial general intelligence (AGI).
GPT-4 has been integrated into various applications, including Microsoft’s Bing and Microsoft 365 Copilot, Duolingo’s language learning app, Icelandic start-up Miðeind ehf, Khan Academy’s tutoring chatbot “Khanmigo,” Be My Eyes’ “Virtual Volunteer” feature for visually impaired individuals, and GitHub’s Copilot X.
Conclusion
GPT-4 marks a significant step forward in AI language models, with its enhanced capabilities and applications spanning various industries. Although its release has generated some controversy regarding the disclosure of technical details, GPT-4’s impact on the AI landscape is undeniably transformative. As AI continues to advance, it will be crucial for researchers, developers, and policymakers to navigate the challenges and opportunities these powerful tools present.
