GPT-4

GPT-4. It is a multimodal large language model created by OpenAI and released in March 2023. Representing a significant advancement over its predecessor, GPT-3, it demonstrates improved performance on a wide array of professional and academic benchmarks. The model's development involved extensive training and safety research conducted by teams at OpenAI in collaboration with partners like Microsoft.

Development and release

The development of this model was led by researchers at OpenAI, building upon years of work on the Generative pre-trained transformer series. Key figures involved in its creation included Sam Altman and Ilya Sutskever. It was officially announced and released to the public on March 14, 2023, with initial access provided through the ChatGPT Plus subscription service and an API for developers. The training process utilized immense computational resources, leveraging the Azure supercomputing infrastructure provided by Microsoft. This release followed a period of extensive internal testing, known as red-teaming, to identify potential risks.

Architecture and capabilities

While specific architectural details are proprietary, it is known to be a transformer-based model significantly larger and more complex than GPT-3. A primary advancement is its multimodal nature, capable of processing both text and image inputs, though initial public releases focused on the text modality. The model exhibits enhanced capabilities in areas like computer code generation, complex creative writing, and nuanced instruction following. Its training incorporated a broader and more diverse dataset, including sources from multiple languages and technical domains, curated by teams at OpenAI.

Performance and evaluation

In standardized evaluations, it achieved high scores on numerous benchmarks, often surpassing prior models. It performed impressively on exams designed for humans, such as the Uniform Bar Exam, the LSAT, and the SAT. In machine learning competitions like MMLU and HellaSwag, it set new state-of-the-art records at the time of its release. Independent analyses by groups like the Allen Institute for AI and researchers at Stanford University confirmed its advanced reasoning abilities. However, evaluations also noted persistent issues with factual accuracy, sometimes called "hallucinations," and biases present in its training data.

Applications and impact

Its integration into ChatGPT revolutionized public interaction with artificial intelligence. Major technology companies, including Duolingo and Morgan Stanley, deployed it to enhance their services. It has been used for tasks ranging from drafting legal documents for firms like Baker McKenzie to assisting with programming on platforms like GitHub Copilot. The model's capabilities accelerated discussions about automation in fields like journalism, education, and software engineering, influencing policy debates in institutions like the European Parliament and the United States Congress.

Ethical considerations and limitations

The deployment raised significant ethical concerns identified by organizations like the Partnership on AI and researchers at the University of Cambridge. Primary issues include the potential for generating disinformation, embedded societal biases, and threats to certain job markets. Its limitations include a lack of true understanding, a knowledge cutoff date, and vulnerability to adversarial prompts. In response, OpenAI established a Superalignment team and engaged with bodies like the UNESCO to address global governance challenges. Ongoing scrutiny from entities like the Federal Trade Commission highlights the regulatory attention focused on such powerful systems.

Category:Artificial intelligence Category:Natural language processing Category:2023 software