OpenAI GPT-4 Release

OpenAI has released GPT-4, a large language model that exhibits significantly improved reasoning capabilities, broader general knowledge, and higher accuracy in solving complex problems compared to its predecessor, GPT-3.5. This release marks a continuation of OpenAI's scaling strategy, leveraging increased data and computation to create more sophisticated deep learning models.

Advanced Reasoning and Problem Solving

GPT-4 demonstrates superior performance in complex reasoning tasks and creative collaboration. It can generate, edit, and iterate on technical and creative writing tasks, such as composing songs or writing screenplays, and can adapt to a user's specific writing style.

In practical reasoning tests, GPT-4 outperforms GPT-3.5 in scheduling and logic problems. For example, in a scenario requiring the coordination of three people's availability for a 30-minute meeting, GPT-4 correctly identifies the common availability window (12 pm - 12:30 pm), whereas GPT-3.5 fails to identify the correct window.

Standardized Testing Performance

GPT-4 shows a marked improvement in scoring percentiles among test-takers on professional and academic exams:

  • Uniform Bar Exam: GPT-4 scored in the 90th percentile, compared to the 10th percentile for GPT-3.5.
  • Biology Olympiad: GPT-4 (with vision) scored in the 99th percentile, compared to the 31st percentile for GPT-3.5.

Safety and Alignment Improvements

OpenAI spent six months focusing on making GPT-4 safer and more aligned. According to internal evaluations, GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5.

These improvements were achieved through several key methods:

  • Training with Human Feedback: The model incorporated more human feedback, including data submitted by ChatGPT users, and early feedback from over 50 experts in AI safety and security.
  • Real-World Application: Lessons from the real-world deployment of previous models were integrated into the safety research and monitoring systems.
  • GPT-4-Assisted Research: The model's own reasoning and instruction-following capabilities were used to create training data for fine-tuning and to iterate on classifiers for training and monitoring.

Infrastructure and Limitations

GPT-4 was trained on Microsoft Azure AI supercomputers, which also facilitate its global delivery to users.

Despite these advancements, OpenAI acknowledges that GPT-4 has known limitations, including social biases, hallucinations, and susceptibility to adversarial prompts. The company aims to address these issues through transparency, user education, and expanded input from the public to shape future models.

Availability

GPT-4 is available to users via ChatGPT Plus and as an API for developers to build applications and services.

Sources