# GPT-4 Technical Report

**Authors:**  OpenAI  
**arXiv:** [2303.08774](https://arxiv.org/abs/2303.08774)  
**Date:** 2023-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.08774v1/

## Key points

- The GPT-4 Technical Report discusses the creation of a large-scale multimodal model capable of processing both image and text inputs to generate text outputs.
- GPT-4 has shown impressive performance on professional and academic benchmarks, ranking among the top 10% of test takers in a simulated bar exam.
- It is a Transformer-based model pre-trained to predict the next token in a document, with performance enhanced through post-training alignment processes for factuality and desired behavior.
- Infrastructure and optimization methods were developed to ensure consistent behavior across different scales during the model's creation.
- Models trained with significantly less compute power than GPT-4 can accurately predict certain aspects of its performance, suggesting scalable and predictable patterns within the model's functioning.
- The report provides valuable insights into GPT-4's development, strengths, and potential applications across various domains.

## Summary

The GPT-4 Technical Report by OpenAI delves into the creation of a large-scale multimodal model with the unique ability to process both image and text inputs and generate text outputs. While it may not match human capabilities in all real-world scenarios, GPT-4 has demonstrated impressive performance on professional and academic benchmarks. In fact, it achieved a score placing it among the top 10% of test takers in a simulated bar exam. Described as a Transformer-based model pre-trained to predict the next token in a document, GPT-4's performance is further enhanced through post-training alignment processes that ensure factuality and adherence to desired behavior. The development of this advanced model also involved creating infrastructure and optimization methods that exhibit consistent behavior across various scales. One notable finding highlighted in the report is the ability to accurately predict certain aspects of GPT-4's performance using models trained with just 1/1,000th of the compute power required for GPT-4 itself. This suggests that there are scalable and predictable patterns within the model's functioning that can be leveraged for efficient training and evaluation processes. Overall, the GPT-4 Technical Report provides valuable insights into the development and capabilities of this advanced multimodal model. It sheds light on its strengths and potential applications in various domains.