What is GPT-4?
GPT-4 is the latest large language model (LLM) from OpenAI. It is the successor of GPT-3.5, which powers OpenAI’s ChatGPT. GPT-4 (Generative Pre-Trained Transformer, version 4) is a multimodal model. It can receive prompts via text and/or images, to which it produces responses in text form. GPT-4 was released on March 14, 2023. It is currently available for use via ChatGPT Plus and an API.
How is GPT-4 different from GPT-3.5?
According to OpenAI, GPT-4 delivers the following improvements over GPT-3.5:
- GPT-4 scored in the top 10% on a simulated bar exam, while GPT-3.5 scored in the bottom 10%
- GPT-4 is far less likely (82%) to produce disallowed content, due to extra reward signals included in the RLHF (Reinforcement Learning from Human Feedback) phase of training
- GPT-4 is more factually accurate (40%) and less likely to hallucinate
- GPT-4 is more capable of understanding complex tasks and nuanced instructions
- GPT-4 can respond reliably to prompts with a combination of text and images such as photos, diagrams, and screenshots
What are the limitations of GPT-4?
GPT-4’s limitations are similar to those of GPT-3.5; however, due to improvements in general knowledge and processing abilities, these limitations have reduced in degree. GPT-4 still exhibits the following issues:
- Factual hallucinations and reasoning errors
- Biased outputs
- Very limited knowledge of events that occurred after September 2021
- Does not learn from experience
- May accept patently false statements from users
As with all LLMs, carefully-written prompts are key to maximizing usefulness and minimizing errors in model outputs. Although LLMs in general are a work in progress (especially in terms of safeguards), GPT-4 is nonetheless poised to continue and expand the conversational AI revolution and embed itself in multiple facets of modern life.