GPT-4

Free

Announcement of GPT-4, a large multimodal model. OpenAI blog, March 14, 2023.

FreeFree tier

Inputs: image, textOutputs: text

Type

Open Source

Company

OpenAI

Visit Website GitHub (40 stars)

About GPT-4

GPT-4 is OpenAI's latest milestone in scaling up deep learning. It is a large multimodal model that accepts image and text inputs and emits text outputs. While less capable than humans in many real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, such as passing a simulated bar exam with a score around the top 10% of test takers. Compared to GPT-3.5, GPT-4 is more reliable, creative, and able to handle much more nuanced instructions. It also demonstrates improved factuality, steerability, and adherence to safety guardrails through six months of iterative alignment. GPT-4 outperforms existing large language models on traditional ML benchmarks and shows strong multilingual capabilities, surpassing GPT-3.5 and other LLMs even in low-resource languages.

Key Features

Accepts image and text inputs, emits text outputs

Human-level performance on professional/academic benchmarks (e.g., top 10% on simulated bar exam)

More reliable, creative, and handles nuanced instructions than GPT-3.5

Improved factuality, steerability, and refusal to go outside guardrails

Strong multilingual capabilities across 26 languages tested

Accessible via ChatGPT Plus and API (text input initially)

Pros & Cons

Pros

Exhibits human-level performance on many professional benchmarks
Significantly more reliable and creative than predecessor GPT-3.5
Strong multilingual performance, including low-resource languages
Improved safety alignment and factuality through iterative training
Open-sourced evaluation framework (OpenAI Evals) for transparency

Cons

Still less capable than humans in many real-world scenarios
Subtle difference from GPT-3.5 in casual conversation; advantage emerges only with complex tasks
Image input capability initially limited to a single partner
Text input requires ChatGPT Plus subscription or API waitlist
Potential for inaccurate or biased outputs despite improvements

Best For

Professional tasks requiring complex reasoning and nuanced understandingAcademic benchmarking and exam preparationMultilingual translation and cross-lingual tasksInternal business functions like support, sales, and coding assistanceAdversarial testing and safety research through OpenAI Evals

FAQ

How does GPT-4 differ from GPT-3.5?

In casual conversation the difference is subtle, but GPT-4 is more reliable, creative, and able to handle much more nuanced instructions when tasks reach sufficient complexity.

What types of inputs and outputs does GPT-4 support?

GPT-4 is a multimodal model that accepts image and text inputs and emits text outputs.

How did GPT-4 perform on professional exams?

GPT-4 passed a simulated bar exam with a score around the top 10% of test takers, whereas GPT-3.5 scored around the bottom 10%.

Is GPT-4 available to the public?

Text input capability is available via ChatGPT Plus and the API (with a waitlist). Image input capability is initially available through a single partner.

What safety measures were taken for GPT-4?

OpenAI spent six months iteratively aligning GPT-4 using lessons from adversarial testing and ChatGPT, resulting in improved factuality, steerability, and refusal to go outside guardrails.