Explained: OpenAI’s new GPT-4 AI model backed by Microsoft, its capabilities and limitations

Microsoft-backed startup OpenAI has begun rolling out GPT-4, a powerful model of artificial intelligence that succeeds the technology behind the hugely popular ChatGPT.

GPT-4 is “multimodal”, meaning it can generate content from image and text prompts.

What is the difference between GPT-4 and GPT-3.5?

GPT-3.5 only takes text prompts, while the latest version of the big language model can also use images as inputs to recognize objects in an image and analyze them.

GPT-3.5 is limited to around 3,000 words, while GPT-4 can generate responses longer than 25,000 words.

GPT-4 is 82% less likely to respond to requests for unauthorized content than its predecessor and scores 40% higher on some factuality tests.

This will also allow developers to decide on the tone and verbosity style of their AI. For example, GPT-4 can take on a Socratic conversational style and answer questions with questions. The previous iteration of the technology had a fixed tone and style.

Soon, ChatGPT users will have the ability to change the chatbot’s tone and response style, OpenAI said.

What are the capabilities of GPT-4?

The latest version outperformed its predecessor on the US Bar Examination and the Graduate Record Examination (GRE). GPT-4 can also help individuals calculate their taxes, showed a demonstration by Greg Brockman, president of OpenAI.

The demo showed he could take a photo of a hand-drawn mockup for a simple website and create a real one.

Be My Eyes, an app that caters to the visually impaired, will provide a GPT-4 powered virtual volunteer tool on its app.

What are the limitations of GPT-4?

According to OpenAI, GPT-4 has similar limitations to its previous versions and is “less capable than humans in many real-world scenarios.”

Inaccurate responses called “hallucinations” have been a challenge for many AI programs, including GPT-4.

OpenAI said GPT-4 can compete with human propagandists in many areas, especially when paired with a human editor.

He cited an example where GPT-4 offered suggestions that sounded plausible, when asked how to get two parties to disagree with each other.

OpenAI chief executive Sam Altman said GPT-4 was “most capable and aligned” with human values ​​and intent, though “it’s still flawed.”

GPT-4 is generally lacking in knowledge of events that occurred after September 2021, when the vast majority of its data was cut. Nor does he learn from experience.

Who has access to GPT-4?

Although GPT-4 can process both text and image input, only text input functionality will be available for ChatGPT Plus subscribers and software developers, with a waiting list, while the ability image capture is not yet publicly available.

The subscription plan, which offers faster response time and priority access to new features and improvements, launched in February and costs $20 per month.

GPT-4 powers Microsoft’s Bing AI chatbot and some subscription-level features of the Duolingo language learning platform.

© Thomson Reuters 2023