The improved iteration of GPT-4 is here. Announced last November during the OpenAI developers conference, this latest version shows, at first glance, notable performances, surpassing – among others – those of Claude 3 Opus or Gemini Ultra 1.0 and Gemini Pro 1.5, promises OpenAI. Now available to users subscribed to the paid version of ChatGPT, GPT-4 Turbo shows improvements in writing, as well as mathematics, logical reasoning and coding.
The version dedicated to ChatGPT obtains the following results on different assessment tests : 72.2% on the MATH benchmark (compared to 63.2% for Claude 3 Opus and 58.5% for Gemini Pro 1.5), 86.5% on the MMLU test (compared to 84.1% for Claude 3 Opus and 83 .7% for Gemini Ultra 1.0) and 87.6% on the HumanEval test (compared to 84.8% for Claude 3 Opus and 74.4% for Gemini Ultra 1.0).
Advertisement
Less verbose responses in ChatGPT
The start-up particularly points to ChatGPT's ability to respond (still in the paid version) in a more direct, less verbose way and using more conversational language. Unlike GPT-4, GPT-4 Turbo's pop-up window is longer (128,000 tokens versus 8192) and can therefore contain the equivalent of over 300 pages of text in a single prompt.
Note also that the LLM was trained on data up to December 2023 – compared to September 2021 for GPT-4. Finally, OpenAI specifies that it has also optimized its performance in order to be able to offer GPT-4 Turbo at a price 3 times cheaper for input tokens and at a price 2 times cheaper for output tokens compared to GPT-4. For those who would like to use it, GPT-4 Turbo is therefore available in ChatGPT Plus, Team, Enterprise as well as via API.
OpenAI has, in parallel, published on April 14 a personalized GPT-4 model optimized for the Japanese language – as part of opening its office in Tokyo, Japan – which provides improved performance in Japanese text and runs up to 3x faster than GPT-4 Turbo.
Advertisement
Vision integrated into large language models
At the same time, OpenAI unveiled “GPT-4 Turbo with Vision”, a large language model with vision capabilities. GPT-4 Turbo with Vision allows the model to take images and answer questions about them. Historically, LLMs have been limited by the adoption of a single input modality, namely text.
For many use cases, this limited the areas in which models like GPT-4 could be used. Previously, the model was sometimes called GPT-4V or gpt-4-vision-preview in the API. The start-up specifies that images are made available to the model in two main ways: by passing a link to the image or by passing the base64 encoded image directly in the request. Images can be transmitted in “user”, “system” and “assistant” messages.
Limits relating to detection in images
“The model is best for answering general questions about what is present in images. While it understands the relationship between objects in images, it is not yet optimized for answering detailed questions about the location of certain objects in an image”, specifies OpenAI.
Taking the example of “what color is this car” or “what dinner ideas might be based on what's in your fridge”, the company insists that the tool does not is not able to answer correctly. “It is important to keep the limitations of the model in mind when exploring the use cases to which visual understanding can be applied.” As for its price, it depends on the size of the input image. For example, transmitting a 1080 × 1080 pixel image to GPT-4 Turbo costs $0.00765.
Vision, the new workhorse of AI start-ups
OpenAI is not the only one working on this vision function. The start-up xAI led by Elon Musk has also made it its hobby horse. On April 13, 2024, it unveiled Grok-1.5V, its multimodal model. In addition to its text capabilities, Grok can now process a wide variety of visual information, including documents, diagrams, graphs, screenshots and photographs, the company says.
Its performance is very close to that of GPT-4V. Initially accessible to beta testers, xAI plans to bring Grok-1.5V to all Grok users subsequently.
Do you want to stay up to date on the latest news in the artificial intelligence sector? Register for free to the IA Insider newsletter.
Selected for you
