Exploring the Capabilities of GPT-4 Turbo
Eight months after unveiling GPT-4, OpenAI has made another leap forward with the release of GPT-4 Turbo. This new iteration, introduced at OpenAI’s inaugural developer conference, stands out as a substantial upgrade in artificial intelligence technology.
What is GPT-4 Turbo?
GPT-4 Turbo is the latest Large language model released by Open AI following the introduction of GPT-4. You can read more about GPT-4 here. So, what can GPT-4 Turbo do?
Enhanced with Greater Knowledge and More Current Information
A key enhancement in GPT-4 Turbo compared to its predecessor is its extensive knowledge base. Unlike the original GPT-4, which incorporated data until September 2021, GPT-4 Turbo includes data up to April 2023. This update equips the model with 19 more months of information, significantly enhancing its understanding of recent developments and subjects.
Here is a response based on GPT-4-Turbo that is updated till April 2023.
However, GPT-4 still has an older cutoff date as shown below:
Longer Context Window
GPT-4 Turbo boasts a remarkable 128k token context window which is capable of processing around capable of processing the equivalent of around 96000 words in a single prompt, a feature that dramatically expands its data handling capacity from the previous capacity of 32,000 tokens. However, one limitation with this is the output is still limited to 4000 tokens. Claude by Anthropic (available on AWS) is another model that boasts of a similar context length limited to 100k tokens. Read about Claude here.
The enhanced context window not only prepares applications for future advancements but also allows for more complex interactions with a reduced likelihood of the model losing track of the conversation. In applications like chatbots, digital assistants, educational systems, and other scenarios involving extended exchanges, this expanded context capacity marks a significant breakthrough.
Multimodal Abilities (in Preview)
GPT-4 Turbo enhances its predecessor’s capabilities by introducing multimodal functions, enabling it to process images. This means you can now feed images into GPT-4 Turbo for automatic caption creation, visual content analysis, and text recognition within images.
Here is an example of a graph from our Insight generation tool — The Intelligencer which I passed to GPT-4 Turbo for analysis:
Prompt: Explain the diagram with a summary and detailed insights.Image Input:
Output
‘The image displays a bar chart titled “Current and Prominent IT Research Domains in Academic Circles (Jan 2023 — Sep 2023).” It visualizes the total number of publications in different IT research categories within academia during the specified period.
Summary: The chart lists the total number of publications for each category, indicating the level of research activity and prominence within the field of Information Technology. The X-axis lists the research categories, and the Y-axis shows the total number of publications.
Key Insights:
- Machine Learning is the leading research area, with the highest number of publications (7,441).
- Computer Vision and Pattern Recognition come in second with a significant number (6,840 publications).
- This is followed by Computational and Language with 3,555 publications and Robotics with 2,247 publications.
- Cryptography and Security is also a prominent research area with 2,210 publications.
- Artificial Intelligence has 1,946 publications, suggesting considerable interest and research output.
- Information Theory (1,688), Image and Video Processing (1,399), and Software Engineering (1,349) are also represented, indicating active research but with fewer publications compared to the top categories.
- Human-Computer Interaction appears to have the least number of publications within the period, totaling 1,349.
Overall, the diagram highlights that machine learning, computer vision, and pattern recognition are the most researched categories within academic circles during this period, reflecting their importance and relevance in current technological advancements and academic pursuits.
This advancement is currently in preview however visual comprehension greatly broadens the scope for computer vision applications and also moves closer to human ways of perceiving and engaging with the environment. OpenAI plans to integrate vision capabilities into the primary GPT-4 Turbo model during its official launch. Currently, the API allows input of images in base64 format or a direct URL of the image to be passed as input. The cost for this feature will vary based on the size of the input image. For example, processing an image of 1080×1080 pixels through GPT-4 Turbo will be priced at $0.00765.
Better Pricing
The pricing for GPT-4 Turbo is set at $0.01 per 1000 input tokens and $0.03 per 1000 output tokens. This reflects a threefold decrease in the cost of input tokens and a twofold decrease in the cost of output tokens, compared to the original GPT-4’s pricing structure as well as Claude's 100k model.
Here is how it compares against other models:
Additional Technical Features
Feature: Enhanced Instruction Compliance and JSON Mode
GPT-4 Turbo surpasses earlier models in executing tasks that demand precise adherence to instructions, particularly in generating designated formats (like consistently responding in XML). It also introduces the innovative JSON mode, guaranteeing valid JSON responses. This is facilitated by the new API parameter, ‘response_format’, which directs the model to produce syntactically accurate JSON objects.
Here is an example of generating 5 dummy products along with their prices in a JSON format:
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-3.5-turbo-1106",
response_format={ "type": "json_object" },
messages=[
{"role": "system", "content": "You are a helpful assistant designed to output JSON."},
{"role": "user", "content": "Generate a dataset containing 5 dummy product and a price"}
]
)
print(response.choices[0].message.content){
"products": [
{
"name": "Alpha Product",
"price": 20.99
},
{
"name": "Beta Product",
"price": 15.49
},
{
"name": "Gamma Product",
"price": 25.99
},
{
"name": "Delta Product",
"price": 18.99
},
{
"name": "Epsilon Product",
"price": 30.49
}
]
}This feature proves especially beneficial in application development scenarios where generating a specific format, like JSON, is essential.
Feature: Predictable Outputs and Log Probability Insights (Beta)
GPT-4 Turbo introduces a ‘seed’ parameter that ensures the model provides consistent completions most of the time, enabling reproducible outputs. This beta functionality is especially beneficial for replaying requests during debugging, crafting detailed unit tests, and gaining greater control over model behavior. OpenAI found this feature invaluable during unit testing and would be useful for ensuring reproducible outputs from the large language model. An example notebook can be found here.
The ‘seed’ parameter in GPT-4 Turbo is like a fixed recipe that ensures you get the same result every time you use it. Imagine if every time you baked a cake with the same recipe, you got a different tasting cake. That would be unpredictable and not very helpful if you wanted to recreate a specific flavor. The ‘seed’ parameter is like having a magic ingredient that guarantees your cake will taste the same every time you bake it using that recipe.
In more technical settings, like when developers are testing software or building applications, having this consistency is very important. It’s like making sure the cake turns out perfect every time because they can repeat their tests or processes and know they’ll get the same result. This makes it easier to check if everything is working correctly and to build more reliable and predictable systems.
Additionally, in the coming weeks, OpenAI plans to introduce a feature that reveals log probabilities for the most likely output tokens produced by both GPT-4 Turbo and GPT -3.5 Turbo. This will be instrumental in developing functionalities like autocomplete in search interfaces.
Other Updates from the OpenAI Devday
Other than the release of GPT-4 Turbo, GPT-3.5 has been improved as well as the introduction of the following features:
- DALL·E 3 Integration: This new feature for ChatGPT Plus and Enterprise users allows for advanced image generation through the Images API, offering diverse formats, quality options, and built-in moderation. Companies like Snap, Coca-Cola, and Shutterstock are already using it, with costs starting at $0.04 per image.
- Text-to-Speech (TTS) API: This API creates human-like speech from text, featuring six voices and two versions (tts-1 for real-time use and tts-1-hd for high quality) at a starting price of $0.015 per 1,000 characters.
- GPT-4 Fine Tuning Experimental Access: An experimental program that presents more complex challenges for significant improvements compared to GPT-3.5. Users of the GPT-3.5 fine-tuning will gain access to this advanced program.
- Custom Models Program: Aimed at organizations requiring deep customization, especially those with large proprietary datasets. This program involves a complete overhaul of the model training process and provides exclusive access to custom models without data sharing or reuse. This exclusive and costly program is open for applications.
In conclusion, OpenAI’s Devday showcased the impressive capabilities of GPT-4 Turbo, an advanced iteration of GPT-4, notable for its expanded knowledge base, a vast 128k token context window, and groundbreaking multimodal functions for image processing. This development, combined with more competitive pricing and enhanced features like improved instruction compliance and JSON format output, marks a significant stride in AI technology. Alongside GPT-4 Turbo, updates to GPT-3.5, the DALL·E 3 Integration, the Text-to-Speech API, and the Custom Models Program were also significant developments. Follow us at the Version 1 AI Labs to hear about the latest updates and developments in the field of AI and how you can use it in your organisation.
About the Author:
Rohit Vincent is a Data Scientist at the Version 1 AI Labs.
