This week, Greg Brockman, President and Co-Founder of OpenAI, took to the public to announce the latest advancements in ChatGPT, and I must say, GPT-4 is mind-blowing!
With a concise presentation devoid of unnecessary theatrics, the President and Co-Founder of OpenAI showcased the latest developments of ChatGPT. He utilized the tool to generate speeches, songs, jokes, and even code in front of millions of viewers, demonstrating impressive capabilities and potential applications of GPT-4 across various domains and tasks.
We had already shown our awe at ChatGPT‘s capabilities before, which was in its 3.5 version, but this week those capabilities reached another level! I mean, if ChatGPT was already revolutionizing everything, GPT-4 is the revolution of the revolution!
How does GPT-4 work?
GPT-4 is based on deep learning, a branch of machine learning that uses neural networks to learn from data. It was trained on a vast amount of texts and images from the Internet, covering various domains and languages, and is capable of predicting the next word or image based on input data.
Using a technique called self-attention, the model can calculate the relevance or similarity between each pair of words or tokens in the input data. For example, for an input sentence “The cat drank milk,” GPT-4 will calculate how much each word relates to every other word in the sentence, such as “cat” with “drink” or “milk.” Based on these calculations, the model will assign different weights or scores to each pair of words, determining how much attention they will receive. By doing this, it can learn which input information is most important or relevant for generating the output. For example, if the task is to complete the sentence, GPT-4 will pay more attention to the word “cat” than to “the” because it is more informative for predicting the next word.
GPT-3.5 vs GPT-4
GPT-4 is not only more capable than its predecessor GPT-3.5, but also more reliable, creative, and able to handle much more complex and nuanced instructions. To demonstrate its abilities, OpenAI tested GPT-4 on a variety of benchmark standards, including simulating exams originally designed for humans. The model was able to pass an exam with a score around 10% higher; in contrast, GPT-3.5’s score was around 10% lower. It also outperformed the previous version in its advanced reasoning capabilities.
OpenAI spent 6 months making GPT-4 safer and more aligned with human values and expectations. It incorporated more human feedback into its training process to improve its behavior and reduce harmful or biased outcomes. It also worked with over 50 experts for early feedback in domains such as security and AI safety.
A major difference is that GPT-4 can accept inputs of both images and text and output text, and has a 10 times greater capacity than its predecessor. To give you an idea, in the presentation, Greg Brockman drew the structure of a web page on paper, and GPT-4 identified what it was and returned the HTML code ready for page programming. Additionally, the tool can also describe images in detail.
The architecture used is also much more evolved, consisting of a deep neural network (called Transformer) that allows multiple layers of attention mechanisms. This enables the intelligence to capture contextual information by drawing parallels and handling large amounts of data.
When I mention the ability to handle large volumes of data, I would like to comment on a part of the demonstration where Greg Brockman copied the entire regulation and instructions for tax filing (over 100 pages) and GPT-4 was able, in seconds, to assist in filing the income tax return! Just imagine how easy our life will be to add entire manuals there and ask if what you want to do is possible! In a matter of seconds, the technology “reads” the manual and answers you, faster than you would take to locate the item in the index!
How to use GPT-4?
There are two main ways to access GPT-4: through ChatGPT Plus or through the API.
ChatGPT Plus is a chatbot service that allows users to interact with GPT-4 and other models through a web interface. Users can ask questions, request tasks, or have casual conversations with GPT-4. To use ChatGPT Plus, users need to subscribe to a plan that costs $20 per month. Benefits include faster response times, priority access to new features, and unlimited usage.
The API is an application programming interface that allows developers to integrate GPT-4 into their own applications and services. Developers can use the API to send requests and receive programmatic responses from GPT-4. To use the API, developers need to join a waiting list and wait for an invitation from OpenAI, as it is currently in beta and has limited availability.
To sign up for ChatGPT Plus or the API, users need to visit the OpenAI website and follow the instructions there. Users can also try out GPT-4 without a subscription in the Microsoft Bing browser, but with limited functionality and availability.
The complete video
Interested in seeing the full demo? Here’s the video for you.
GPT-4 is an impressive breakthrough that showcases the potential of artificial intelligence for various applications and domains. However, it is not without limitations or challenges. As OpenAI acknowledges, GPT-4 is still less capable than humans in many real-world scenarios, and exhibits human-level performance only in certain professional and academic benchmarks. It also requires careful oversight and governance to ensure its ethical and responsible use. As such, OpenAI invites researchers, developers, users, policymakers, and society at large to engage with them in exploring the opportunities and risks of this technology.