Keeping up with the rapid development of artificial intelligence and technology can seem very difficult. New stuff drops every week or month, and now you're here to learn something new again!
This time it is GPT-5.
GPT-4 was released in May 2023, and since then everyone has been waiting for the release of GPT-5. Chen Siqi tweeted on March 27 that "gpt5 is scheduled to finish training in December this year." However, OpenAI CEO Sam Altman clarified this statement at an MIT event in April, when asked about GPT-5 by saying "we won't and won't be for a while."
So this clarifies that. However, some experts suggest that OpenAI release GPT-4.5 in Q3/Q4 of 4.5, which is an intermediate version between GPT-4 and GPT-5. Improvements have been made to the current model, which could be a potential release of GPT-4.5. Many have said that GPT-4.5 has the potential for multi-modal capabilities, which was already demonstrated during the GPT-4 developer livestream in April 2023.
Despite the high hopes for GPT-5, GPT-4 still needs to iron out some creases. For example, GPT-4 has very high inference times and is computationally expensive to run. There are other challenges, such as accessing the GPT-4 API.
While there is still work to be done, what we can say is that each version of GPT pushes the boundaries of AI technology and its capabilities. AI enthusiasts are excited to explore the breakthrough capabilities of GPT-5.
So what features can we expect from GPT-5? Let's find out.
reduce hallucinations
It's all about trust, and it's the main reason why most users don't trust AI models. For example, GPT-4 scores 40% higher than GPT-3.5 on the internal truth evaluation across all nine categories, as shown in the figure below. This means that GPT-4 is less likely to respond to disallowed content and 40% more likely to generate factual responses than GPT-4.
As new versions will continue to improve on the current challenge, GPT-5 is said to reduce hallucinations below 10%, making LLM more trustworthy.
Image from OpenAI
Computational efficiency
As mentioned earlier, GPT-4 is computationally expensive at $0.03 per token. This compares to GPT-3.5's cost of $0.0002. This is a big difference. GPT-4 trained on a trillion-parameter dataset and infrastructure reflects cost.
While Google's PaLM 2 model only trains 340 billion parameters and has efficient performance. If OpenAI plans to compete with Google's PaLM 2, they will need to investigate ways to reduce the cost and parameter size of GPT-4 - while being able to maintain performance.
Another aspect that needs to be investigated is better inference time, or the time it takes for deep learning models to make predictions on new data. The more features and plugins in GPT-4, the more computationally efficient it is. Developers are already complaining to OpenAI that the GPT-4 API frequently stops responding, forcing them to use GPT-3.5.
Taking all of these factors into account, we can expect OpenAI to overcome these challenges with a smaller, cheaper, and more efficient version of GPT-5.
multisensory
Before GPT-4 was released, many people were crazy about its multi-modal capabilities. While it hasn't been added to GPT-4 yet, this is where GPT-5 could come in and be the star of the show and really make it multimodal.
Not only can we expect it to handle images and text, but also audio, video, temperature, and more. Sam Altman said in an interview: "I'm very excited to see what happens when we can do video. There is a lot of video content in the world. There are many things that are easier to learn through video than text.
Adds data types that can be used to make conversations more dynamic and interactive. Multimodal capabilities will be the fastest link to artificial general intelligence (AGI).
long term memory
GPT-4 has an impressive maximum token length of 32,000 tokens. But as the world releases model after model, we have models like Story Writer that output 65,000 tokens.
To keep up with the current competition, we can expect GPT-5 to introduce longer context lengths, allowing users to have AI friends who can remember their characters and history over the years.
Improve contextual understanding
As a large language model (LLM), the first thing we can expect is improved and enhanced ability to understand context. If we merge this with the above point about long-term memory, GPT-5 has the potential to maintain context over long conversations. As a user, you will get more catered and meaningful responses tailored to your requirements.
With that comes a higher-level understanding of language, and a major component of natural language is emotion. The potential capabilities of contextual understanding in GPT-5 can make it more empathetic and generate appropriate responses to continue engaging in a conversation.
pack up
There is much more to learn about the potential capabilities of GPT-5, and we won't be able to find out much until closer to launch. This paper builds on the current challenges facing GPT-4 and GPT-3.5, and how OpenAI leverages these obstacles to overcome and produce a high-performance version of GPT-5.