In March 2023, almost a year has passed since the launch of GPT-4. OpenAI announced a faster and improved GPT-4 Turbo model in November 2023, but there is no sign of GPT-5 in sight. We have seen how GPT-4 performs in various tasks, such as image analysis, code interpretation, and more. Now AI aficionados are looking forward to learning about the upcoming OpenAI model, GPT-5, the potential of AGI, and more. So follow our explanation below to find more in-depth information about GPT-5 release date and expected features.
GPT-5 Release Date Speculation
While GPT-4 will be released in March 2023, OpenAI is expected to launch its next generation model in December 2023. However, we are already in 2024, but there is no official announcement about the launch of GPT-5 by OpenAI. The company filed for the GPT-5 trademark with the USPTO on July 18, 2023, but there has been no word since then.
Speaking recently at the World Economic Forum on January 17, 2024, Sam Altman told Ina Fried of Axios that his main priority now is to launch a new model that shows GPT-5. He also added that the next big model “will have the capability to do much, much more.“
Additionally, Altman confirmed that the GPT-5 will have “generalized intelligence” and that the new model will be smarter than its predecessor. In a recent podcast with Bill Gates, Sam Altman spilled the beans on new capabilities coming to GPT-5. Advanced imaging, better image analysis and video capabilities come to GPT-5.
In addition, OpenAI is working to make the next generation of ChatGPT more customizable and personal. He added,
“Customizability and personalization will also be very important. People want very different things out of GPT-4: different styles, different sets of assumptions. We’ll make all that possible, and then also the ability to have it use your own data. The ability to know about you, your email, your calendar, how you like appointments booked, connected to other outside data sources, all of that.”
So, we don’t know the exact release date of GPT-5, OpenAI will try to release GPT-5 in 2024, maybe by the end of the year. With the release of the Gemini and Gemini 1.5 Pro models, Google has put pressure on OpenAI. Additionally, Antropic recently launched the Claude 3 model, and the Opus model has shown great promise.
In particular, Mark Zuckerberg said that Llama 3 is currently being trained and will become one of the leading AI models in the industry. Yes, 2024 would be a safe bet for the release of GPT-5, as this would allow OpenAI to continue its lead in the AI race.
That said, recently leaked information suggests that OpenAI may release an interim GPT-4.5 Turbo model first, followed by a GPT-5 model by the end of the year. In addition, OpenAI introduces Sora, an excellent text-to-video model. The company is currently working on a red team with experts to assess the damage and model the risk. Again, OpenAI has a lot on its plate right now, and by 2024 we’ll probably see a lot of big sales of AI models.
Expected Features and Capabilities of GPT-5
Reduced Hallucinations
The most interesting conversation in the industry will reach GPT-5 AGI (Artificial General Intelligence), but we will go into more detail about it. Additionally, GPT-5 is supposed to reduce research time, increase efficiency, reduce hallucinations, and more. Let’s start with hallucinations, one of the main reasons why most users don’t easily trust AI models.
According to OpenAI, GPT-4 outperformed GPT-3.5 by 40% in real-time benchmarks generated by internal adversarial across nine categories. Now, GPT-4 is 82% less likely to respond to illegal and banned content. It came very close to achieving the 80% mark in the category-by-category accuracy test. This is a giant leap forward in the fight against hallucinations.
It is anticipated that OpenAI will reduce hallucinations to less than 10% in GPT-5., which will be huge to make the LLM model reliable. In an interview with Bill Gates, Sam Altman said that the company is working to improve the accuracy of ChatGPT to make it a reliable AI chatbot. In my personal experience, the GPT-4 model provides mostly factual answers. Therefore, it is possible that GPT-5 is less hallucinating than GPT-4.
A computational model
Furthermore, we know that GPT-4 is expensive to run ($0.03 per 1K digits) and the search time is higher. Meanwhile, the old GPT-3.5 turbo model is 30x cheaper than GPT-4 ($0.0010 per 1K numbers). OpenAI has managed to improve performance and reduce costs with the latest model GPT-4 Turbo ($0.01 per 1K numbers), but it is still not available in ChatGPT due to the lack of computing resources.
According to the latest report by SemiAnalysis, GPT-4 is not a compact model, but based on the “Mixture of Experts” architecture. That means GPT-4 uses 16 different models for different problems and has 1.8 trillion parameters.
With such a large infrastructure, the GPT-4 model is very expensive to operate and maintain. In recent comments about Google’s PaLM 2 model, we saw that PaLM 2 is very small in size, which makes it quick to execute. Similarly, Google’s latest Gemini model is taught to be a multimodal model of text, image, audio, video, and code designed together to build a strong text system from scratch.
A recent report by CNBC confirmed that PaLM 2 was designed with 340 billion parameters, which is less than the large parameters of GPT-4. Google says that bigger is not always better, and that research creativity is the key to creating a great model.
So, if OpenAI wants to computationally optimize its future models, it needs to find creative new ways to reduce model size while maintaining output quality.
A significant part of OpenAI’s revenue comes from businesses and companies, so GPT-5 is not only cheaper, but also faster to return products. Developers claim that GPT-4 API calls often stop responding and are forced to use GPT-3.5 in production. Therefore, it should be on OpenAI’s wish list to improve performance in the upcoming GPT-5 model.
Multi-sensor AI model
Of course, GPT-4 is a multimodal AI model, but it currently only deals with two types of data, namely images and text. However, with GPT-5, OpenAI can take a big leap and expand its multimodal capabilities even further. It can also handle text, audio, photos, videos, depth information, and temperature.
It can combine data streams from different methods to create an internal space. In a podcast with Bill Gates, Sam Altman said that the OpenAI team is working to add video capabilities to ChatGPT, which is why it exists.
Meta recently released ImageBind, an open-source AI model for research purposes that integrates data from six different methods. In this space, OpenAI is not yet fully open, but the company has a strong foundation for vision analysis and image generation.
OpenAI also developed CLIP (Contrast Language – Image Preparation) for image analysis and DALL-E 3, a popular Midjourney alternative that can generate images from text descriptions.
This is an area of ongoing research and its applications are not yet clear. According to Meta, it can be utilized for designing and crafting immersive content for virtual reality. We will have to wait and see what OpenAI does in this space, and with the release of GPT-5 we will see more AI applications in various multimodalities.
Long term memory
With the release of GPT-4, OpenAI first introduced a maximum context length of 32K characters, at a cost of $0.06 per 1K character. Looking at the competition, Anthropic increased the context window from 9K to 100K characters in the May 2023 Claude AI chat. Now, OpenAI has increased the context length to 128K by introducing GPT-4 Turbo. Without suffering anthropic defeat, GPT-4 Turbo introduces a 200K context window at launch.
Now, OpenAI is expected to rise again in the competition and create a bigger context with the release of GPT-5. Leaked GPT-4.5 Turbo has indicated that it will support a 256K context window.
It can help you create AI characters and friends that remind you of your personality and memories that can last for years. You can also load a library of books and text documents into one context window. A variety of new AI applications are possible thanks to long-term memory support, and GPT-5 makes this possible.
Reasonable price of GPT-5
We already know that the GPT-4 is more expensive to use compared to the GPT-3.5-turbo model. Recently, OpenAI has reduced the price of both models and released GPT-4 with 32K context length and turbo GPT-3.5 with 16K context length. For GPT-4 with a context length of 32K, the sampled token cost is $0.12/1k. The GPT-4 Turbo model is cheaper. $0.03 per 1K characters for 128K context length.
By comparison, Anthropic AI’s recently released Clod 2.1 costs about $0.02 to generate 1000 words and supports contexts larger than 200K in length.
So if OpenAI wants developers to adopt GPT-5 in the future, the company needs to keep prices competitive and reasonable. As mentioned above, the computational cost of the GPT-4 mixed model is very high due to the large infrastructure. OpenAI needs to find a way to create a more capable and advanced density model than the current GPT-4 model.
GPT-5 Release: AGI Scared?
In February 2023, Sam Altman blogged about AGI and how it could benefit all of humanity. AGI (Artificial General Intelligence), as the name suggests, is the next generation of AI systems that are usually smarter than humans. OpenAI’s upcoming GPT-5 model will be the next step towards AGI, and there seems to be some truth to this.
We already have several autonomous AI agents, such as Auto-GPT and BabyAGI, which are based on GPT-4 and can make autonomous decisions and draw the right conclusions. It is entirely possible that some version of AGI is deployed with GPT-5.
In the blog post, Altman suggests “We believe we have to continuously learn and adapt by deploying less powerful versions of the technology in order to minimize ‘one shot to get it right’ scenarios”. He also acknowledges the significant risks involved in managing powerful systems like AGI. Prior to the recent Senate hearing, Sam Altman also called on US lawmakers to establish regulations around newer AI systems.
During the hearing, Altman expressed concern, stating, “I think if this technology goes wrong, it can go quite wrong. And we want to be vocal about that.” He also remarked, “We want to work with the government to prevent that from happening.” OpenAI has been actively advocating for regulations on advanced AI systems that possess high levels of intelligence and power.
Note that Altman is looking for security rules on incredibly powerful AI systems, not open source models or AI models developed by small startups.
It should be noted that Elon Musk and Steve Wozniak, Andrew Yang, and Noah Harari, among others, called for a pause in massive AI experiments in March 2023. Since then there has been a massive backlash. Fight AGI with a new AI system – more powerful than GPT-4.
Elon Musk recently filed a lawsuit against OpenAI for being a closed source AI company, claiming that the company is developing AGI to make money for Microsoft. Musk is seeking an injunction to prevent OpenAI and Microsoft from paying money for its AGI technology.
If OpenAI is indeed going to bring AGI capabilities to the GPI-5, expect more delays in its general release. Regulation will certainly start and work on security and alignments will be reviewed carefully. The good thing is that OpenAI already has a strong GPT-4 model, and it is constantly adding new features and capabilities.
OpenAI GPT-5: Future Position
After the release of GPT-4, OpenAI became more secretive about its operations. He no longer shares his research on learning frameworks, architectures, hardware, learning computing, and learning methodologies with the open source community. It is surprising for a company founded as a non-profit (now profitable) based on the principle of free cooperation.
Speaking to The Verge in March 2023, OpenAI chief scientist Ilya Sutskever said: “We were wrong. Flat out, we were wrong. If you believe, as we do, that at some point, AI — AGI — is going to be extremely, unbelievably potent, then it just does not make sense to open-source. It is a bad idea… I fully expect that in a few years, it’s going to be completely obvious to everyone that open-sourcing AI is just not wise.“
Now, it is clear that neither GPT-4 nor the upcoming GPT-5 will be open source to compete in the AI race. However, Meta takes a different approach to AI development. Meta released several AI models under the Llama 2 public license standard (with some restrictions for commercial use) and gaining traction among the open source community. Zuckerberg wants to open source Meta AGI in the future.
All in all, GPT-5 will be a frontier model that will truly push the boundaries of what is possible with AI. With the release of GPT-5, we might see a spark of AGI. If so, OpenAI should be prepared for stricter regulation (and possible bans) around the world. As for the estimated release date of GPT-5, the safe bet is 2024.
0 Comments