What to expect from the future incarnation of the OpenAI generative model for artificial intelligence. What changes compared to the current versions.
Sam Altman, the number one of OpenAI, recently declared that the era of large-scale generative models is already over, suggesting that there would not be a GPT-5 for quite some time, at least not in 2023.
In reality, several are boiling in the pot novelty that seem to look to a new generative model even more powerful than the GPT-4 also used in the most recent incarnation of the ChatGPT chatbot. According to rumors coming from sources inside OpenAI itself, around the month of October 2023, it should first of all debut GPT-4.5 that is, an intermediate version of the generative model that will adopt the approach multimodal.
We saw this recently with Meta ImageBind: a generative model for AI with multimodal capabilities is able to generate and understand more information mode. In general, modalities refer to different sensory channels or types of data, such as text, images, audio or even video. A multimodal generative model can work with more than one of these modalities simultaneously, integrating and combining information from different sources to generate output consistent and complete.
For example, a multimodal generative model might be able to generate a text description detail of an image or to generate an image starting from a textual description. This requires the model to understand both the image content and the meaning of the text to produce a consistent and meaningful output.
Multimodal capabilities can also be extended to other modalities, such as multimodal translation, where the model can translate text from one language to another by incorporating contextual audio or visual information. Similarly, a multimodal generative model can be used for the speech synthesis of text or for the generation of dialogues that include text, images and audio.
Second Metamultimodal generative models will be able to create immersive content for the virtual reality: the main player of the sector are in turmoil so it will be interesting to understand what OpenAI will do in this segment and if we will see AI applications with a strong innovative load in many sectors, hitherto unexplored, precisely with the release of GPT-5.
The next step will be to develop one AGI (Artificial General Intelligence) that is, a type of artificial intelligence that is able to “understand”, “learn” and perform a wide range of intellectual activities similar to those performed by humans. Unlike specialized AI systems, which are designed to perform specific tasks or deal with specific problems, AGI aspires to a form of general artificial intelligence that can be flexibly applied to a wide range of activities.
An AGI should be able to face new situations or tasks he has never encountered before, using his general intelligence to draw conclusions and make decisions on his own.
The intent of OpenAI would be to shape this superintelligence precisely with the arrival of GPT-5, which according to some estimates would not be presented on the market before 2024. Also because GPT-4 itself is an “open construction site”: the time to inference of GPT-4 is very high and is quite expensive to run. The same access to Plugins in ChatGPT and Internet browsing functionality are still in beta.
The future GPT-5 model should however reduce the inference time, improve the efficiency, counter the problem of hallucinations that is, the tendency of models to generate output that do not correspond to reality or that contain imaginary or invented information.
According to OpenAI, GPT-4 is 82% less likely to respond to prompt of users with inaccurate content or out of place; it also has an accuracy close to 80% in all test categories, highlighting a clear progress in the fight against hallucinations. With GPT-5, the company is expected to be led by Altman it will reduce the hallucinations at less than 10%: a real “milestone” for LLM models.
A huge chunk of OpenAI’s revenue comes from enterprises and enterprises so GPT-5 doesn’t have to be alone anymore economic but also faster to return the output. The developers are already scolding the fact that calls API GPT-4 they often stop responding and are forced to use the GPT-3.5 model in production: with future evolutions of the generative model, this should change.
Finally, another novelty: registering the enormous success of Meta’s LLaMA open source generative model and other solutions, including derivatives, open source, OpenAI would have changed its position. Do you also remember what a Google employee said that open source models for artificial intelligence would surpass those of OpenAI and Google itself?
Well, always according to sources close to OpenAI, the company would be working on a new open source artificial intelligence model that will soon be released to the public. There’s no information on its capabilities and how competitive it will be against GPT-3.5 or GPT-4, but it’s a welcome change indeed.