what should we expect from the next OpenAI model?

what should we expect from the next OpenAI model?
what should we expect from the next OpenAI model?

OpenAI’s future artificial intelligence model crystallizes the expectations of the generative AI community.

Will come out, will not come out? Since the beginning of April, rumors about the potential announcement of GPT-5 have proliferated on social networks. However, a live event is being held this Monday, May 13 at 7 p.m. to present new capabilities for ChatGPT and potentially an update to GPT-4. But OpenAI tells us that no announcement around GPT-5 will take place. Before this arrives, the JDN decided to take stock of the capacities expected by users and more generally by the market for the next version of the flagship model of the San Francisco start-up.

What we certainly know

Where the smartphone world can pride itself on obtaining quality leaks, generative AI has few credible sources. It is therefore impossible to reliably anticipate the characteristics of GPT-5. However, a few clues outline the broad outlines of OpenAI’s next flagship model. Starting with the words of the company’s CEO. Interviewed by Lex Fridman in March, Sam Altman explained that OpenAI is working to overcome hardware challenges. The technique would also be drastically worked on. We can anticipate a model with an even more optimized architecture and tokenizer and much faster inference.

More recently, the new Valley star gave concrete details during a conference in front of Stanford students. “GPT-5 will be a gigantic step in terms of capabilities compared to GPT-4. GPT-4 is the stupidest model anyone will ever have to use again,” teases Sam Altman. The already explosive calculation costs of GPT-4 training ($400 million) would likely be exceeded in the long term, he confided. OpenAI CEO says GPT-5 will be smarter, with “better reasoning, better understanding and advanced support for complex tasks”.

To avoid “surprising” the world by unveiling GPT-5 too quickly, OpenAI plans more iterative and progressive product releases to give society and institutions time to adapt (i.e. avoid direct lightning from regulators). “AI and surprise do not go well together,” justifies the manager.

The most credible hypotheses

Quite obviously, we can already anticipate a significant increase in the number of parameters of GPT-5. When GPT-4 had between 175 and several thousand billion parameters depending on the sources, GPT-5 could see this number increase significantly (up to a trillion?). As a result, we would obtain an improvement in our understanding of language, complex tasks and more generally the world around us. Claude 3 and Gemini 1.5 having started the ball rolling, it is a safe bet that GPT-5 hasa context size equal to or greater than 1 million tokens. Such a window size would make it possible to process new modalities currently not supported by GPT-4. We think of video and audio, two new modalities which should be supported by the model.

GPT-5 could also act with a certain degree of autonomy. According to credible rumors regularly relayed in recent weeks, GPT-5 would have the capacity tosummon developed agents by OpenAI to accomplish tasks automatically. Advanced search functionalities could also be implemented. According to several technical indices (domain name registration, code implementation), OpenAI would work on an advanced web search system driven by AI. A brick that would find its place within the next OpenAI model.

Hypotheses (a little) less credible

Other probable but less credible hypotheses assert that GPT-5 could constitute an AGI model. As a reminder, artificial general intelligence refers to an AI system capable of matching or surpassing humans in most, if not all, cognitive tasks. Unlike current AI, AGI would be able to learn, reason, and generally adapt, just like humans do. With the Q* project, OpenAI would have succeeded in developing several months ago an AI model with logical and mathematical reasoning, which could approach artificial general intelligence. Some experts therefore expect GPT-5 to demonstrate AGI or at least pre-AGI capabilities.

These new reasoning capabilities for global understanding of the world could lead GPT-5 to reach the singularity in several key areas. AI would then be more competent than humans in a large number of areas. It could evolve in partial or complete autonomy. In theory, AI would also have the ability to self-improve at an exponential rate. For the moment, the arrival of technological singularity with GPT-5 clearly does not seem to be on the agenda. A majority of researchers believe that we would have to wait several more years. The very principle of technological singularity is also regularly subject to controversy, as is its timetable.

The fact remains that OpenAI is expected by the market. Meta has struck a major blow with Llama 3 and is already planning to deploy, soon, a version with 400 billion parameters. Microsoft is also preparing with an LLM called MAI-1 with 500 billion parameters. A competition which makes the GPT-5 forecast schedule even more complex.

-

-

PREV Death of the French Ambassador to Sri Lanka and the Maldives
NEXT Belgian Championship: Amoura finishes 3rd top scorer of the season