GPT-5, the latest evolution of OpenAi artificial intelligence technology at the origin of the Chatgpt phenomenon in 2022, is about to be revealed. Users are already wondering about the extent of progress compared to GPT-4, and if this new leap will be up to the previous advances in the research laboratory.
Two first testers of the new model told Reuters that he had been impressed by his programming capacities and his resolution of scientific and mathematical problems. However, according to them, the jump from GPT-4 to GPT-5 would not be as spectacular as that observed between GPT-3 and GPT-4. These testers, subject to privacy agreements, wanted to keep anonymous.
Openai refused to comment on this information.
The improvement of GPT-4 was based on increased computing power and a larger amount of data. The company hoped that this “scaling” strategy would obtain ever more efficient AI models.
However, Openai, supported by Microsoft and currently valued at 300 billion dollars, has encountered difficulties in continuing this rise. One of the major obstacles was the “data barrier”, as the former scientific director Ilya Sutskever explained last year: if the computing power progresses, the amount of data stagnates.
It referred to the fact that large language models are trained on huge datasets from the internet set, and that AI laboratories have little options to access large corpus of texts generated by humans.
In addition to this lack of data, another problem lies in the “training cycles” of large models, which are more subject to hardware failures due to the complexity of systems. Researchers can often assess the final performance of the model only at the very end of training, a process that can last several months.
OPENAI did not communicate an official date for the release of GPT-5, but according to several media, the industry expects an imminent announcement. Boris Power, head of research applied to Openai, said on Monday on X: “Look forward to seeing how the public will welcome GPT-5. »»
“Openai has made a considerable leap between GPT-3 and GPT-4, which aroused enormous expectations around GPT-5”, analyzes Navin Chaddha, managing partner of the risk capital fund Mayfield, investor in AI but not shareholder of Openai. “Hope is that GPT-5 allows you to go beyond simple conversation to achieve the autonomous execution of complex tasks. »»
“Test-time compute”: a new step
Almost three years ago, Chatgpt revealed to the world the potential of the generative AI, impressive by its ability to write texts and poems of almost human quality, becoming one of the fastest growth applications in history.
In March 2023, Openai launched GPT-4, a language model having crossed a major course in terms of intelligence. While GPT-3.5, previous version, obtained a score among the lowest 10 % on the bar exam, GPT-4 succeeded in the best simulation.
GPT-4 has thus become the reference to beat, and the world has realized that the models of AI could surpass humans on many tasks.
Other companies quickly followed. The same year, Google (alphabet) and Anthropic-supported by Amazon and Google-launched competing models from GPT-4. In less than a year, open source models of an equivalent level, such as Meta Platforms Llama 3, have emerged.
In parallel with the training of large models, Openai has invested in a new path, the “test-time”, which consists in allocating more computing power to resolve complex tasks, such as mathematics or operations requiring reasoning and decision-making close to humans.
Openai CEO Sam Altman said earlier this year that GPT-5 would combine test-time and major models. He also admitted that the offer of models and products from Openai had become “complex”.