Artificial intelligence (AI) enters a new phase of development thanks to innovative training techniques. These approaches, led by OpenAI and other leading companies, aim to create models that are smarter, more efficient, and capable of reasoning similar to human ones.
Let’s see all the details in this article.
More efficient models and human reasoning: how is the training of AI progressing
Artificial intelligence (AI) is facing a crucial turning point, thanks to the introduction of innovative training techniques that promise to revolutionize the sector.
Companies like OpenAI are indeed working to overcome the limitations of current methods, addressing issues of scalability, costs, and energy consumption.
The o1 model of OpenAI, one of the main innovations, represents a concrete example of how AI can evolve towards a more human and sustainable approach.
In recent years, the expansion of large language models (LLM) has reached a critical point. Despite the significant progress of the 2010s, researchers have encountered increasing difficulties.
Ilya Sutskever, co-founder of OpenAI and Safe Superintelligence, emphasized that the focus now shifts to quality rather than quantity.
“Scaling in the right direction is what matters most,” he stated, indicating that simply expanding the models is no longer sufficient to achieve significant improvements.
In this sense, the o1 model by OpenAI stands out for its unique approach. Instead of relying solely on an increase in computational resources, it uses techniques that mimic human reasoning.
By dividing tasks into phases and receiving feedback from experts, o1 manages to process complex data in a more accurate and strategic way.
Furthermore, the adoption of a method called “test time calculation” allows for the allocation of computational resources in a more targeted manner, improving performance without an exponential increase in costs.
A concrete example of this innovation was presented by Noam Brown, researcher at OpenAI, during the TED AI conference.
It has indeed demonstrated that a bot, reasoning for only 20 seconds in a hand of poker, achieved results equivalent to a model trained for 100,000 times longer.
This result highlights the potential of new techniques to make AI more powerful and efficient.
The challenges of energy and data
In addition to the high costs, training large AI models also involves a significant energy consumption. The training runs require enormous amounts of computational power, with tangible consequences on electrical grids and the environment.
Another crucial problem is represented by the scarcity of data: language models have now used up most of the information accessible online, creating an unprecedented challenge for future development.
To address these issues, researchers are exploring more sustainable methods. The o1 model, for example, uses specialized data and optimizes processing only for tasks that require complex reasoning, reducing the overall consumption of resources.
In other words, the new techniques not only redefine the way models are trained, but they could also transform the bull market of IA hardware.
Companies like Nvidia, a leader in the production of AI chips, might have to adapt their products to meet new demands.
Nvidia, which in October became the most valuable company in the world thanks to the demand for AI chips, could face increasing competition from new players offering alternative and more efficient solutions.
Competition and innovation
Other laboratories, including Google DeepMind, Anthropic, and xAI, are developing their own versions of the techniques adopted by OpenAI. This competition is set to stimulate further innovations, paving the way for increasingly advanced and diverse AI models.
The growing competition could also reduce the costs associated with AI, making these technologies more accessible for a larger number of companies and sectors.