Artificial intelligence companies like OpenAI have begun to innovate training techniques for advanced language models, seeking a more “human” way to make algorithms think. In a context in which the expansion of massive models faces unexpected challenges, the development of techniques that optimize the “reasoning” of AIs during the use phase is emerging as a promising alternative.
A shift in focus: Beyond “bigger is better”
Although advances in AI over the past decade have been driven by the philosophy of “scaling up” in terms of data and computational power, some of the most influential AI scientists are pointing out the limitations of this mentality.
Ilya Sutskever, co-founder of OpenAI, recently acknowledged that the results of increasing the pre-training phase have reached a saturation point, suggesting the need for a renewed approach. «The 2010s were the era of scale; “Now we are back in the age of discovery,” Sutskever said.
“Test-time computing” techniques: The future of reasoning in AI
To improve existing models, OpenAI has incorporated a technique known as “trial-time computing,” allowing models to process multiple responses before making a final decision.
In the new “o1” model, recently launched, this approach allows for step-by-step, human-like reasoning, especially being applied in complex areas such as mathematics or programming. According to Noam Brown, researcher at OpenAI, this method achieves significant improvements without the need to massively expand the models.
Microsoft and META invest heavily in the AI sector.
Global competition and the impact on AI infrastructure
As OpenAI and other leading companies such as Anthropic, Google DeepMind and xAI explore this approach, the race to optimize the AI market intensifies.
Kevin Weil, product director at OpenAI, assured that the company is focused on remaining “three steps ahead” of its competitors. This shift toward inference techniques and reduced reliance on pre-training clusters could also impact demand for AI chips, a sector dominated by Nvidia.
Nvidia and the inference market: a new challenge?
Until now, Nvidia has led the market for AI model training chips, but with this shift toward inference, other competitors could gain ground.
“We are facing a new law of scalability, one that applies to inference time,” commented Jensen Huang, CEO of Nvidia. This evolution could shift demand for massive pre-training clusters toward distributed inference clouds, something venture capital investors are watching closely.
Implications for AI development and enterprise use
With these new techniques, AI labs could reduce the time and resources required for training models, driving the creation of more complex applications in a more accessible framework. The transition to artificial intelligence that “thinks” in real time represents a crucial change that, in the long term, could transform the way AIs interact and solve problems in various sectors.
Do you want to always be updated in the world of cryptocurrencies? Subscribe now to the CriptoTendencia WhatsApp channel! Here you will instantly receive the most relevant information about Bitcoin, Altcoins, DeFi, NFTs, Blockchain and the Metaverse.
Related
Crypto Keynote USA
For the Latest Crypto News, Follow ©KeynoteUSA on Twitter Or Google News.
