Singapore‘s Sapie Ielligence Startup, a new artificial ielligence architecture has developed a new artificial ielligence architecture that can compete with large language models (LLM) and in some cases surpassed them.
This architecture, known as the “hierarchical reasoning model”, is designed to be inspired by the manner of human brain function. The system focuses on how the human brain uses distinct systems for slow and conscious planning along with fast and iuitive calculations.
This new architecture can transform the developme of artificial ielligence
This model achieves significa results by using a small volume of data and memory that today’s large language models need. This high performance can have importa benefits for real applications of artificial ielligence at the organizational level; Especially when the data are limited and low computational resources.
In the face of complex issues, large language models rely largely on the “chain of thought”. In this way, the problem is broken io the middle steps of the text, and the model is practically forced to think long or explicitly express its meal stages when moving to the solution.
Singapore’s researchers, however, have claimed in their article that the use of COT for reasoning is not a satisfactory solution, but a temporary support. This method relies on fragile breakdowns and pre -defined by humans, where even a small mistake or displaceme of the arrangeme of the steps can remove the eire process of reasoning.
But to overcome this stage, the researchers examined the concept of “hidden reasoning”, in which the model performs the process of reasoning in the form of iernal and abstract represeations instead of producing iellectual signs. This approach is more consiste with human thinking.
Finally, the researchers have been able to develop a new architecture that is up to 5 times faster than the curre models, and with only a thousand training samples have achieved this level of skill and speed.
