01.Ai challenges OpenAI with a low -cost high performance model | OpenAI ChatGPT | Chat GPT login free | OpenAI Login | Turtles AI

A Chinese company, 01.ai, challenged the Global Competition in the field of AI, managing to train a high-level model with a budget of only 3 million dollars, against 80-100 million spent by Openi for GPT -4. The key to the success of 01.ai was an optimized combination of engineering, innovations in processes and a highly efficient use of resources.

Key points:

• 01.ai ha addestrato un modello di AI avanzato con solo 2.000 GPU, contro le decine di migliaia usate da OpenAI.
• The training cost of the YI-Lightning model of 01.ai was only 3 million dollars, compared to the 80-100 million spent by OpenAI for GPT-4.
• The 01.AI strategy included optimizations in the inference processes and the use of a dedicated engine to improve efficiency and reduce operating costs.
• The cost of inference for 01.ai is just 10 cents per million token, significantly lower than competing models.

AI has become one of the most competitive and expensive sectors, but a new challenge was launched by an actor who could change the rules of the game. The Chinese company 01.ai recently revealed that it had trained an advanced model of AI, Yi-Lightning, with a budget of only 3 million dollars, a figure that appears considerably lower than the development costs of the best known models, such as GPT-4 by OpenAI, whose training is said to have cost between 80 and 100 million dollars. This result shows that, contrary to what has been suggested by customs in the sector, reaching high performance in AI does not necessarily require immense capital, but it can be the result of an engineering innovation and targeted optimizations. Kai-Fu Lee, founder and CEO of 01.Ai, explained that his company, while operating in a context of restrictions and limitations related to access to advanced technological resources such as the Nvidia GPUs, managed to obtain excellent results thanks to a highly efficient use of available resources.

The number of GPUs used by 01.ai to train Yi-Lightning was considerably lower than the huge hardware fleets used by companies like OpenAI. While the latter has relied on tens of thousands of advanced GPUs such as the Nvidia A100 for its previous models and many others for GPT-4, 01.Ai was able to count on "only" 2,000 units, which leads to questioning about differences in approach to the design and optimization of training processes. Despite the reduced number of GPU, 01.ai has managed to position itself among the leaders of the sector, with its Yi-Lightning model that occupies the sixth position in the performance of the model, as measured by LMSIS at the University of Berkeley.

Kai-Fu Lee stressed that one of the keys to this success lies in the way his company faced the bottlenecks during the inference process, that is, the use of the resources necessary to ensure that the model could respond to requests in real time. The company has implemented advanced engineering solutions to reduce these bottlenecks, transforming some of the computational operations into less expensive activities in terms of memory. One of the most innovative approaches was the development of a multilayer caching system, which allows you to enormously speed up access to the data requested during the inference. In addition, a dedicated inference engine was designed, capable of optimizing resources management specifically for the needs of the Yi-Lightning model. This approach has allowed 01.Ai to reduce operating costs: inference, which is one of the most expensive aspects of the AI ​​models, costs only 10 cents per million token, a figure that represents about 1/30 of the typical cost requested by competing models.

This efficiency, which is the result of targeted technical choices, fits into the context of a corporate reality that has had to face difficulties related to access to latest generation hardware resources. Chinese companies such as 01.Ai, in fact, must deal with the restrictions imposed by US regulations, which limit access to advanced GPUs such as those produced by Nvidia, for example, one of the main companies in the AI ​​sector. Despite these limitations, 01.ai has managed to plan a path of growth and development that has allowed her to remain competitive globally, focusing on detailed resources planning and on innovative solutions in the software field.

If on the one hand the numbers relating to the cost of the GPUs do not seem to return (for example, if GPU Nvidia H100 are assumed, a cost of about $ 30,000 each, 2,000 GPUs would cost around 6 million dollars, well above the 3 million declared ), on the other the story of 01.AI offers an interesting starting point on the potential of a more agile and optimized approach in training and in the use of AI models. The choice to focus on the efficiency of the processes, in particular in the inference phase, could be a way to other companies in the sector, which have to deal with high production costs and limited resources.

The incredible result of 01.ai shows that, even in a highly competitive context such as that of AI, innovation does not depend exclusively on the amount of resources invested, but on the ability to exploit them at best.