The development of artificial intelligence has long been not only something that concerns and interests only corporations and the typically commercial sector. The governments of great powers such as the USA or even China also have their ambitions. As always in such cases, such ambitions lead to friction and disagreements between the biggest players.
The greatest competition and rivalry is, of course, visible between the powers mentioned above, which see enormous potential in AI. This is also visible in attempts to sanction one side and effectively bypass prohibitions on the other. All this so as not to be pushed into the background.
While most eyes are turned towards the United States, Chinese models are “silently” achieving further impressive results. The latest tests show that one of the models, created in the “Middle Kingdom”, is hot on the heels of American giants. Is the competition heating up?
Cheap but still good
The founder and head of the Chinese company 01.ai boasted last week that his company had managed to accomplish an impressive but breakneck challenge. The company managed to train one of its most advanced AI models called Yi-Lightning on 2,000 systems. The entire project was also extremely economical, because according to the press release, it cost about $3 million.
We are talking here about a very significant difference in costs between what the AI industry leader, OpenAI, spent on training ChatGPT-4. In this case, the amount is approximately USD 80-100 million. Even more interesting are the reports about ChatGPT-5, the cost of which is estimated at up to $1 billion.
The differences are therefore gigantic and show that, if necessary, it is possible to optimize activities related to model training to the most acceptable amounts. However, there is a question about the effects, and they turn out to be… surprising.
Yi-Lightning ranks 6th in the performance ranking measured by LMSIS. This is an exceptionally good result, and although it falls short of ChatGPT-4 in September 2024, it is better than the version that appeared in May.
Some may wonder how the Chinese company managed to achieve such results. In short, the answer lies in the word “optimization”. It turns out that engineers were able to find ways for better buffering, better engines and allocation of resources, of which there were not very many.
Optimization works wonders
So it turns out that if you want and have the need, you can achieve real miracles. After all, Chinese companies do not have access to the latest technologies from NVIDIA, or at least it is officially so. Scale and the possibility of operating in the so-called the gray zone cannot be thoroughly examined.
It is true that Chinese equipment manufacturers have the ambition to create their own chips, and Huawei is an example here. Recently, however, it came to light that this giant used the services of the Taiwanese TSMC and even if it was indirect cooperation, it is not possible at the moment.
It will be interesting to see what the capabilities of Chinese AI models will be like in the long term. Will it be a real alternative to what the USA will offer? How will companies cope with sanctions? We will only know the answers to these questions for quite some time.
