Kai-Fu Lee: Compared to the giants in Silicon Valley, China's advantage in AI large models lies in faster and lower-cost commercial implementation

Zero One AI recently launched a new flagship pre-training model, Yi-Lightning, which surpassed OpenAI GPT-4o-2024-05-13 and Anthropic Claude 3.5 Sonnet on the international authoritative blind test list LMSYS, ranking sixth in the world and first in China.

Li Kaifu, the founder of Zero One AI and former president of Google China, told Titanium Media App, "It is the first Chinese large model to achieve a very high ranking in an international authoritative list, surpassing most American large models, and becoming the first Chinese large model to surpass the global leading OpenAI GPT-4o (May version). Yi-Lightning's lightning model not only has world-class model performance and very fast reasoning, but also has a very low price, making it very suitable for both App calls and enterprise application scenarios."

Li Kaifu candidly admitted that China lags behind the United States in AI, but some say it is behind by ten or twenty years. Based on the GPT4o model, calculating how far China is from surpassing the United States, Zero One AI is only 5 months away from OpenAI's model.

Li Kaifu later stated in an article on the front page of the Financial Times in the UK that China's advantage in AI lies in creating truly affordable reasoning engines, which is the most important thing for the vigorous development of AI applications. At the same time, China has a large number of technically skilled and hardworking engineering talents, giving it an advantage over the United States in this regard.

However, Li Kaifu also emphasized, "China's advantage may not necessarily make groundbreaking research without a capped budget, but it can definitely achieve landing better, faster, more reliably, and at a lower cost."

It is understood that Zero One AI (01.AI) was established on May 16, 2023, dedicated to building a new AI 2.0 platform and a global company of AI-first productivity applications. It was founded by Li Kaifu, Chairman and CEO of Innovation Works, who also serves as the CEO of Zero One AI.

On the financing front, Zero One AI became a "unicorn" with a valuation of $1 billion last November. According to public reports, it completed a new round of financing in August, with an amount reaching hundreds of millions of dollars. Participants in this round of financing include an international venture capital firm, a Southeast Asian consortium, and many other institutions. (See previous article on Titanium Media App: "Dialogue with Li Kaifu: The gap between Chinese and American large models is getting smaller, I haven't 'cashed out' in 10 years")

Currently, many Chinese AI large model companies such as Zero One AI, DeepSeek, MiniMax, and StepStar, adopt the so-called "Mixture of Experts" (MoE) model architecture. Some researchers believe that the MoE architecture is a key technology that achieves the same level of intelligence as dense models with less computing power. However, there is a higher risk of training failure with this method because it requires coordinating multiple "expert" models simultaneously during the model training process, rather than focusing on training a single model. Therefore, companies like Meta Llama in the United States have not developed related models, while Chinese companies like Zero One AI have created the world's fastest MoE model Li Kaifu believes that Yi-Lightning is a "top model at a bargain price". In terms of reasoning speed and price, Yi-Lightning's highest generation speed has increased by nearly 40%, costing only 0.99 yuan (14 cents) per million tokens, while OpenAI's smaller model o1-mini requires 26 cents per million tokens, and the inference cost of GPT-4o is $4.4 per million tokens, with the model pricing of Zero One Million still profitable.

Li Kaifu mentioned that the "pre-training" cost of the Yi-Lightning model is $3 million, which refers to the cost of the key training phase of the model, only 3% of training GPT-4 by OpenAI, and can be fine-tuned or customized according to different application scenarios afterwards.

Li Kaifu pointed out to Titanium Media App that Zero One Million is currently accelerating the commercial landing of large models, focusing on the overseas To C (consumer-level) paid market and the domestic To B (enterprise-level) paid market, such as releasing industry application product AI 2.0 Digital Human, focusing on domestic retail and e-commerce To B business scenarios. "**" All the generated answers rely on our Yi-Lightning large model, and the GMV sales of a wine and travel company have increased by 170%.

In March of this year, Li Kaifu pointed out at the Fortune Innovation Forum that too many AI large model startups focus on making breakthrough progress, rarely paying attention to the commercialization of their achievements. With the maturity of new technologies, AI companies that cannot make a profit are about to face a "reckoning". He emphasized, "The science fair phase must end."

In fact, if there is one common point among the three major tech giants in the United States, it is that they have successfully turned an emerging technology into reality—Microsoft with personal computers, Apple with smartphones, and Google with search ads and the Android system for smartphones, gaining a significant advantage in the internet and mobile internet era and landing emerging technologies.

Li Kaifu admitted, "**Google is a warning. Despite having the most dense AI talent network in the world today, he believes that Google lost to OpenAI because it wasted time and resources, indulging in competitive plans for all employees. 'If you have too many researchers and have formed a culture where everyone can try their own ideas, then as a startup, your funds will quickly run out.'"

Therefore, Li Kaifu stated that in order for Zero One Million (his company) to become a world leader in the field of AI one day, it must make extremely efficient use of every dollar. "We are taking the same approach, working very, very hard to save GPU computing costs."

"Investors will ask: What do you have to show? What is your profit and loss statement? What is your revenue? What is your growth rate? When will you break even?" Li Kaifu said that if an AI startup cannot provide convincing answers, then its "science fair" era is over Li Kaifu emphasized that the research goal of Zero One Infinity is not "regardless of how expensive or large, to build the world's number one model", but to build a world-class model with extremely low costs, capable of creating high cost-effective models, allowing developers to build applications without being overwhelmed by inference costs