Author丨Zhao Xiaoxiao
Editor丨Guan Ju
Source丨Zero 10,000 things
On May 13, at 10:00 a.m., Kai-Fu Lee appeared on time on the 20th floor of Block A of Dinghao Building, wearing a navy blue suit, a light-colored shirt with a tie, and a hairstyle, which was almost the same as that worn at a sharing meeting in March last year, with the only difference being an extra pair of glasses on his face.
This is the first public appearance in front of the media by Kai-Fu Lee a year after the establishment of Zero One Things, where he answered everything about the new product and Zero One Things, and responded for the first time to the controversy involving Llama's architecture last year.
Zero One Everything was launched in July last year, the first model Yi-34B was released in November, and half a year later, a new 100 billion parameter closed-source model Yi-Large and the first domestic to C product "Wanzhi AI" were released, and Kai-Fu Lee also personally served as the CEO of Wanzhi.
In Stanford's open-source evaluation project AlpacaEval 2.0, Yi-Large's main indicator of English proficiency, LC Win Rate, ranked second in the world, only slightly behind OpenAI's new version of GPT-4 Turbo released in April this year, and ranked first in Win Rate. In terms of Chinese ability, in SuperCLUE's updated April benchmark performance, Yi-Large ranked first among domestic large models.
斯坦福AlpacaEval 2.0 Verified认证模型类别，英语能力评测：零一万物的Yi-Large Preview 的 LC Win Rate为51.9%，排第二，Win Rate为57.5%，排第一。
One of the four overseas products that were launched last year has achieved an ROI (return on investment) of 1, which means that the growth can be more aggressive. The revenue of overseas products mainly comes from the payment of C-end users, "this year there will be 1-200 million yuan in revenue and nearly 10 million users." Kai-Fu Lee said.
It's a contest of speed and skill. Today's AI environment is completely different from the last time the Four Tigers era, it can bring real value, the critical point of the large model is the AI inclusive point, whoever ignites this point first, it will become a giant in the industry.
Some people call Kai-Fu Lee the "oldest" Chinese model entrepreneur, he is 62 years old this year, and he has become one of China's six large-scale model unicorns. Forty years ago, Kai-Fu Lee wrote in his doctoral dissertation application letter, "AGI is my dream. ”
"I started this business, and I won't cash out for ten years." Kai-Fu Lee said.
Microsoft in the AGI era
At a sharing session in March last year, Kai-Fu Lee defined AI 2.0 as "a revolution that must not be missed." "Its huge leap forward is that it overcomes the limitations of AI 1.0 in a single domain, multiple models, and no commercialization, and can bring real value.
"For example, in the past, AI applications for productivity could help me write reports, write articles, and make PPTs, and there was never a software that could help me analyze and summarize a thousand or even 10,000 pages of text." Kai-Fu Lee said.
At the end of 2022, ChatGPT was the fastest way to educate the market after its appearance. After the release of GPT4, it became "a gold standard" in the field of large models, and at the same time ignited the opportunity to B and to C.
If a company catches one wave of technology, it becomes a giant in the industry, and if it can catch both waves, it becomes a world-class company of excellence.
"So we're going to do it ourselves." Kai-Fu Lee said.
Kai-Fu Lee began his Ph.D. in computer science at Carnegie Mellon University in 1983 and has been researching artificial intelligence for nearly 40 years. He has been a supporter of entrepreneurs for more than a decade, and this time he came to the stage because he believed that this was the most important technology in human history and that he could not miss this revolution.
零一万物的起点是海外，to B。
In September last year, 010000 launched 4 products overseas, and the ROI (return on investment) of one of the productivity products has been achieved1. This means that the subscription fee received by Zero One World from users has been lower than the cost of acquiring users plus the cost of GPUs, getting rid of the "burning money to acquire customers." ”
Kai-Fu Lee and his team have a standard, generally they must see at least 0.6 ROI before they start pushing, otherwise the more users, the more money they will lose. If the product pushes to 1, it means that the growth can be more aggressive. Among the 4 overseas products, the ROI of entertainment and companionship is still at a low stage, and productivity products have brought 100 million revenue to 010 million and nearly 10 million users.
The open-source model Yi-34B, released in November last year, surpassed Llama2-70B, which was developed and open-sourced by Meta AI, and Falcon-180B, which was developed by the Institute of Technological Innovation in the United Arab Emirates, and became one of the strongest open-source basic models in the world at that time.
?? Hugging Face API ——Yi-Large Yi-Large-Turbo Yi-Medium Yi-Medium-200K Yi-Vision Yi-Spark," Reddit ”
"When we first launched, our self-developed model was not ready, and after Yi-Large was launched overseas, it will become a platform for us to compare and test our self-developed models in real time." Cao Dapeng, head of zero 10,000 things productivity products, said.
"Now we have the confidence to push the products that have reached PMF (Product Market Fit) back to China." Kai-Fu Lee said.
The overseas AI ecosystem has been very prosperous, many AI native applications have been born, and even reached tens of millions of dollars in revenue, and many mature products and companies have made their own second-curve new businesses using OpenAI's APIs. In contrast, the country will be a little slower, and it is still in the exploration stage.
The reason why Zero One Everything temporarily does not do domestic to B business and does not accept project-based orders is because Kai-Fu Lee feels that China has not escaped the curse of AI 1.0 last time - some very good startups have fallen into a vicious circle of high growth but low quality.
To C is relatively easy to do, Wanzhi is the first C-end product of 010000000, which accurately cuts into the three major needs of "finding, reading, and writing" people in the workplace, and is also the first to C application product in the industry with the concept of "AI-First" and the perfect integration of model capabilities and application scenarios.
"All applications should be AI-First, AI-Native, and it wouldn't be true without AI." Kai-Fu Lee said, "For example, the PPT production function is rarely seen at home and abroad. Former Google CEO Eric Smith affirmed that "information packaging is a $50 billion market."
Zero One will do a lot of things, overseas, domestic, to B, to C, it Day One is a global company. The company has launched the next generation of Yi-XLarge MoE model training, which will impact the performance and innovation of GPT-5.
"Our philosophy is to be the Microsoft of the AGI era." Kai-Fu Lee said.
Every company has to find its own TC-PMF
At any time, entrepreneurs must learn to seek the optimal solution of cost and efficiency in relatively sufficient resources.
Kai-Fu Lee defined a new term: TC-PMF - on the basis of finding the matching point between the product and the market, adding technology and cost, which is a four-dimensional match.
The emphasis on technology and cost is because Kai-Fu Lee believes that PMF can no longer fully define the entrepreneurship of large models, because it is necessary to take into account the current technical capabilities of the model, the future development and the estimated cost.
"There is a big difference between the current AI and the mobile Internet, that is, the mobile Internet has no inference cost." Kai-Fu Lee said.
For example, some applications require AI to not have illusions, so the current technology has challenges; Some applications require large models or long text windows, but this can introduce a 50-second inference delay, which can cost 75 bucks to process a million tokens.
In the era of mobile Internet, PMF used to be the core goal pursued by many start-ups, and the marginal cost of user growth was very low. However, in the era of large models, model training and inference are very expensive, which is a growth trap that every startup will face, and many companies will burn money indiscriminately and go bankrupt.
"So every company has to find its own TC-PMF." Kai-Fu Lee said. He called the point where TC-PMF is about to appear in an application field as the AI inclusive point, and only when this point is reached, can the most perfect match between technology, model and cost be achieved. Before this point comes, companies must first make products, wait for the opportunity, and become a Popular App.
There are two points in time when Douyin has become a Super app, the first point is the popularization of 4G, and the second point is the popularity of Android thousand-yuan machines, and the two points have come one after another, expanding the short video market by 100 times at once.
The fastest way to find it is to build your own inference engine, which requires a strong model team, an AI Infra (base layer) team, and an application team, and the three teams become one and cooperate with each other. This is something that sounds simple, but it is difficult to do, and it is also the biggest challenge encountered by enterprises in building their own inference engines.
In addition to the tacit cooperation between the teams, they pay special attention to the optimization of AI Infra: from selecting servers to designing computer rooms and deploying GPUs, to building vector databases, they are all doing it themselves, and they are all mastering the rhythm of reducing inference costs.
"The upside is that when someone else's reasoning costs are reduced tenfold, we may be twenty or thirty times lower." Kai-Fu Lee said. After optimizing the GPU cost in many aspects, the training of the Yi-Larege model with 100 billion parameters decreased by about double year-on-year. After connecting to the self-developed vector database, the cost is reduced to 18% compared with using a third-party vector database. Overseas TC-PMF has entered the stage of positive ROI.
I have to find TC-PMF by myself, but also because Kai-Fu Lee believes, "No technology can last a lifetime, and we must find barriers outside of technology." ”
For example, in the era of mobile Internet, why are Android, Apple, and WeChat so strong, and Douyin can still be killed?
It is because ByteDance predicted the use of AI to recommend short videos very early, found its own PMF, grasped the time window, and created a very high user stickiness and growth rate. When it became an inclusive app with a large number of users, WeChat could no longer replicate Douyin and kill it.
The decline in the inference cost of large models will promote China's AI large models to enter the stage of "landing is king", and Kai-Fu Lee judged that this year will usher in the "first year of the explosion of large model applications", and in the second half of next year, it will usher in a blowout period of inclusive application of large models.
"It's my dream to do AGI"
Kai-Fu Lee now has four identities: the founder of Sinovation Works, an investor, the founder of Zero One Everything, and the CEO of Wanzhi (Chief Experience Officer).
In 2009, Kai-Fu Lee established Sinovation Works in Zhongguancun, Beijing, which opened up a new model of venture capital and entrepreneurship in China, and he has been an investor for 15 years. In the AI 1.0 era, Sinovation Works has invested in more than 10 AI unicorn companies such as 4Paradigm.
Last year, at the age of 61, Lee chose to become an entrepreneur, which is already the retirement age for most people, and it was only because he saw the most important technology in human history that he decided to "try harder." ”
This is also the dream that Kai-Fu Lee wrote in his doctoral application letter more than 40 years ago, and I hope to have the opportunity to do something in the field of AGI in the future. Over the past few decades, he has been researching the possible impact of AI on humans, and has published three books on the subject, and most people's impression of his "elegance" stems from his status as a "writer". He also advised young people to learn more about AI and not just focus on "whether AI will replace our jobs." ”
"From the very beginning, I considered being the CEO myself," said Kai-Fu Lee, who felt that he needed to put more effort into finding problems and fixing them in time for the first to C product. From the launch of Wanzhi to the present, he will test more than 100 questions every day.
The most valuable modification suggestion is the lengthening of the search box, which is a problem he encountered when he was Google, the user could not search for something by typing two or three words in the search box, and the result became better by typing ten words, and the final solution was to make a longer search box.
According to Lan Yuchuan, the person in charge of the API platform of Zero One Things, the revenue of Zero One Things on the C-side this year will be 1-200 million yuan, and Microsoft's Office will have billions of dollars in revenue every year. "The ceiling in the productivity product market is very high, and we're thinking about how we can make the AI workspace bigger." Lan Yuchuan said that as a result, zero 10 things can expand overseas markets, such as the Middle East and Southeast Asia.
In the past year, Kai-Fu Lee has become a very popular figure, more than half of which is because Zero One Everything was launched in a short period of time, and a number of new products were launched in a short period of time, and he became a unicorn in the angel round. The other half is because of the controversy over the Llama architecture.
"The casing claim is purely false, and many technical experts wrote detailed articles to explain it at the time." Kai-Fu Lee said that the architecture of China's large models is based on the architecture of Transformer or Llama, and the technical community does not agree with the term Llama architecture.
"For large models, the higher weight is the training itself, as well as the parameters and code obtained through training. The YI model is trained from scratch and obtained its own parameters and code, without referencing the parameters of any Llama or other models. Kai-Fu Lee said.
Today's AI is a fully competitive market, and the competition is not only about technology, team, but also mentality. For AI, Kai-Fu Lee is a little paranoid, he focuses on doing a good job of the product, and leaving other things to the market to verify.
He also opened a Douyin account recently, but not to bring goods for live broadcasts, but to share good technologies and products. From the perspective of being an investor, Kai-Fu Lee said that if he had a lot of money in his hands, he would invest all of it in Zero 10,000 Things.
In this venture, Kai-Fu Lee's expectation is to "rush a trillion dollars", AGI is a big, big dream, and it can't just be to make some money.
He also promised investors that he would not cash out for ten years.