On May 13th, 01.AI, a large-scale startup company founded and led by Kai-Fu Lee, officially released the closed-source billion-parameter large model Yi-Large. In the AlpacaEval 2.0 English proficiency ranking of Stanford’s open evaluation project, this model’s LC Win Rate (win rate controlling text length error) is second only to Open AI’s GPT-4 Turbo, while its overall Win Rate ranks first in the world.
Although this company has mainly focused on releasing open-source medium and small-scale models in the past, the team recognizes the value of closed-source large models in practice. Kai-Fu Lee also revealed that 01.AI has initiated training for the next generation trillion-parameter Yi-XLarge MoE model.
In the past, one of the most talked-about terms among entrepreneurs in the field of large models was PMF (Product-Market Fit), but Kai-Fu Lee believes that this concept from the era of mobile internet is no longer suitable for the AI era.
In the era of mobile internet, the low marginal cost brought by user scale growth. However, in the era of large models, model training and inference costs are challenges that every startup must face.
“The high-quality applications required for user growth cannot be separated from powerful base models, which often come with high training costs. In addition, one must also consider the inference costs that increase with user scale.” Kai-Fu Lee said that Product-Market Fit (PMF) is no longer enough; one must also consider the matching of Technology and Cost.
Therefore, he proposed a new concept – TC-PMF, the technical cost X product market fit. ‘Doing TC-PMF, the reduction in reasoning costs is a ‘moving target,’ which is 100 times more difficult than traditional PMF,’ said Kai-Fu Lee.
Precisely because of this, Kai-Fu Lee does not approve of the current practice of some large model companies buying traffic for promotion. He stated that the industry should collectively avoid falling into irrational practices like burning money in an ofo-style manner.
In order to develop and compete in the next stage, 01.AI has identified four clear directions including global layout, model-base collaboration, integration of model and application, Al-First.
In terms of globalization, 01.AI has launched 4 consumer products overseas. Among them, the productivity product line is expected to generate revenue exceeding one hundred million yuan this year with an ROI (Return on Investment) of 1. The user base for overseas productivity applications is close to tens of millions.
Joint construction of the model mainly refers to self-developed artificial intelligence infrastructure technology, which can optimize training methods from the bottom layer and improve computing efficiency. 01.AI stated that after optimization, the training cost of its billion-parameter model decreased by half compared to the previous year. The integration of models and applications mainly indicates that the capabilities of basic models should be systematically linked with application effects.
The industry is currently quite certain that there will soon be an explosion in application layers in the field of large models. Kai-Fu Lee has clearly stated before that 2024 will be the first year of explosive growth for AI applications. Allen Zhu, managing partner at GSR Ventures, has also publicly stated multiple times that AI large model applications will explode in 2025.
Is the time difference between two conclusions significant? Kai-Fu Lee said that Allen Zhu’s point makes sense. Currently, the cost of reasoning has decreased by 10 times in a year, which means it will be 100 times lower in two years. If GPT-4’s ability to reason costs only one percent of the price, many applications are sure to explode.
“But I think this year is the first year because some fields will explode this year.” Kai-Fu Lee believes that not all areas will progress simultaneously; some fields have high demands for applications. He used productivity tools as an example and mentioned that they have already reached TC-PMF but can still improve further.
01.AI released its productivity application ‘Wanzhi’ for consumers a week ago, and it has now officially started providing services. The ‘API Open Platform’ targeting enterprise ecology is also launched simultaneously.
SEE ALSO: 01.AI Launches Personal AI Workspace ‘Wanzhi’
As for when the real super app will appear, Kai-Fu Lee said that the team can already predict the completion time of Yi-XLarge training and the growth of reasoning ability. There is an internal methodology where they make final predictions based on interim results while refining products based on those predictions.
The field is moving so fast that we can’t wait for the model to be completed before thinking about TC-PMF,” he said. “Every second counts now.”
Since its establishment a year ago, 01.AI has received two public investments from Sinovation Ventures and Alibaba Cloud. It is reported to have a valuation of over 1 billion US dollars, making it one of the members of the domestic unicorn club in the large-scale model field.
Founder and CEO Kai-Fu Lee seems to have fully adapted to the transition from investor to entrepreneur. Recently, he just registered a Douyin account and started rhythmically posting short videos. In fact, he has been invited many times, but now in order to share his own products and technology, he is clear that “this is still necessary”.
We (China’s large models) have been sprinting for a year, catching up with the top large models in the United States at least in terms of modeling. Of course, we need to continue to work hard in the future, but we need to have accumulation and vision for the future. We cannot just run desperately; instead, we need to start thinking about how to enter into a long-distance running mode,” said Kai-Fu Lee.
For this new battle in life, he does not hesitate to express his confidence and determination.
“This has been my dream for over 40 years. I wrote in my doctoral degree application letter that AGI is my dream,” he said. “As for realization, I did make a commitment with my investors a year ago, which is that I will not cash out for ten years.”
>>> Read full article>>>
Copyright for syndicated content belongs to the linked Source : Pandaily – https://pandaily.com/01-ai-releases-a-trillion-parameter-closed-source-large-model-yi-large/
