Published on

Lee Kai-fu's 01.AI Shifts Strategy: Focus on Practical Models Over AGI

Authors
  • avatar
    Name
    Ajax
    Twitter

01.AI's Strategic Adjustment: No Longer Chasing Super-Large Models

In a recent interview with LatePost, 01.AI CEO Lee Kai-fu detailed the company's recent strategic shift. The core change is that 01.AI will no longer pursue training super-large models. Instead, it will focus on developing models with moderate parameters that are faster and more economical, and build commercial applications on this basis. This shift marks the first time a Chinese large model unicorn has publicly made a significant adjustment to its development direction. It also reflects a significant turning point in the large model boom of the past two years.

Lee Kai-fu emphasized that 01.AI is not seeking to be acquired and will continue to conduct pre-training. The company has established a "Large Model Joint Laboratory for Industry" with Alibaba Cloud. Most of 01.AI's training and AI infrastructure teams will join the laboratory as Alibaba employees. This cooperation model aims to use the resources of large companies to train larger models, thereby improving the capabilities of 01.AI's own smaller models.

Challenges of Large Model Startups in China

Lee Kai-fu summarized several major challenges faced by large model startups in China:

  • Chip Restrictions: Chinese companies face restrictions in obtaining chips, resulting in lower financing amounts and valuations compared to their American counterparts.
  • Scaling Law Slowdown: The effect of the Scaling Law is weakening, transitioning from belief to skepticism in just one year.
  • Competition with Large Companies: Startups find it difficult to compete with large companies in terms of model scale, ultimately leading to a lack of success.
  • Commercialization Difficulties: How to transform technology into commercial value and achieve profitability is the core question for all large model companies.
  • Market Difficulties: There are bottlenecks that are difficult to break through in the To B, To C, domestic, and international markets.

01.AI's Response Strategy

Lee Kai-fu believes that 2025 will be a year of application explosion and commercialization elimination. 01.AI's opportunity lies in exploring the Product-Market Fit (PMF) of To B large models. He pointed out that in some niche areas, large models can help customers double their revenue, which is the real PMF.

After adjusting its strategy, 01.AI will focus on:

  • Training faster and cheaper models, such as MoE (Mixture of Experts) models.
  • Leveraging its advantages in AI infrastructure and inference engines to reduce training and inference costs.
  • Co-creating with industry companies to establish joint ventures and jointly develop niche industry models and solutions.

Reasons for Abandoning the Pursuit of AGI

Lee Kai-fu admitted that 01.AI had abandoned the pursuit of AGI (Artificial General Intelligence) early on. He explained that pursuing AGI requires a large amount of resource investment, and 01.AI's highest priority at this stage is to consolidate its own strength and achieve commercial profitability.

He reviewed the experience of 01.AI releasing the Yi-Large model last May and said that he realized at that time that the model was slow and expensive. This prompted 01.AI to make a decision: instead of burning money to train super-large models, it would focus on developing commercialized models that can be implemented and make money.

Cooperation with Alibaba

Establishing a joint laboratory with Alibaba Cloud is a key step in 01.AI's strategic adjustment. Lee Kai-fu said that this cooperation model can make full use of the advantages of both parties and accelerate the sharing and co-construction of technology, platforms, and applications, opening a new paradigm of cooperation between "large factories + small tigers" in China.

Although some pre-training and AI infrastructure teams will join Alibaba, 01.AI will still retain a smaller training team and infrastructure team to continue model development. Lee Kai-fu emphasized that 01.AI will not stop pre-training, but will no longer be obsessed with super-large models.

Slowdown of Scaling Law

Lee Kai-fu pointed out that the Scaling Law is slowing down. This means that the returns from investing more computing power and data are decreasing. He gave an example that adding from one card to ten cards may achieve the value of 9.5 cards, but adding from 100,000 cards to 1 million cards may only achieve the value of 300,000 cards.

He also mentioned that internet data resources are like fossil fuels, which are gradually being depleted. This makes the training cost of super-large models increasingly high and the returns increasingly low.

The Role of Super-Large Models

Despite the slowdown of the Scaling Law, Lee Kai-fu believes that super-large models still play an important role, especially as teacher models. He pointed out that Anthropic's Opus model is used to train smaller models.

Super-large models can improve the capabilities of small models in the following ways:

  • Labeling results to improve post-training effects.
  • Generating synthetic data for training new models.

The Soul-Searching Question of Commercialization

Lee Kai-fu believes that everything is accelerating in the era of large models, and the question of commercialization is coming faster. He emphasized that AI companies must answer the core question: how to transform technology into commercial value and achieve profitability.

He proposed that AI companies need to:

  • Understand business operations.
  • Achieve revenue growth.
  • Control costs.

Lee Kai-fu also emphasized avoiding investing in commercialization directions with no return, such as To C applications that require continuous blood transfusions and losses to maintain industry status, and To B bidding projects with low payments and do not create core value.

01.AI's Commercialization Path

01.AI is actively expanding into the To B market and is experimenting in areas such as gaming, energy, automobiles, and finance. They will co-create with industry companies, establish joint ventures, and jointly develop niche industry models and solutions.

Lee Kai-fu stated that 01.AI's actual revenue in 2024 has exceeded 100 million RMB and is expected to double several times in 2025.

The Future of AI-First Applications

Lee Kai-fu believes that disruptive AI-first applications will definitely emerge. He pointed out that these applications need to have the following core characteristics:

  • Interact using natural language.
  • Possess general reasoning and understanding capabilities.

He also gave a judgment method: if an application cannot be established without a large model, then it is definitely an AI-first application.

Lee Kai-fu's Entrepreneurial Insights

Lee Kai-fu said that he embarked on AI entrepreneurship to seize the opportunities of the AI era and transform his experience and abilities into value. He believes that challenges are inevitable in the entrepreneurial process, but a good CEO should not easily regret it.

He summarized his entrepreneurial insights:

  • Do not blindly invest in impossible goals.
  • Seize opportunities and make decisive decisions.
  • Have a clear prediction of the future and make adjustments in advance.

Outlook for 2025

Lee Kai-fu is full of confidence in 2025. He predicts:

  • A large number of To C applications will explode.
  • The PMF of To B large models will be discovered, and many niche industry models will also emerge.

He also stated that 01.AI is exploring the application of Agents and will jointly develop industry models + Agents in vertical fields with partners.