Robin Li of Baidu discusses choosing closed-source for AI development in a visual from China.

Robin Li, Baidu’s co-founder, chairman, and CEO, recently shared his stance in an internal speech on whether to open-source their large-scale models, a hot topic within the tech industry.

On April 11th, news emerged that Li discussed the reasons behind keeping their AI model, ERNIE, closed-source. He recalled, “When ERNIE was first released a year ago, we had intense internal discussions. Ultimately, as you all know, we decided to keep it closed-source. Why? At that time, we believed that the market would undoubtedly have open-source models from not just one but several entities. In such a scenario, whether Baidu opens its model or not really doesn’t make a big difference.”

Li acknowledged the significant influence of models like Meta’s Llama and Mistral’s (a model by a French AI startup) overseas, as well as domestic models like Zhiji, Baichuan, and Alibaba’s Tongyi, noting that the market isn’t lacking in open-source models without Baidu’s contribution. He argued that maintaining an open-source version would not be cost-effective for Baidu.

Li views open-source models as scattered and small-scale in their approach to validation without the rigorous testing larger computation power affords. He pointed out that open-sourcing doesn’t necessarily lead to a collaborative development that significantly pushes the model forward, remarking that the main contributors to models like Llama are still primarily the developers at Meta, rather than a community-driven effort.

From Li’s perspective, maintaining a closed-source model is crucial for a viable business model that can attract talent and computational power. “Closed-source is profitable. And profitability allows us to gather computational power and talent. It’s pretty straightforward—if you compare, say, a talented individual in Silicon Valley with offers from OpenAI, Meta, and Llama, their choice is clear,” explained Li.

Besides attracting talent and computational power, Li highlighted the cost advantages of closed-source models. He asserts that closed-source models at equal capability have lower inference costs and faster response times, offering stronger capability under the same parameters.

Li stated, “Whether in China or the U.S., the most powerful foundational models today are closed-source. Smaller models, even the best of them, have been distilled from these larger models. This approach not only results in better models but also provides cost and efficiency advantages.”

Following the official announcement of ERNIE and the ERNIE Bot in August last year, Baidu has been integrating AI capabilities into various consumer-facing products and launching a one-stop large-scale model platform for enterprises through its cloud service.

Baidu’s 2023 R&D spending reached 24.2 billion yuan, up 4% from 2022. This was primarily due to depreciation expenses for servers supporting generative AI research and costs associated with server racks. Baidu noted a significant increase in the use of ERNIE and its API calls, with inference costs of the ERNIE model reduced to 1% of the original costs from March through December 2023.

Li has consistently emphasized the application aspect of large models. He believes that for startups, chasing large models is less meaningful than seizing opportunities in applications. The “dual drive” of working on both models and applications is not an ideal model for startups, which should rather focus on leveraging their knowledge and data in specific domains to creatively use a mix of available models, offering value that can differentiate them.

Concerning worries that foundational models might monopolize AI applications, Li used Pinduoduo and Didi as examples to illustrate that vibrant ecosystems can support multiple players who provide unique value and have their competitive strengths. He argued there’s no need to fear foundational models overshadowing all AI applications.

Baidu’s fourth-quarter and annual report for 2023 revealed a total revenue of 134.59 billion yuan, a 9% year-over-year increase; net profit attributed to Baidu (non-GAAP) reached 28.7 billion yuan, up 39%. The company’s annual revenue and profits exceeded market expectations. Li predicts that generative AI and foundational model services will contribute billions in revenue to Baidu in 2024, positively impacting the company’s total income.

As of December 2023, Baidu App’s MAU reached 667 million, up 3%.

At the time of this report, Baidu’s stock saw a 0.31% increase in the U.S. stock market. The previous trading day, Baidu shares dropped by 1.83%, closing at $101.61 with a total market value of $35.63 billion. As of the 11th, Baidu’s Hong Kong stock fell by 1.18%, closing at HK$100.80 with a total market value of HK$282.8 billion.