Reporter: Ke Yang, Editor: Zhang Haini
On the evening of April 11th, the “Daily Economic News” learned from an internal speech given recently by Robin Li, who addressed pivotal industry topics such as the choice between open-sourcing and keeping AI models proprietary, and whether AI entrepreneurs should focus on developing models or applications.
In his speech, Li emphasized that proprietary models consistently lead in capabilities over time rather than just temporarily; the open-sourcing of models does not necessarily lead to superior outcomes as seen with traditional software like Linux or Android.
Li also argued that proprietary models not only have a genuine business model that allows them to generate revenue, attracting computing power and talent, but they also have cost advantages. With equivalent capabilities, proprietary models invariably incur lower operational costs and are faster in response times.
Furthermore, Li pointed out that whether in China or the US, the most advanced foundational models are proprietary. Models derived from these foundations tend to be superior, giving proprietary models an edge in terms of cost and efficiency. For AI entrepreneurs, the focus should not be on the model itself, as it is resource-intensive and requires long-term commitment.
He believes that adopting a dual approach of developing both models and applications is not effective for startups due to their limited resources and capacity. Concentrating on one over the other significantly increases the chance of success.
Excerpts from Robin Li’s internal speech:
Why Not Open Source?
The market already has enough open-source large models
A year ago, when Wenxin was first released, there was intense internal debate at Baidu. Ultimately, we decided against open-sourcing. The market will inevitably have open-source models from not just one but multiple providers. Adding or subtracting one Baidu wouldn’t make much difference.
Proprietary models will continue to lead
More importantly, in our view, proprietary models will maintain a lead in capabilities. The open-sourcing of models isn’t particularly impactful as these are often piecemeal efforts lacking extensive computational validation.
Proprietary models support a real business model and gather talent
When comparing, for instance, offers from OpenAI, Meta, and Llama, the choice for top talent in Silicon Valley is clear. This obvious preference is because proprietary models can generate revenue, which in turn attracts computational power and talent.
Proprietary models also have cost advantages
While it’s assumed that open-source models are cost-effective due to their free nature, proprietary models, by contrast, often cost less in terms of operational expenses for equivalent capabilities.
Entrepreneurs should rely on Wenxin
Dual focus is not a good model
The so-called “dual-wheel drive” of working on both models and applications spreads a startup’s limited resources too thin. Focusing on one area clearly leads to higher success rates.
The core competitiveness for AI entrepreneurs isn’t the model itself
Instead, it lies in domain-specific knowledge and data. For instance, if you were searching for “yellow men’s swim trunks without pockets” on today’s e-commerce platforms and couldn’t find any, a large model with domain knowledge could potentially fulfill this unique demand.
Foundational models won’t monopolize AI applications
There’s no need to fear that successful applications will be copied or overtaken by foundational models, as has been demonstrated in the mobile era with platforms like WeChat not overshadowing entities like Pinduoduo or Didi, which thrive by offering unique values within a closed platform.
Daily Economic News