Robin Li: Baidu Has Stepped in These Pits During the AI Transformation Era
At the Baidu World Conference, Robin Li stated that the next act for the industry is the intelligent organization, and reviewed the pits Baidu has stepped in during the AI transformation: in the early days, engineers were keen on using open-source models for SFT, but after model upgrades the results quickly fell behind, so Baidu required business departments not to train models on their own, with a dedicated model department taking unified responsibility. After new hot spots such as agents emerged, business departments rushed to follow suit, leading to scattered resources and lower success rates; Baidu subsequently required focus through monthly meetings, narrowing its efforts "from many to few."
Topics
Body · AI translation
On November 12, at the Baidu World Conference, Baidu Chairman Li Yanhong said that the next act for the industry is the intelligent organization, and he discussed some of the pitfalls Baidu has encountered in the era of AI transformation.
He said that when large models first emerged, Baidu engineers were keen on using open-source models for SFT (supervised fine-tuning), and the results in business applications were once good, but after the models were upgraded, the original results quickly fell behind. For this reason, Baidu required business departments not to train models on their own, but to have a dedicated model department handle it uniformly, to avoid "entropy increase" and improve efficiency. Later, when new hot spots such as agents emerged, business departments rushed to follow up one after another, leading to scattered resources and a lower success rate instead. Baidu then required focus through monthly meetings, turning the direction of efforts "from many to few," so that everyone would do fewer things in order to improve efficiency and the success rate. (First Financial Reporter Liu Jia)