news

"the ceiling of big models is very high, and we are still very far from the ideal situation," baidu ceo robin li said in an internal speech on september 11. he bluntly said that the outside world has three cognitive misunderstandings about big models: "the gap between different models is not getting smaller, but getting bigger," "computing power is a key factor in determining the success or failure of big models, and open source models cannot solve this problem," and "it is not a consensus that intelligent agents are the most important development direction of big models."

the topic of big models is always new and always discussed. the debate over open source and closed source, the value of charts, etc. are all reasons. li yanhong, who saw the potential in the big model track early on, is happy to share his views.

"every time a new model is released, people want to say how good it is. every time they compare it with gpt-4o, they use test sets or make some rankings to say that my score is almost the same as its, or even that my score in some individual items has exceeded it. but this does not prove that these newly released models are not that far behind the most advanced models of openal," said robin li, who believes that the outside world has a misunderstanding of large models.

in his opinion, the gap lies in capabilities and costs. "from the rankings or test sets, you think the capabilities are very close, but there is still an obvious gap in actual application. i don't allow our technical staff to compete for rankings. what really measures the capabilities of the wenxin big model is whether you can meet user needs and generate value gains in specific application scenarios. this is what we really care about."

value and scenarios are also often mentioned by peers. "enterprises embrace ai not to pursue cool technology, nor to 'find nails with a hammer'. the core is to solve business pain points. since last year (2023), everyone has been a little too optimistic about big models, thinking that they can quickly change the world. recently, there has been some pessimism, thinking that big models look good but are not practical. 'overestimating progress in the short term and underestimating results in the long term' are actually not advisable," said tang daosheng, senior executive vice president of tencent group and ceo of the cloud and smart industry business group, from the perspective of customers. scenarios are the key to unlocking ai. it is best for enterprises to combine unique professional data to find opportunities to reduce costs and increase efficiency in existing workflows and business scenarios, and then continue to improve and invest in the long term.

open source and closed source has been a long-debated topic. as a representative of closed source, robin li made another comparison, "in addition to capability or effect, a model also needs efficiency. the open source model is not efficient."

"to be precise, the closed-source model should be called a commercial model. the commercial model is that countless users or customers share the same resources, share the r&d costs, and share the machine resources and gpus (graphics processors) used for inference, while the open-source model requires you to deploy a set of things yourself." he further said, "before the era of big models, everyone was used to open source meaning free and low cost. but these things are not true in the era of big models. in the era of big models, people often talk about how expensive gpus are. computing power is a key factor in determining the success or failure of big models. does the open-source model provide you with computing power? if it doesn't provide you with computing power, how can computing power be used efficiently? the open-source model cannot solve this problem."

two months ago, robin li stated at the 2024 world artificial intelligence conference that he was most optimistic about intelligent entities. in early september, many entrepreneurs and experts at the 2024 inclusion bund conference believed that intelligent entities are a new terminal form that will give birth to a new generation of super platforms.

regarding this increasingly heated topic, robin li spoke out again, "why do we emphasize intelligent agents so much? because the threshold for intelligent agents is indeed very low. intelligent agents provide a very direct, very efficient and very simple way. it is quite convenient to build intelligent agents on top of the model," he asked and answered himself.

he believes that "the judgment that 'intelligent agents are the most important development direction of big models' is actually a non-consensus. to date, intelligent agents are still not a consensus, and there are not many companies like baidu that regard intelligent agents as the most important strategy and the most important development direction of big models."

beijing business daily reporter wei wei