Baidu’s bet on AI could make or break China’s fallen tech group

Robin Li stepped up to podiums across China this autumn with a message to the country’s entrepreneurs, chief executives and officials: stop building large language models.

In a series of speeches, Baidu’s founder told potential rivals “there’s no point in competing” to build generative artificial intelligence technology that can spew out humanlike text, code and images in seconds.

He urged them to instead build apps for existing models, such as his company’s “Ernie”. With each speech, Li also ruefully counted greater numbers of homegrown rivals, from about 70 large models in September to 238 in November.

That surge in competition is, once again, threatening the Chinese search group’s ambitions in emerging tech. Over the past decade, Baidu has tried everything from ecommerce to building a mobile operating system, before subsequently retreating amid an onslaught of new entrants.

“It’s a do-or-die moment,” said one engineer on Baidu’s large language model team. “Whether Ernie succeeds or not, will play a decisive role in the company’s future.”

More than a dozen current and former Baidu employees who spoke with the Financial Times said issues with Li’s leadership and company culture had contributed to the group’s difficulty incubating new business lines over the past decade.

They praised Li’s devotion to advancing technology and his work ethic but said Baidu was plagued by a lack of product focus and high turnover especially among its most talented staff — problems that threaten its nascent efforts in generative AI. Some added that Li and his wife Melissa Ma’s joint presence at the company caused issues for subordinates.

The company declined to make Li or Ma available for an interview. In response to a request for comment on detailed points before publication, Baidu said “the article presents a heavily biased and factually inaccurate depiction of the company and its CEO”.

Baidu warned that it could not comment on the claims of those who declined to be named, as it was impossible to “verify their claims or identify any personal prejudices which may colour their accounts”.

The return of Ma

Baidu was once known as the “B” in BAT, beside Alibaba and Tencent in the acronym synonymous with Big Tech in China. But over the past decade revenues at its two rivals have surged, while Baidu’s sales have only nudged upwards.

About five years ago, Ma returned to help right the ship. Officially given the title of special assistant to her husband, her authority within the company soon exceeded that modest designation. Today she is involved in investments, talent management and recruiting, according to people close to the company.

Baidu’s SEC filings indicate Ma has received about $1.7mn annually on average in stock compensation since rejoining. She is also the second-largest shareholder by voting rights and has sold more than $500mn worth of shares since 2020, according to data provided by the Washington Service.

Baidu, without providing specifics, said details that the FT put to the company about Ma included “many errors of fact”. The company said she provided advisory services to Baidu’s HR department and was not involved in day-to-day operations nor among its core management.

“Suggesting Ms Ma is an executive officer of Baidu, based on her share-based compensation, is mistaken,” the company added.

Several staff said the husband and wife relationship had, at times, complicated decision making. “It’s like there are two factions,” said one employee. “Sometimes they disagree with each other and are evenly matched.”

Ma’s return also contributed to the exit of one of Baidu’s star hires, Lu Qi, an AI expert whom Li poached away from Microsoft in 2017 to run the company’s day-to-day operations as president and COO, according to two people familiar with the matter.

Lu “couldn’t figure out who to deal with, whether Robin or Melissa was the boss”, said one of the people. He left the company 16 months after joining. Baidu said Lu departed “to deal with a personal family matter”. Lu did not respond to a request for comment.

Since 2017 more than 20 executives of vice-president level or higher have cycled through the company, according to an FT review of public departures. 

Baidu said it “believes in enabling every talented individual to reach their full potential”, and “it is this strong, empowering culture that has helped us retain so much of our core talent for extended periods”.

Switching focus

Faltering strategy has claimed many of Baidu’s business aims. A former mid-level manager attributed the frequent course corrections to Li’s preoccupation with technology rather than making money.

Ticking off several businesses that Baidu had tried, the manager noted: “When [Li] saw we didn’t have a big advantage from tech, he decided it was just a business and lost interest.”

The company is now resetting ambitions for its intelligent driving unit, which faces regulatory headwinds and a crowded market.

Baidu in recent years has repeatedly trumpeted the launch of “fully driverless” robotaxis in cities including Beijing and Wuhan, but Chinese officials still largely require human drivers to remotely monitor every ride, undermining the economics of a self-driving taxi fleet.

Over the summer, when the FT visited the group’s autonomous driving centre in Beijing, remote safety monitors sat in arcade-like driving simulators ready to hit the brakes or take the steering wheel to override Baidu’s self-driving software.

The group’s plan for its own electric car also hit a speed bump in the autumn, when Baidu was forced to rename its first consumer vehicle and turn over control to partner Geely in order to obtain regulatory approval to begin production. Baidu hopes the car will spur interest in its self-driving software from other automakers.

With Li shifting his attention to large language models, lay-offs have begun to trim staff numbers and several auto executives exited. Li in November reassigned the unit’s longtime head Li Zhenyu to a rotational role as his assistant and chair of a “technology ethics committee”.

“Resources are being shifted to Ernie,” said one person close to the self-driving project. Two employees said that shift also included many of the driverless car team’s powerful Nvidia chips.

The long journey to Ernie

Li’s determination to build Baidu into an AI giant emerged at least a decade ago, when he led the group to bid for DNNresearch, the start-up founded by AI pioneer Geoffrey Hinton. Hinton ultimately sold to Google.

Yet, despite both Baidu and Google spending a decade investing in AI, they were left flat-footed when Microsoft-backed OpenAI released ChatGPT in November 2022.

After its staff worked around the clock for weeks, the company released Ernie in March and tens of thousands joined a waiting list to test the country’s first generative AI chatbot. But Baidu was soon backpedalling, trying to explain why Ernie had generated an image of a “patriotic cat” cloaked in an American flag or cited US intelligence findings to claim Huawei products contained backdoors.

The problem stemmed from the reams of data Baidu’s team fed into Ernie’s models. Employees said a lot of the material was in English and from western open-source databases.

Some Baidu staff were concerned that the chatbot would inadvertently cross China’s redlines for permissible speech, especially after the country’s powerful cyber regulator released new rules demanding generated content “embody core socialist values” and stipulated that tech companies would be held responsible for the output.

Staff wrote to the regulator to lobby against holding tech groups responsible for AI content. One employee said Baidu was now using other internal models to check content generated by Ernie.

Since March, Baidu has released a series of upgrades, most recently unveiling Ernie 4.0, which Li claimed was comparable to OpenAI’s GPT-4.

But the local competition is now rapidly catching up, too. SuperCLUE, a benchmark set up by a group of AI experts evaluating foundational models that use Chinese, ranked Ernie 4.0 just a few points ahead of rival models from Tencent and Alibaba in November.

Li told investors in November that Baidu’s stash of powerful American AI chips provided a competitive moat and predicted that the difficulty obtaining semiconductors, data and talent would soon lead to consolidation in the industry.

“We believe there will only be a select few foundation models in the market and Baidu will certainly be one of them,” he said. He predicted “millions” of apps would eventually be built on top of Ernie, with outside groups paying each time their users run queries.

Baidu said it had “successfully” transitioned from being an information services company to becoming a leading AI company, adding: “Strategic shifts like these take courage, vision, and most importantly, time.”