LinkedIn Turns 20. Its Next Career Move: A Big AI Push

In 2009, LinkedIn’s new engineering chief, David Henke, assembled his full crew of coders and managers for the first time and fired off tough questions. “What the heck is wrong with all of you?” was the sanitized gist of it. The fast-growing professional social network had about 50 million users, but every Thursday afternoon it went completely offline as engineers launched new features and fixed bugs. Job seekers couldn’t browse openings. Recruiters paying to trawl the website for candidates were forced to twiddle their thumbs. Clicking on a profile would summon the “wizard of [in],” a jaunty, staff-wielding mascot akin to Twitter’s famous “fail whale.”

Henke wasn’t amused. The weekly downtime, which included an office happy hour to lubricate a long night’s work, stemmed from LinkedIn’s scrappily built technology. That was standard practice in startups, but the network was now big and important enough that unreliability threatened to stunt its growth and scare off clients, Henke thought. That could choke a healthy business that had sought profits early on to win over venture capitalists who were still spooked by dotcom busts.

During the next few weeks of around-the-clock coding, the engineering team rebuilt LinkedIn’s technology from the ground up, significantly reducing downtime. Company veterans recall the project as one of the most pivotal in the company’s 20-year history. Without the focus on “site up,” LinkedIn may have struggled to become the giant it is today, says Deep Nishar, its product head from 2009 to 2014. “The success definitely has underpinnings in the technologies that we invested in at that time,” he says.

Launched on May 5, 2003, LinkedIn will be the first of a string of social media giants born from the spoils of the early 2000s tech crash to transition out of its teens, with Facebook, YouTube, Reddit, and Twitter also turning 20 over the next three years. Although those incumbents are proven survivors, troubles at Twitter under Elon Musk and the rise of challengers such as TikTok and Bluesky make social media feel more competitive than it has for years.

Looking back on LinkedIn’s coming of age, several current and former executives recalled some defining moments to WIRED. They also spoke about LinkedIn’s preparations to survive another 20 years by integrating AI “copilots,” as Microsoft likes to describe virtual assistants, and helping make hiring more fair.

LinkedIn emerged from a small group of Silicon Valley product managers and engineers, including founding CEO Reid Hoffman, who wanted a way to track their connections in the startup community to find the hot places to work and bring friends along. Hoffman, who later became chairman, stepped away to focus on investing after Microsoft bought the service for $26 billion in 2016.

Today, LinkedIn has more than 930 million users. Its revenue rose 34 percent to nearly $14 billion in the year ended last June 30—a much smaller business than Meta, whose annual revenue was nearly $117 billion last year, but one that has a near-monopoly on its business-minded user base.

Back in 2009, Henke had joined LinkedIn as engineering boss from Yahoo. LinkedIn’s then CFO, Steve Sordello, pitched Henke for four hours on the potential revenue growth from effectively digitizing résumés and workplace friendships, and the lack of real competition. “He opened up the books,” Henke says.

LinkedIn had begun selling premium services such as more powerful searches of user profiles to recruiters, researchers, and businesses much earlier than other social networks such as Facebook rolled out ads or subscriptions. It had worked, but Henke found that poor engineering practices festered along the way and once-modern technology had become old-fashioned.

The issues were hard to miss. Office TVs displaying usage statistics would fill with squiggly red lines, denoting failed sign-ups due to overloaded servers, Nishar says. The year before Henke joined, the opening of LinkedIn’s second data center had gone so badly that executives almost had to call the board in the middle of the night to say the project had failed because of a database bug, says Igor Perisic, LinkedIn’s former chief data officer.  

In another case, functionality for one of LinkedIn’s paid search tools depended on a data scientist’s laptop that she accidentally took offline while a LinkedIn salesperson was making a demo to a potential client. “It's one of these moments where you say, ‘Oh, people are actually using this,’” Perisic says.

Henke’s order to urgently end maintenance outages and prioritize uptime above all else was not instantly carried out. Mohak Shroff, then a senior software engineer but now LinkedIn’s engineering chief, went to Henke soon after that first meeting and asked him to approve some updates that required downtime. “He's like, ‘Didn’t you hear what I said? The site doesn't go down,’” Shroff says. To underline his point, Henke temporarily barred all updates. 

Instead, engineers rewrote code, redesigned data storage systems, and replaced complex hardware with more affordable and flexible options so that LinkedIn could better handle spikes in usage, Shroff says. Finally, updates could be made without user interruption, and any errors could be undone. LinkedIn went public in 2011—a year ahead of Facebook but with a disclosure to investors that downtime was a significant risk. Its shares more than doubled on their first day, and the service crossed 100 million users that year.

But LinkedIn’s newfound stability begat a new issue. Fearing being the one to knock LinkedIn offline, teams grew risk averse, drawing hundreds of people into meetings to approve launches. “Terrible memories,” Shroff says of that time. 

During a smoke break in 2011, Henke tapped Shroff to solve the logjams blocking innovation. Shroff asked for 30 people and six months. Henke gave him 300 people and six weeks. “My head just starts spinning,” Shroff says. Project InVersion, as it was called, was born. 

LinkedIn paused code updates even longer this time around. Henke’s engineers built new infrastructure to support A/B tests and other mechanisms that allowed product teams to make and test incremental changes and measure how users responded. “It led to 2012 and beyond being some of the most productive years of our history, in terms of just the sheer volume of innovation,” Shroff says. 

In those subsequent years, LinkedIn added multimedia support to its home feed and profile pages and redesigned search to deliver more personalization and filtering options. Like equivalents at Facebook, the newly created internal tools at LinkedIn soon were emulated at other fast-expanding internet services.

As LinkedIn picked up the pace, some cracks unsurprisingly surfaced. LinkedIn suffered a data breach in 2012, compromising millions of users’ personal information just as it had begun adding protections to make data undecipherable to hackers, according to Nishar. In a later incident, a bug allowed a recruiter to click thousands of pages deep into search results, frying the memory on LinkedIn’s servers and bringing down the search feature, Perisic says.

Meanwhile, usage from smartphones began soaring, but LinkedIn’s mobile apps were buggy and disjointed. In 2014 the company, as part of what it called Project Voyager, started over and centralized functionality into one app for consumers and separate apps for recruiting and sales clients paying for premium functions.

LinkedIn retrained large cohorts of engineers on mobile development and for a year crammed them all onto one floor as they rewrote every line of mobile code over long nights and weekends, says Erran Berger, vice president of product engineering. The new LinkedIn app, which was easier for the company to update and better monitored user behavior, launched in 2015. 

About the same time, LinkedIn became one of the earliest big tech companies besides Google to increase investment in a revolution taking place in machine learning. It had tapped the technology early on, launching  its signature “People You May Know” feature in 2007, which recommends new connections. But tight computing capacity initially meant “you would see the same three names for three months,” Perisic says. 

That limitation led LinkedIn to create a new data-logging technology, called Kafka, to count how many times someone had seen a particular suggestion, to ensure they wouldn’t get tired of the same old faces. Kafka, made open source in 2011 and named after the Metamorphosis author Franz Kafka, is now widely used across the tech industry for tracking activity, like views of an ad. 

LinkedIn’s precocious embrace of algorithms also gave the company an early taste of the ethical questions AI can raise. The company became the butt of jokes for encouraging spammy connection requests through “People You May Know.” LinkedIn had trained the system to recommend users with whom a viewer would likely try to connect with, not connections welcome on both sides. An update around 2015 changed the feature to show people likely to accept an invitation to connect. “That simple change made such a big difference,” says Ya Xu, LinkedIn’s head of data and AI.

Tuning connection requests has been just one part of LinkedIn’s grappling with its power over the job market. Any unfair skew in LinkedIn’s algorithms for search results or candidate or job recommendations could sway who gets hired and by whom. 

Around 2018, Xu’s team decided to prioritize equal treatment for job seekers, meaning they don’t allow LinkedIn’s AI systems to be biased by factors such as demographics. Managers screening candidates on the site can also choose to hide candidates’ profile photos, to cut the potential for bias in hiring decisions. In addition, the company has created a tool that helps engineers spot when a change increases inequality among users.

These days, LinkedIn’s top brass, including CEO (and college dropout) Ryan Roslansky, are championing an effort to motivate recruiters to review candidates based on skills such as budget management or speaking Spanish, not educational backgrounds or past employers, to create more equitable hiring practices. As of last month, more than 45 percent of hirers trawling LinkedIn dialed in on skills data through screening questions or search filters. Many more of LinkedIn’s users now add skills to their profiles alongside conventional job and education data, and the company’s algorithms can link those to job descriptions even when they don’t explicitly list the skills required. “It's a huge win for us to be changing how customers are starting to hire,” Xu says.

LinkedIn has long been interested in AI, but pressure to embrace it now comes from high up as well. Microsoft CEO Satya Nadella has said he expects every application to be powered by AI, and LinkedIn’s features have been no exception.

Microsoft introduced a chatbot for Bing search that uses startup OpenAI’s GPT-4 language model, and it also announced features it calls “copilots” for software such as Word, where the tool can draft an entire document from just a brief prompt. LinkedIn’s copilots so far include AI programs that use GPT models, which power ChatGPT, to help companies write job descriptions and help users polish profile copy or author a message to hiring teams.

It’s all just the start. Berger, the product engineering executive at LinkedIn, says he imagines users someday explaining the jobs and companies they are interested in, their preferred workplace culture, and compensation criteria to a conversant copilot, which could show them matching job opportunities and then talk them through crafting applications and preparing for interviews. To make such an AI-fueled vision reality, Berger says LinkedIn has assembled engineers for another intense effort of starting with a clean slate, like those that came before in the company’s history. This time, he says, the project is driven more by ambition than impending danger—and the site will stay up.