LinkedIn is launching its new AI-powered people search this week, after what seems like it should have been a natural offering for generative AI.
The chat comes three years after GPT launched and six months after LinkedIn launched its AI job search offering. For tech leaders, this timeline illustrates an important enterprise lesson: Deploying generative AI in real enterprise settings is difficult, especially at scale to 1.3 billion users. It is a slow, brutal process of pragmatic reform.
The following account is based on several exclusive interviews with the LinkedIn product and engineering team behind the launch.
First, here’s how the product works: The user can now type a natural language query, "Who is going to know about curing cancer?" In the LinkedIn search bar.
The old LinkedIn search, based on keywords, would get stumped. It would have been just looking for references "Cancer". If a user wanted to be sophisticated, they would have to search for separate, hard keywords "Cancer" And then "Oncology" And try to aggregate the results manually.
The new AI-powered system, however, understands this intention Searched because the LLM under the hood captures the semantic meaning. For example, it recognizes "Cancer" Conceptually associated with it "Oncology" and even less directly, "Genomics Research." As a result, it’s at the level of a much more relevant list of people, including oncology leaders and researchers, even if their profiles don’t use the exact word. "Cancer"
The system also balances this compatibility utility. Instead of only showing the world’s top oncologists (who might be an inaccessible third-degree connection), it will also weigh who you have in your immediate network—like a first-degree connection. "Very relevant" And can serve as an important bridge to this specialist.
See the video below for an example.
The argument, though, has a more important lesson for enterprise practitioners "Cookbook" LinkedIn has evolved: a replicable, multistage pipeline of Asuncion, co-design, and countless optimizations. LinkedIn had to perfect one product before trying it on another.
"Don’t try to do too much at once." writes Wenjing Zhang, LinkedIn’s VP of engineering, in a post about the product launch, and who also spoke with VentureBeat in an interview last week. She notes that earlier "Broad ambitions" To build a unified system for all LinkedIn products "Stalled progress."
Instead, LinkedIn focused on winning verticals first. Prior to that, its success was driven by AI job search—which led to job seekers born without a four-year degree. 10% are more likely to be hiredaccording to Iran Berger, VP of Product Engineering – provided the blueprint.
Now, the company is applying that blueprint to a much bigger challenge. "It’s one thing to be able to do that in tens of millions of jobs." Berger told VentureBeat. "It’s quite another to do it north of a billion members."
For enterprise AI builders, LinkedIn’s journey provides a technical playbook for just that Actually What it takes to move from a successful pilot to a billion-user scale product.
The new challenge: a 1.3 billion member graph
Berger explained that the job search product created a strong recipe on which to build new people search products.
The recipe started with one "The Golden Data Set" Only a few hundred to a thousand actual query pairs, scored in detail on 20 to 30 pages "Product Policy" Document. To measure this for training, LinkedIn used this small golden set to signal a large foundation model to generate large volumes. artificial Training data. This synthetic data was used for training 7 billion parameter "Product Policy" A high-fidelity judge of model-fit that was too slow for direct production but perfect for teaching small models.
However, the team hit a wall early on. For six to nine months, they struggled to train a single model that could balance strict policy adherence (compliance) against user engagement signals. "Ah the moment" That’s when they realized they needed to break the problem. He distilled the 7B policy model into one 1.7B Teacher Model Completely focused on compatibility. They then paired it with a separate teacher model trained to predict specific member actions, such as job applications for job products, or contacting and following up to find people. This "Multi-Teacher" Ensemble produced smooth probability scores that the final student model learned to simulate via KL divergence loss.
The resulting architecture works as a two-stage pipeline. First, a big one 8b parameter model Handles extensive retrieval, casting a wide net to draw candidates from the graph. Then, the highly refined student model takes over for fine grain classification. While the job search product successfully deployed a 0.6b (600 million) Parametric student, new people search products require even more aggressive compression. As Zhang notes, the team wiped their new student model down to 440m 220m parametersachieving the speed required for 1.3 billion users with less than 1% sync loss.
But applying it to people’s quests broke the old architecture. The new problem was not the only one Classification But also recovery.
“A billion records," There is one, Berger said "Different animals"
The team’s previous recovery stack was built on CPUs. To handle the new scale and latency requirements of one "snappy" Search experience, the team had to move their indexing GPU based infrastructure. This was a fundamental architectural change that the job search product did not require.
Organizationally, LinkedIn benefited in a number of ways. For a time, LinkedIn had two separate teams b (b ( Job search and people search b (b ( Trying to solve the problem in parallel. But once the job search team made progress using a policy-driven approach, Berger and his leadership team intervened. He took Jeet Architects in search of employment b (b ( Product Lead Rohan Rajeev and Engineering Lead Wenjing Zhang b (b ( Transplanting their ‘cookbook’ directly to the new domain.
Distilling for 10x throughput gain
Along with solving the retrieval problem, the team faced the challenge of classification and performance. This is where the cookbook was adapted with new, aggressive optimization techniques.
Zhang’s technical post (I will insert the link once it goes live) Provides specific details that our audience of AI engineers will appreciate. One of the more important optimizations was the input size.
To feed the model, the team trained Another one LLM with Reinforcement Learning (RL) for the same purpose: to abstract the input context. This "Summary" The model was able to reduce the input size of the model 20 times with minimal information loss.
The combined result of the 220M-parameter model and the 20x input reduction? a 10x increase by rankingallowing the team to effectively roll out the model to their massive user base.
Pragmatism over hype: building tools, not agents
During our talk, Berger was adamant about something else that might grab people’s attention: the real value in businesses today is in perfecting the recommendation system, not chasing it. "Agentic hype." He also declined to discuss the specific models the company used for the search, suggesting it didn’t matter. The company chooses the models based on what it finds most efficient for the job.
The new AI-powered people search is a manifestation of Berger’s philosophy that it’s better to improve recommender systems first. The architecture includes a new "Intelligent query routing layer," As Berger explained, he himself is an LL.M. This router practically decides whether a user’s query is -like "Trust expert" – Should go for the new semantic, natural language stack or the old, reliable lexical search.
This entire, complex system is designed to be a "Tool" That a The future The agent will not use the agent itself.
"Agentic products are only as good as the people they use to accomplish tasks," Berger said. "You can have the best reasoning model in the world, and if you’re trying to use an agent to find people but the people search engine isn’t great, you’re not going to be able to deliver."
Now that people search is available, Berger suggests that one day the company will offer to use agents. But he did not provide details about the timing. He also said that the recipe used for finding jobs and people will spread to the company’s other products.
For enterprises to build their AI roadmap, LinkedIn’s playbook is clear:
Be practical: Do not try to boil the sea. Win a vertical, even if it takes 18 months.
Codify "Cookbook": Turn this win into an iterative process (policy documents, alignment pipelines, co-design).
Improve without limits: The actual 10x gains come After An RL-trained summarizer such as initial models, pruning, smoothing, and generative corrections.
LinkedIn’s journey shows that for real-world enterprise AI, the emphasis on specific models or cool agentic systems should take a back seat. Sustained, strategic advantage comes from mastery The pipeline -An ‘AI-native’ cookbook of collaborative design, automation and ruthless optimization.
.