Tomer Cohen on LinkedIn Product Decisions and the AI Shift

LinkedIn chief product officer Tomer Cohen tells Harry Stebbings that he has always loved building: identifying problems, designing, applying technology and learning from feedback. He worked on semiconductor design and embedded systems before founding a consumer company. He joined LinkedIn in 2012 and became its CPO in 2020.

Cohen recalls hearing Reid Hoffman discuss professional communities and economic opportunity in 2008. Hoffman later became a mentor. Cohen says a conversation about rebuilding LinkedIn for mobile led to an invitation to join the company; he values Hoffman’s ability to offer a new frame for a difficult problem.

Asked whether product is art or science, Cohen resists assigning percentages. He sees design, data, experimentation and business skills as a learned craft, while vision, imagination, intuition and judgment help distinguish exceptional work. Those capabilities must be practiced by building, he argues.

Cohen enjoys the visionary, craft and operational dimensions of product work, from imagining a change to refining a launch through feedback. He says early releases are often incomplete and require iteration toward product-market fit. He values people who love the act of building rather than merely making decisions.

A roadmap should draw on what teams learn from customers, including unmet needs customers may not articulate directly. Cohen uses the jobs-to-be-done framework to distinguish what people seek from the audience segment to which they belong. He says functional, emotional and social needs all matter.

For example, buying enterprise software can be emotionally stressful: the person responsible must secure agreement and live with the consequences of an expensive decision. A professional creator may care about reputation and opportunities as much as likes. Cohen cites an internal creator-program observation; it is not evidence that all creators obtain economic outcomes.

He says the founder’s original insight remains valuable, but the founder’s operating role can change as a company matures. In an early startup he would be concerned if a founder were not deeply involved in product; in a mature company the founder may guide rather than run daily operations. His examples are his view of company stages, not a universal rule.

Stebbings asks whether LinkedIn’s interface feels dated. Cohen acknowledges room to simplify and rethink it, including through AI. He cites company growth in revenue, members and engagement over several years; those are a LinkedIn executive’s 2023 claims, not independently verified metrics here.

His operating maxim is that the team may be wrong but should not be confused: clarity about the problem, principles and execution makes a hypothesis testable. He describes feed and messaging as ways to help members accomplish professional tasks rather than goals in themselves.

Cohen sees both social connections and topic-based recommendations as useful. A network can reflect colleagues and people willing to help one another; following can expose someone to expertise outside that network. He acknowledges LinkedIn had not always clearly distinguished connecting from following.

He says LinkedIn was increasingly suggesting follows where people did not appear to know each other, while emphasizing stronger mutual relationships in the network. He cites follow-growth figures as company metrics. Stebbings questions whether large connection lists still convey a meaningful relationship.

LinkedIn Stories is Cohen’s example of being clear but wrong. The team hypothesized that ephemeral posts would reduce professional anxiety and encourage sharing, but results were mediocre. Follow-up conversations suggested some creators wanted their work to last and reinforce professional identity, rather than disappear.

He argues that being first to product-market fit matters more than being first to launch. The Stories test, he says, was about whether the format served professional expression, not simply beating Snapchat. He recommends defining success before launch and looking for adoption and return use rather than retrospectively fitting metrics to a story.

Some bets begin with conviction rather than strong evidence. Cohen uses skills-based matching as an example: the team believed demonstrating skills could improve labor-market matches and then looked for early validation before investing further. The claimed potential is a hypothesis and product strategy, not proof that the skills gap is solved.

Cohen calls product reviews “product jams” to emphasize feedback instead of approval. A presenting team brings local expertise and a proposed solution; cross-functional leaders contribute different perspectives. He prefers in-person sessions for creative discussion, while acknowledging remote sessions can work.

A jam begins with a nuanced problem, the job to be done and principles that specify tradeoffs, then spends substantial time on the product demonstration. For spam, growth and trust can conflict: reducing friction may admit bad activity, while adding it may obstruct legitimate activity. He wants teams to state their approach explicitly.

He gives advertiser return on investment as an example of a product principle: optimize for customer value rather than immediate LinkedIn revenue, on the belief that satisfied customers will return. That is LinkedIn’s stated product philosophy, not an independent proof of advertiser results.

At the end of a jam, Cohen summarizes one to three areas for progress. The presenting team sends a brief-back with feedback, actions and estimated timing. He says accountability requires clear roles, including who recommends, approves and ultimately decides when teams disagree.

Cohen describes transforming LinkedIn’s feed from a surface used by internal teams to promote features into a place for professional conversation. He says the change was controversial because other products relied on the feed for discovery and revenue. The team used evidence over time to support the new direction.

As conversation and virality grew, he says, clickbait, spam and shallow engagement also appeared. LinkedIn chose to restrain growth and redesign for trust and quality, in his account. As a product executive he has an interest in that characterization; the interview does not independently measure the tradeoff.

Cohen declines to give the feed a numeric score, saying it remains a work in progress. His goal is for members to find insights and relationships useful to their work, while creators gain opportunities. He describes such outcomes for millions of members, a company claim rather than a documented effect in the interview.

On the balance between revenue and innovation, he says LinkedIn connects annual goals and longer-term product bets to a broader strategy, while revising plans as conditions change. Stebbings worries that a broad mission can obscure a specific user need. Cohen agrees teams need to specify the particular job and audience before expanding to the next one.

Cohen sees AI as a major product shift and argues product leaders must understand the technology instead of treating it as a separate technical department. He recalls combining teams across algorithms and interface, and personally reviewing training data to improve quality. These are his leadership experiences and priorities.

He compares the mindset change to moving from desktop to mobile. With generative AI, outputs are less predictable than a fully specified interface, so a team sets objectives, data and constraints while continuing to evaluate behavior. His analogy does not mean a model reliably learns or improves itself without oversight.

Stebbings asks whether enterprises will prefer bundled AI products or specialized providers. Cohen expects a bundling wave in which a model coordinates multiple tools and automates parts of workflows. His dinner-party and enterprise examples are future-oriented possibilities, not proof that these systems can reliably complete every task.

He expects substantial productivity gains but says a five-year forecast is difficult at the speed of change. When asked whether startups or incumbents will win, he points to incumbents’ capital, computing, distribution and data, while arguing startups can rethink old problems. These are predictions made in 2023, including by an executive of Microsoft-owned LinkedIn.

Cohen says pre-trained models provide a broad shared baseline, potentially reducing some historical data advantages, while genuinely proprietary data can still help a company build specialized products. Stebbings pushes back that valuable private data remain scarce. Cohen agrees that unique information can still confer an edge.

He expects teams to need new skills in working with models, data and prompts. His example asks a model to explain quantum physics to a child in a playful style; he observes that a more detailed prompt changes the answer. The example illustrates interaction design, not a validation of the scientific explanation.

Stebbings asks whether AI systems that answer questions directly may take traffic from publishers whose content they use. Cohen says LinkedIn exposes some information for discovery but has private elements, and notes that citation links can lead users back to sources. He acknowledges the future of information discovery and publisher business models remained unresolved.

On safety, Cohen says powerful technology can be useful or harmful and emphasizes transparency, inclusivity and privacy in product development. Asked about political correctness, he says information-seeking products should aim for accuracy and objectivity, while product context affects behavior. Those are his company’s stated principles, not proof that every model is unbiased or safe.

Looking beyond present models, he imagines AI helping generate new scientific hypotheses and discusses possible changes to human–machine interaction. Examples involving medicine and other fields are speculative, not established discoveries or clinical capabilities. He acknowledges that defining AGI is difficult.

For his children, Cohen emphasizes learning how to learn and a growth mindset as skills change. He cites figures about changes in job skills and predicts further acceleration, but the interview does not supply the methodology needed to assess those numbers. His educational recommendation is a personal view.

Asked whether there will be one model or many, he expects foundational models to support numerous applications and anticipates more foundational models over time. This is a 2023 forecast, not a settled market structure. He describes a continuing cycle of bundling and unbundling.

In hiring product managers, Cohen asks about a complex problem and how the candidate handled it, then about failure and what they learned. He wants clarity, depth and an evolving mindset. He considers LinkedIn’s mobile transition and feed among meaningful product releases.

He recalls an instant-articles effort that improved loading speed but struggled with publishers’ different needs. He says the hypothesis was sound while execution and timing were not. The account is his reflection on a product that did not succeed, rather than a comprehensive assessment of publishers’ positions.

His advice to new product leaders is to understand the whole LinkedIn ecosystem instead of optimizing only a single feature. He cites the platform’s large membership and company base as 2023 company figures and argues cross-product work can create distinctive value.

Cohen describes his own intermittent-fasting routine and says he is not a doctor and is speaking from personal experience. His claims about sharper thinking, health span and longevity are not established by this interview and should not be generalized as medical advice.

He praises Microsoft’s OpenAI integration, explicitly acknowledging his perspective is subjective and that he has been involved. Asked whether Microsoft controls OpenAI, he describes Microsoft as a partial owner and OpenAI as independent. Those statements reflect his 2023 account and do not resolve the full governance relationship; the conversation closes with thanks.