Scrap Your Roadmap: A First Principles Approach to AI

A conversation with LinkedIn CPO Tomer Cohen

Hello and Happy Holidays!

It’s been a while since my last post. I’ve been heads down at Glean building a Work AI platform with Agentic reasoning. As a part of this, we launched Glean’s new prompt library - it’s been fun to see customers building and sharing their favorite prompts to help their co-workers get more value out of the AI platform.

Building these AI powered products has been exciting and also has its set of unique challenges, which is why I was eager to chat with LinkedIn’s Chief Product Officer, Tomer Cohen. Tomer is an early adopter of AI and he’s credited with dramatically improving LinkedIn’s newsfeed by integrating machine learning into the feed.

Tomer reminds me a lot of Intercom’s Fergal Reid, who spoke to me last year about what it means to take an AI-first approach to building. Like Fergal, Tomer saw that LLM technology would be a game changer. From its earliest innings, Tomer encouraged his team to embrace AI, even though by doing so, they would need to scrap their existing roadmaps. This is not easy to do! I appreciate Tomer’s candor and all the improvements he’s made to LinkedIn over the years.

Here’s our conversation:

You played a leading role in transforming LinkedIn’s newsfeed into the resource it is today. How did the newsfeed evolve, and how did you know when you were on the right track in terms of developing it?

When I started leading the LinkedIn newsfeed team in 2014, it was mostly used as a springboard for products, so reorienting it around knowledge sharing was a significant change.

LinkedIn’s primary mission is to create economic opportunity by connecting professionals to work and to each other. Before the newsfeed, these professional connections were more transactional: if you needed to fill a role, you’d reach out to a candidate; if you needed a partner for a business venture, you’d message someone who looked like they had the right experience.

At some point we zoomed out and saw that we had about ten billion years of cumulative expertise at our fingertips. What if this expertise could be used to help other professionals by introducing knowledge sharing? The big question we wanted to solve was: How do we help people connect to conversations that matter to them professionally?

How did you initially start that flywheel of change?

At the start, it was unsuccessful. In building out the newsfeed with AI, we quickly came to understand that while there were a lot of important elements at play, the one that mattered most was the ranking algorithm. How do I get to know what a user is interested in? How do I show that content to them first and foremost when they log in? If I can do that, I’m on a path to a great feed experience.

At the beginning, we were trying to do this at the scale of LinkedIn. But this was very hard because every other team had their own traffic needs to fulfill. So we were running into walls left and right as we attempted to make these big changes.

To solve for this, we carved out two million members and focused on them exclusively. We were trying to show that, for these two million members, we could dramatically change how they used LinkedIn. It didn’t happen overnight, but we did significantly change the engagement with the feed.

We started off by focusing on conversations and expertise, and we showed content these users liked based on the parameters built specifically for them. Within several weeks of introducing these changes, we saw their behavior begin to change. For these users, LinkedIn was becoming an organic place to have a conversation and to discover new opportunities. This feeling of serendipity is still a big part of the newsfeed today. You log on with a specific goal, but then you learn about something valuable that ends up expanding your career.

Did you have any kind of North Star metric as you were building out the feed?

When we first started, we spent a lot of time iterating on the objective function of AI. This is probably the most important part of being an AI product leader, because if you’re off, then you’re directing an amazing machine toward the wrong goal. I believe the best objective functions involve “Thinking Fast and Slow,” where you’re thinking “fast” about what it means to have a great experience but also thinking very slowly about the parameters it takes to create that great experience.

Over time, these parameters changed. At first, we wanted to have a high click-through rate. Then we shifted our parameters to engagement which involved viral actions, conversation, comments, and reshares. Then a lot of spammers joined because they realized they could use the platform to build virality, and this caused us to recalibrate again. From here, we shifted to downstream conversations and people interacting with their own network.

What we ultimately discovered is that objective function is a multi-factor process that erases negativity while enhancing positivity. It’s a continuous iteration. Feed algorithms are always iterating by nature because you’re always trying to get better at driving a great experience for your users. Change is key here, because the same things that would be a good fit for me in my feed a few years ago would not be the best fit for me today.

When you first started working with AI in the newsfeed, it was a bit of a contrarian bet. There were no LLMs, and no ChatGPT. Did you have any kind of “aha! moment” where you realized that the technology was really taking off?

When I took the CPO role at Linkedin I wanted to make sure that we were an AI-first product and an AI-first organization, which meant that everybody got trained in AI. They thought about AI. They thought about data. They thought about objective functions. Today, not only do they not shy away from the “AI black box” but they don't think of AI in terms of being a “black box” filled with spells and charms. They understand the box and they understand how to make sure it operates in a certain way.

We always had AI functions built into LinkedIn mostly through matchmaking job seekers and employers. But in the fall of 2022, I saw that everything was going to change. It was shocking, too, because these were big changes that weren’t somewhere off on the horizon.They were happening right now.

So I got my team together and I said, “Scratch your roadmaps. Start rethinking what we want to build.” We still had the same objectives that we were trying to solve, but now we were thinking differently about how to accomplish those objectives.

It was a tough transition. Everyone had an established roadmap, and while they were excited about their crazy CPO who just saw some remarkable new technology, they still wanted to cling to their roadmaps because it took them a long time to build those roadmaps and they didn’t want to miss their numbers. So it took a strong, top down leadership approach. I had to imprint on them that I wanted to scratch our roadmaps and go back to the drawing board.

I encouraged them to return to the very beginning by literally writing down: What is the job to be done? How can we do that job better now?

What are some examples of the changes you made to your product roadmap?

There are so many! One that’s top of mind is how we assisted recruiters in reaching out to passive candidates. It’s really time-consuming to reach out to candidates one by one and craft an email with all the right information. LLMs help out here by populating a lot of that basic information so recruiters can focus on finding the right candidates. The new tool had a big impact: we actually saw the acceptance rate of these messages go up by 40%. In the past we were celebrating 2% engagement lifts, so a 40% increase was huge.

Did you have to change your product development process because you were dealing with a technology that's non-deterministic?

This was a mindset shift I tried to instill in the team even before LLMs came on the scene. In becoming an AI-first organization, you have to start moving away from deterministic products into products that are probabilistic. That takes time to learn, and you want to make sure you’re giving the algorithm room to learn. So you rethink how you’re building. You start asking, “How can I get high quality data into this algorithm to help it learn?”

We had access to LLMs pretty early on and we were iterating a ton on prompt engineering, toying with how to write the perfect prompt. But there was still no best practices playbook for the ultimate prompt. It almost felt like being a cognitive scientist attempting to re-engineer someone’s brain.

Throughout this process we wanted to make sure we were keeping the product experience high quality. We’re dealing with people’s careers and life opportunities, so we had to make sure we weren’t sacrificing quality in the midst of re-engineering. Internally, we call people’s positive response our “thumbs up rate” and overall that rate was really high, around 95%.

What I'm finding in organizations is that people still aren’t quite sure how to leverage AI tools. Are there any tools you’re using that increase your productivity?

Glean is actually a good example here, because I’ve found it to be a really powerful way to access knowledge. I could see a world where there’s an agent that reflects the product organization of LinkedIn and you can learn about roadmaps and activities. These are products that give you a big picture view of the organization. For example, I could ask, “What are the three most challenging customer conversations that happened for the sales team this week?” or “What are some of the best opportunities that happened companywide this week?” and get that data as a report. That used to be a process that was done by multiple people, so it’s remarkable that we now have this technology. All the information is right there; it's in Gong, it's in your CRM, it's in your email. Just being able to pull all that information together is extremely powerful. So for me, that's a big change in the product development process. But I still feel like we're very much in the early innings of how this technology is going to develop.

As you said earlier, LinkedIn’s fundamental mission is to create economic opportunities. What do you think about the future of AI and work in terms of changing economic opportunities? What do you think about the role LinkedIn plays in helping people navigate this transition?

I’m very much a techno-optimist so I see AI as an enabler in terms of connecting people, empowering them, and uplifting their capabilities. At the same time, there’s a dual effect when it comes to the workforce: this technology will create new opportunities while also requiring people to adapt and acquire new skills. Something we say internally at LinkedIn is that whether or not you're changing jobs, your job is changing.

Using LinkedIn’s own forward projection, we estimate that 70% of the skills required to do a given job will change dramatically by 2030. This is something we need to get people ready for. Our hope is to give them access to the tools, programs, and communities they need to learn these new skills. The upside here is that a lot of skills are transferable. As long as you have the aptitude, you can learn.

Are there any ways in which LinkedIn is thinking about helping people get over their fears about these big changes that are coming for their jobs?

We’ve introduced a learning coach on LinkedIn where you can take courses and interact with the course. You can even roleplay situations at work and have an AI talk back to you. This is especially effective when it comes to job seeking, which is often a really lonely, emotional experience. Sometimes I do sessions with job seekers to see how they’ll perform in a given role, and it’s not unusual for people to cry during those sessions. There’s a lot of insecurities that come up when people are looking for work. They face a lot of rejection. It’s not easy.

One thing we sought to do was provide a partner for that experience. A lot of people get stuck as they’re seeking a job.They don't understand if their skills translate to what the job is looking for, so this partner can provide support and help them assess whether or not they’re a good fit. And then we have the learning tools that can help level them up and provide some pretty magical transitions for people in their professional lives.

The best part about all of these changes is that the rate of learning is accelerating. The goal here is to get people prepared so they feel they can learn and adapt and evolve. Once they get to that place, we will have tools to help them out.