https://www.lennysnewsletter.com/p/how-perplexity-builds-product
The original text comes from the unaltered version of the Deep Intelligence Translation Engine.
Johnny Ho, co-founder and product lead of Perplexity, shared valuable experiences on how he organized the team inspired by slime molds and how to empower AI companies through AI.
In less than two years since its establishment, Perplexity has become a product I use multiple times daily, replacing many of my Google searches—and I am not alone. The company has fewer than 50 employees, but its user base has grown to tens of millions. They are also generating over $20 million in annual recurring revenue and competing with Google and OpenAI in the future search battle. Their recent $63 million fundraising has valued the company at over $1 billion, with investors including Nvidia, Jeff Bezos, Andrej Karpathy, Garry Tan, Dylan Field, Elad Gil, Nat Friedman, Daniel Gross, and Naval Ravikant (but unfortunately not me). Nvidia’s CEO Jensen Huang stated that he uses the product “almost every day.”
I had an in-depth conversation with Johnny Ho, the co-founder and product lead of the company, to give you insights into how Perplexity builds its products—which, in my view, may represent the trend of product development for many companies in the future:
1. AI First: They integrate AI into every aspect of the company’s construction and actively seek solutions from AI, such as “How to launch a product?” They encourage employees to ask AI before consulting colleagues.
2. Organize Like Slime Molds: They optimize and reduce coordination costs by maximizing the parallelization of each project.
3. Small Teams: Their typical team size is two to three people. Their highly praised AI-generated podcast is built and operated by just one person.
4. Streamlined Management: They focus on hiring self-driven individual contributors (ICs) and intentionally avoid hiring those who excel at guiding others’ work.
5. Predictions for the Future: Johnny said, “If I had to guess, technical project managers or engineers with product taste will become the most valuable people in companies over time.”
For more information, please visit Perplexity. They are hiring! If you want to learn about how other excellent product teams operate, don’t miss my in-depth analysis of Linear, Shopify, Figma, Notion, Duolingo, Ramp, Miro, Coda, Gong, and Snowflake.
Appendix: I am working with Perplexity to explore how product managers can use Perplexity, and we sincerely invite you to share your valuable experiences. If you are a loyal user of Perplexity, please fill out this short survey, and they will contact you for a user interview.
From left to right: Perplexity co-founder Johnny Ho , Aravind Srinivas and Denis Yarats
1. How do you utilize AI tools internally to build Perplexity?
If we relied solely on our own exploration, it might take several days to clarify things, but with the help of AI and some prompts, we can quickly take action within five minutes.
We still use this method today. For example, this week I asked Perplexity, “How to write an email inviting someone to join Perplexity Pro?”
We even tried using AI tools to build products, but found that they are still far from mature in coding. AI can help us write scripts, but it struggles to generate the stable code required for building platforms. Even today, with technology evolving rapidly, the latest AI models can only generate template code and cannot genuinely design durable new abstract concepts.
2. How many product managers does your company have?
We only have two full-time product managers on a team of 50 people.
3. I believe your success is largely due to excellent hiring strategies and strict standards. What traits do you value most in candidates (that others might overlook)?
Given our current work pace, the foremost consideration is the flexibility and initiative of candidates. The ability to work constructively in a resource-limited environment (which may require wearing multiple hats) is crucial for us.
Looking at product managers’ resumes, many emphasize their ability to assist others and reach consensus. I believe that with the rise of AI, the importance of these skills will decline. Therefore, you don’t need to focus too much on management processes or team leadership skills. We seek outstanding individual contributors who can make a clear quantitative impact on users, rather than being limited to contributions within the company. If I see terms like “Agile Expert” or “Scrum Master” on a resume, that candidate may not be suitable for our team.
Moreover, AI empowers product managers to take on more independent contributor roles, especially in data analysis and customer insights. Of course, a solid professional foundation (such as mathematics, statistics, programming basics) is still essential, but the threshold for becoming a truly “technical” product manager has significantly lowered.
In terms of personnel selection, we still value cultural fit and teamwork abilities, but the requirement for guiding others’ work has decreased, as AI can take on that part of the responsibility to some extent. This trend may change as the company scales, but at this stage, the number of products to be developed far exceeds the available manpower.
Looking ahead, I anticipate that the management hierarchy across the industry will gradually flatten. Boldly speculating, technical product managers or engineers with product thinking will become the most valuable talents in companies.
4. Is your team organized around products, user types, user journeys, outcomes, or other dimensions? Has this organizational structure changed over the years?
My goal is to build teams that minimize “coordination friction,” as Alex Komoroske explained in his presentation about viewing organizations as slime molds. The core idea is that coordination costs (arising from uncertainty and disagreements) increase with scale, and adding more managers does not effectively solve the problem, but rather leads to misalignment of goals and distortion of information. For example, employees tend to conceal the truth from superiors, while superiors filter information to higher management. Cross-departmental communication requires layers of reporting and approvals, resulting in inefficiency.
Conversely, what we should do is ensure consistency of overall goals and advance projects aimed at that goal in parallel through shared reusable guidelines and processes. Especially with the push of AI, we can leverage AI for “rubber duck debugging” to minimize coordination costs and avoid over-reliance on perfect coordination and consensus. Additionally, we maintain a real-time updated internal “expert directory” to encourage employees to directly contact relevant personnel when needed. This model is based on a high level of trust.
However, more importantly, with the assistance of AI, you do not need to frequently communicate with others. Before asking others, you can try spending a minute asking AI, which not only reduces communication costs but also provides a reasonable starting point for everyone to independently begin their work.
Spending some time reflecting on meta-tasks at the beginning of each week can clarify thoughts and avoid overreacting or making hasty decisions. Over time, our ability to estimate scale and determine priorities based on return on investment has also improved.
6. Does your company adopt some form of OKR goal management?
We are committed to implementing rigor and a data-driven approach in quarterly planning. All goals are measurable, whether through quantitative metrics or Boolean “X completed” indicators. We set challenging goals, and typically by the end of a quarter, we can only achieve about 70% of the goals. The remaining 30% helps us identify gaps in prioritization and personnel allocation. For example, if infrastructure goals are not met, it indicates that we have not invested enough in hiring infrastructure engineers.
7. How does your company conduct product/design review meetings?
After determining core goals and high-level design plans, we tend to adopt a fairly decentralized decision-making approach. Projects are driven by a single Directly Responsible Individual (DRI), and execution steps are conducted in parallel as much as possible.
The primary step of any project is to decompose it into parallel tasks to minimize coordination issues. We use Linear for task decomposition, led by me and the product managers (PMs) in the team (or anyone responsible for PM work). We strive to ensure each task is independent, allowing executors to complete tasks without obstacles. Controversial decisions may arise during execution, but these can be resolved in subsequent stages.
A brief kickoff meeting is held at the beginning of each project to ensure team goals are aligned, after which iterative work proceeds asynchronously without constraints or review processes. When individuals feel ready to receive feedback on design, implementation, or the final product, they share relevant information in Slack, and other team members will provide candid and constructive feedback. The iterative process unfolds naturally as needed, and products are only released after gaining internal approval through internal trials.
I encourage team members to work in parallel as much as possible without waiting for others to eliminate all obstacles. Ideally, design, front-end, and back-end teams should all participate in the same project simultaneously. Now that we have a business team, all four teams can work in parallel, whereas traditional workflows may require waiting for design or models to be completed.
8. How does the reporting line work?
Currently, teams are divided by function (product, research and development, design, business, etc.), with different teams responsible for various aspects of the company and tech stack. However, all teams focus on improving the core product. We set goals and translate them into common top-level metrics to comprehensively enhance user experience. For example, all teams share the same top-level metrics when conducting A/B testing, even if the testing scope is limited to their respective tech stack layers. Given the rapid pace of product iteration, we want to avoid binding personal identity to any specific component of the product to mitigate potential political issues.
Given the current size of the company, we have adopted a flattened organizational structure. In this structure, the reporting line’s role in determining priorities is less significant than the commitment to top-level goals. Our two full-time product managers (one responsible for web and one for mobile) report directly to me (the product lead). We find that when teams are not equipped with dedicated product managers, team members proactively take on product manager responsibilities, such as adjusting project scope, making user-facing decisions, and trusting their judgment.
9. You have built one of the most popular and successful products. What do you think makes your product approach so unique or core to achieving such success?
The core of our product approach is to collect user and internal feedback and distill it into a few intuitive and easy-to-use products that meet the needs of a broad customer base. We also strive to distill feedback in a way that motivates and inspires the team, setting a macro vision while allowing individuals the autonomy to decide how to best serve the original goals. We adopt a decentralized decision-making approach, passing responsibility to every member, enabling rapid iteration without cumbersome approval processes. Individuals can quickly make locally optimal decisions based on actual conditions. Any deviations are swiftly corrected afterward.
10. What are the main tools you use for task management and bug tracking?
Linear. For AI products, the boundaries between tasks, bugs, and projects become blurred, but we find many concepts in Linear, such as Leads, Triage, Sizing, etc., to be very important. One of my favorite features is auto-archiving—if a task hasn’t been mentioned for a long time, it may not actually be important.
We mainly use Notion to store roadmaps, milestone plans, and other important information sources. During development, we use Notion to write design documents and RFC; after development is complete, we use it for documentation, post-mortem analysis, and historical archiving. Recording ideas (documenting the thought chain) helps clarify decision-making and facilitates asynchronous collaboration, avoiding unnecessary meetings.
Unwrap.ai is a tool we recently launched to integrate, document, and quantify qualitative feedback. Due to the nature of AI technology, many issues often lack sufficient certainty, making it difficult to categorize them explicitly as bugs. Unwrap aggregates scattered feedback into more specific themes and improvement directions.
11. Do you think the roadmap conception mainly comes from top-down (the team is told what to build) or bottom-up (the team proposes ideas autonomously)?
High-level goals and directions are usually set top-down, but a large number of new ideas emerge from bottom-up creative surges. We firmly believe that engineering and design teams should have ownership of ideas and details, especially in AI product development, where constraints are often unknown before ideas turn into code and models. We encourage and continuously brainstorm. We have a dedicated brainstorming channel in Slack, and subsequent ideas are collected into Linear, and typically, optimized ideas directly enter the coding implementation phase without anyone’s approval.
The best example of bottom-up ideas is reflected in Perplexity‘s discovery, collection, and sharing experience. For example, as I shared earlier, our brand designer Phi is responsible for building the Discover Daily podcast and simultaneously makes decisions on scripts, ElevenLabs integration, branding, and audio engineering. For AI, predicting use cases before product iteration release is almost impossible. A year ago, we never anticipated that the Discover experience would eventually evolve into a podcast format.
12. When people observe companies like yours from the outside, everything seems perfect, as if you have mastered everything. However, what aspects have progressed poorly or encountered significant challenges?
Currently, our significant challenge is how to scale the company from its current level to the next stage, covering recruitment, execution, and planning aspects. We do not want to lose the core values of working in a flattened and collaborative environment. Even minor decisions, such as how to organize Slack and Linear, can become tricky during scaling. How to expand the number of channels and projects while maintaining transparency without causing notification overload is an issue we are currently working to resolve.
13. Does the product team or company have any interesting or unique rituals or traditions?
At Perplexity, many features and products are born from week-long (or even shorter) hackathons. It turns out that this sprint activity focused on building new features is the most exciting and memorable. Our first interactive search prototype Pro Search (formerly Copilot) was built in just a few days and refined through multiple iterations.
Thanks to Johnny! Also, thanks to Phi Hoang for the visual effects support.Wishing you a fulfilling and efficient week!
