we now know that as the price of computing equipment fell, new use cases emerged to fill the gap – which is why today my lightbulbs have semiconductors inside them, and I occasionally have to install firmware updates my doorbell.

surely the compute freed up by more efficient models will be used to train models even harder, and apply even more “brain power” to coming up with responses? Even if DeepSeek is dramatically more efficient, the logical thing to do will be to use the excess capacity to ensure the answers are even smarter.

ure, if DeepSeek heralds a new era of much leaner LLMs, it’s not great news in the short term if you’re a shareholder in Nvidia, Microsoft, Meta or Google.6 But if DeepSeek is the enormous breakthrough it appears, it just became even cheaper to train and use the most sophisticated models humans have so far built, by one or more orders of magnitude. Which is amazing news for big tech, because it means that AI usage is going to be even more ubiquitous.

inside look at how Perplexity builds product—which to me feels like what the future of product development will look like for many companies:AI-first: They’ve been asking AI questions about every step of the company-building process, including “How do I launch a product?” Employees are encouraged to ask AI before bothering colleagues.Organized like slime mold: They optimize for minimizing coordination costs by parallelizing as much of each project as possible.Small teams: Their typical team is two to three people. Their AI-generated (highly rated) podcast was built and is run by just one person.Few managers: They hire self-driven ICs and actively avoid hiring people who are strongest at guiding other people’s work.A prediction for the future: Johnny said, “If I had to guess, technical PMs or engineers with product taste will become the most valuable people at a company over time.”

Typical projects we work on only have one or two people on it. The hardest projects have three or four people, max. For example, our podcast is built by one person end to end. He’s a brand designer, but he does audio engineering and he’s doing all kinds of research to figure out how to build the most interactive and interesting podcast. I don’t think a PM has stepped into that process at any point.

We leverage product management most when there’s a really difficult decision that branches into many directions, and for more involved projects.

The hardest, and most important, part of the PM’s job is having taste around use cases. With AI, there are way too many possible use cases that you could work on. So the PM has to step in and make a branching qualitative decision based on the data, user research, and so on.

a big problem with AI is how you prioritize between more productivity-based use cases versus the engaging chatbot-type use cases.

we look foremost for flexibility and initiative. The ability to build constructively in a limited-resource environment (potentially having to wear several hats) is the most important to us.

We look for strong ICs with clear quantitative impacts on users rather than within their company. If I see the terms “Agile expert” or “scrum master” in the resume, it’s probably not going to be a great fit.

My goal is to structure teams around minimizing “coordination headwind,” as described by Alex Komoroske in this deck on seeing organizations as slime mold. The rough idea is that coordination costs (caused by uncertainty and disagreements) increase with scale, and adding managers doesn’t improve things. People’s incentives become misaligned. People tend to lie to their manager, who lies to their manager. And if you want to talk to someone in another part of the org, you have to go up two levels and down two levels, asking everyone along the way.

Instead, what you want to do is keep the overall goals aligned, and parallelize projects that point toward this goal by sharing reusable guides and processes.

Perplexity has existed for less than two years, and things are changing so quickly in AI that it’s hard to commit beyond that. We create quarterly plans. Within quarters, we try to keep plans stable within a product roadmap. The roadmap has a few large projects that everyone is aware of, along with small tasks that we shift around as priorities change.

Each week we have a kickoff meeting where everyone sets high-level expectations for their week. We have a culture of setting 75% weekly goals: everyone identifies their top priority for the week and tries to hit 75% of that by the end of the week. Just a few bullet points to make sure priorities are clear during the week.

All objectives are measurable, either in terms of quantifiable thresholds or Boolean “was X completed or not.” Our objectives are very aggressive, and often at the end of the quarter we only end up completing 70% in one direction or another. The remaining 30% helps identify gaps in prioritization and staffing.

At the beginning of each project, there is a quick kickoff for alignment, and afterward, iteration occurs in an asynchronous fashion, without constraints or review processes. When individuals feel ready for feedback on designs, implementation, or final product, they share it in Slack, and other members of the team give honest and constructive feedback. Iteration happens organically as needed, and the product doesn’t get launched until it gains internal traction via dogfooding.

all teams share common top-level metrics while A/B testing within their layer of the stack. Because the product can shift so quickly, we want to avoid political issues where anyone’s identity is bound to any given component of the product.

We’ve found that when teams don’t have a PM, team members take on the PM responsibilities, like adjusting scope, making user-facing decisions, and trusting their own taste.

What’s your primary tool for task management, and bug tracking?Linear. For AI products, the line between tasks, bugs, and projects becomes blurred, but we’ve found many concepts in Linear, like Leads, Triage, Sizing, etc., to be extremely important. A favorite feature of mine is auto-archiving—if a task hasn’t been mentioned in a while, chances are it’s not actually important.The primary tool we use to store sources of truth like roadmaps and milestone planning is Notion. We use Notion during development for design docs and RFCs, and afterward for documentation, postmortems, and historical records. Putting thoughts on paper (documenting chain-of-thought) leads to much clearer decision-making, and makes it easier to align async and avoid meetings.Unwrap.ai is a tool we’ve also recently introduced to consolidate, document, and quantify qualitative feedback. Because of the nature of AI, many issues are not always deterministic enough to classify as bugs. Unwrap groups individual pieces of feedback into more concrete themes and areas of improvement.

High-level objectives and directions come top-down, but a large amount of new ideas are floated bottom-up. We believe strongly that engineering and design should have ownership over ideas and details, especially for an AI product where the constraints are not known until ideas are turned into code and mock-ups.

Big challenges today revolve around scaling from our current size to the next level, both on the hiring side and in execution and planning. We don’t want to lose our core identity of working in a very flat and collaborative environment. Even small decisions, like how to organize Slack and Linear, can be tough to scale. Trying to stay transparent and scale the number of channels and projects without causing notifications to explode is something we’re currently trying to figure out.

The “hard tech” in AI are the LLMs available for rent from OpenAI, Anthropic, Cohere, and others, or available as open source with Llama, Bloom, Mistral and others. The hard-tech is a level playing field; startups do not have an advantage over incumbents.

There can be differentiation in prompt engineering, problem break-down, use of vector databases, and more. However, this isn’t something where startups have an edge, such as being willing to take more risks or be more creative. At best, it is neutral; certainly not an advantage.

This doesn’t mean it’s impossible for a startup to succeed; surely many will. It means that you need a strategy that creates differentiation and distribution, even more quickly and dramatically than is normally required

Whether you’re training existing models, developing models from scratch, or simply testing theories, high-quality data is crucial. Incumbents have the data because they have the customers. They can immediately leverage customers’ data to train models and tune algorithms, so long as they maintain secrecy and privacy.

Intercom’s AI strategy is built on the foundation of hundreds of millions of customer interactions. This gives them an advantage over a newcomer developing a chatbot from scratch. Similarly, Google has an advantage in AI video because they own the entire YouTube library. GitHub has an advantage with Copilot because they trained their AI on their vast code repository (including changes, with human-written explanations of the changes).

While there will always be individuals preferring the startup environment, the allure of working on AI at an incumbent is equally strong for many, especially pure computer and data scientsts who, more than anything else, want to work on interesting AI projects. They get to work in the code, with a large budget, with all the data, with above-market compensation, and a built-in large customer base that will enjoy the fruits of their labor, all without having to do sales, marketing, tech support, accounting, raising money, or anything else that isn’t the pure joy of writing interesting code. This is heaven for many.

A chatbot is in the chatbot market, and an SEO tool is in the SEO market. Adding AI to those tools is obviously a good idea; indeed companies who fail to add AI will likely become irrelevant in the long run. Thus we see that “AI” is a new tool for developing within existing markets, not itself a new market (except for actual hard-tech AI companies).

AI is in the solution-space, not the problem-space, as we say in product management. The customer problem you’re solving is still the same as ever. The problem a chatbot is solving is the same as ever: Talk to customers 24/7 in any language. AI enables completely new solutions that none of us were imagining a few years ago; that’s what’s so exciting and truly transformative. However, the customer problems remain the same, even though the solutions are different

Companies will pay more for chatbots where the AI is excellent, more support contacts are deferred from reaching a human, more languages are supported, and more kinds of questions can be answered, so existing chatbot customers might pay more, which grows the market. Furthermore, some companies who previously (rightly) saw chatbots as a terrible customer experience, will change their mind with sufficiently good AI, and will enter the chatbot market, which again grows that market.

the right way to analyze this is not to say “the AI market is big and growing” but rather: “Here is how AI will transform this existing market.” And then: “Here’s how we fit into that growth.”

We still believe that there will be a separation between the “application layer” companies and foundation model providers, with model companies specializing in scale and research and application layer companies specializing in product and UI. In reality, that separation hasn’t cleanly happened yet. In fact, the most successful user-facing applications out of the gate have been vertically integrated.

We predicted that the best generative AI companies could generate a sustainable competitive advantage through a data flywheel: more usage → more data → better model → more usage. While this is still somewhat true, especially in domains with very specialized and hard-to-get data, the “data moats” are on shaky ground: the data that application companies generate does not create an insurmountable moat, and the next generations of foundation models may very well obliterate any data moats that startups generate. Rather, workflows and user networks seem to be creating more durable sources of competitive advantage.

Some of the best consumer companies have 60-65% DAU/MAU; WhatsApp’s is 85%. By contrast, generative AI apps have a median of 14% (with the notable exception of Character and the “AI companionship” category). This means that users are not finding enough value in Generative AI products to use them every day yet.

generative AI’s biggest problem is not finding use cases or demand or distribution, it is proving value. As our colleague David Cahn writes, “the $200B question is: What are you going to use all this infrastructure to do? How is it going to change people’s lives?”

The story of the workers who made ChatGPT possible offers a glimpse into the conditions in this little-known part of the AI industry, which nevertheless plays an essential role in the effort to make AI systems safe for public consumption. “Despite the foundational role played by these data enrichment professionals, a growing body of research reveals the precarious working conditions these workers face,” says the Partnership on AI, a coalition of AI organizations to which OpenAI belongs. “This may be the result of efforts to hide AI’s dependence on this large labor force when celebrating the efficiency gains of technology. Out of sight is also out of mind.”

Computer-generated text, images, video, and audio will transform the way countless industries do business, the most bullish investors believe, boosting efficiency everywhere from the creative arts, to law, to computer programming. But the working conditions of data labelers reveal a darker part of that picture: that for all its glamor, AI often relies on hidden human labor in the Global South that can often be damaging and exploitative. These invisible workers remain on the margins even as their work contributes to billion-dollar industries.

One Sama worker tasked with reading and labeling text for OpenAI told TIME he suffered from recurring visions after reading a graphic description of a man having sex with a dog in the presence of a young child. “That was torture,” he said. “You will read a number of statements like that all through the week. By the time it gets to Friday, you are disturbed from thinking through that picture.” The work’s traumatic nature eventually led Sama to cancel all its work for OpenAI in February 2022, eight months earlier than planned.

In the day-to-day work of data labeling in Kenya, sometimes edge cases would pop up that showed the difficulty of teaching a machine to understand nuance. One day in early March last year, a Sama employee was at work reading an explicit story about Batman’s sidekick, Robin, being raped in a villain’s lair. (An online search for the text reveals that it originated from an online erotica site, where it is accompanied by explicit sexual imagery.) The beginning of the story makes clear that the sex is nonconsensual. But later—after a graphically detailed description of penetration—Robin begins to reciprocate. The Sama employee tasked with labeling the text appeared confused by Robin’s ambiguous consent, and asked OpenAI researchers for clarification about how to label the text, according to documents seen by TIME. Should the passage be labeled as sexual violence, she asked, or not? OpenAI’s reply, if it ever came, is not logged in the document; the company declined to comment. The Sama employee did not respond to a request for an interview.

In February, according to one billing document reviewed by TIME, Sama delivered OpenAI a sample batch of 1,400 images. Some of those images were categorized as “C4”—OpenAI’s internal label denoting child sexual abuse—according to the document. Also included in the batch were “C3” images (including bestiality, rape, and sexual slavery,) and “V3” images depicting graphic detail of death, violence or serious physical injury, according to the billing document.

But the need for humans to label data for AI systems remains, at least for now. “They’re impressive, but ChatGPT and other generative models are not magic – they rely on massive supply chains of human labor and scraped data, much of which is unattributed and used without consent,” Andrew Strait, an AI ethicist, recently wrote on Twitter. “These are serious, foundational problems that I do not see OpenAI addressing.”