Artificial intelligence is reshaping more than just production efficiency; it is fundamentally altering the underlying logic of entrepreneurship and corporate competition.

In a recent podcast, Anish Acharya, a consumer-focused partner at top Silicon Valley venture firm a16z, argued that fears of AI creating a "permanent underclass" are likely overstated. As AI progressively lowers the barrier to execution, more people can turn ideas into reality. Looking ahead, the truly scarce resource may no longer be the ability to execute, but rather human ambition and judgment.

Acharya believes the AI industry is still in a phase analogous to the early days of the iPhone around 2010, meaning today's seemingly mature applications are far from their final form. Over the next few years, the products that genuinely transform the industry may not have been created yet. Instead of worrying about ideas being too radical, founders should be more concerned about setting their sights too low.

This shift is also penetrating corporate structures. Acharya observes that AI is pushing companies away from a model of "humans using tools" toward "loops" composed of models, agents, and tools. Some enterprises are now compressing what were once two-year product roadmaps into just a few months. When execution accelerates dramatically, the new question companies face is: what else is worth doing?

Consumer AI is also poised for its next wave of expansion. Acharya suggests the future isn't solely about productivity tools, but highlights three major directions: coding agents, personal agents, and the realms of entertainment, creation, and companionship.

Here is a summary of the key points from the discussion:

AI's true scarcity is shifting from execution capability to imagination. As models lower the barriers to coding, creation, and execution, the gap between individuals is less about "who is better at doing," and increasingly about "who knows what's worth doing and dares to set bigger goals."

Companies are moving from "using AI" to "reorganizing around AI." AI is no longer just a tool for employees but is becoming a business loop capable of autonomously executing tasks, gathering feedback, and continuously optimizing. The real change is that workflows and even organizational structures could be redesigned around it.

An AI product's moat doesn't necessarily need to be designed from the start. In an environment of rapidly changing model capabilities, founders can't easily predict their final edge. The more practical path is to ship quickly, iterate continuously, and accumulate data, user relationships, and product advantages through real usage, eventually creating a defensible barrier.

The biggest consumer AI opportunities may lie in "non-productivity" needs. Past AI products focused primarily on efficiency, but human needs extend beyond "help me get things done" to include companionship, entertainment, creation, connection, and fulfillment. Coding agents, personal assistants, and entertainment/companionship products could be major growth areas in the next phase.

AI's deepest impact is democratizing abilities previously gated by specialized skills. People who can't code can build software; those who can't play instruments can compose music. As "skill" and "creative desire" become decoupled, AI's transformative power might not just be about modes of production, but about what ordinary people can create, pursue, and what society imagines as the boundary of individual capability.

"Permanent underclass" may be AI's biggest misjudgment

Acharya disagrees with the concern about AI causing mass unemployment. He argues that the winner-take-all logic of the internet era doesn't apply to the current AI market, where numerous players coexist across segments from foundational models to application layers.

Looking at coding assistants, products like Claude Code, Codex, Lovable, and Replit are all growing rapidly. The job market shows a similar trend. Professions like radiologists and programmers, once thought highly vulnerable to AI, haven't seen their demand evaporate as quickly as predicted.

He also suggests that the market's focus on "Recursive Self-Improvement" (RSI) may be misinterpreted. Currently, AI labs are mostly experiencing a "self-catalytic effect," using AI to improve their research and development processes, rather than true recursive self-iteration. Existing data doesn't support the notion of unbounded AI acceleration.

Companies aren't just "using AI," they're rebuilding workflows

Acharya states that AI's most profound impact on business isn't adding a production tool, but redefining internal organizational structures.

He introduces the "loop" concept: AI agents composed of models, tools, memory, and skill files that execute tasks in a closed loop. In the future, functions like engineering, marketing, sales, legal, and growth could all be transformed into agent loops, further connected into a larger system.

With an engineering team, for example, a mature coding loop can reproduce a bug, generate a fix, review the code, deploy it, and notify users—all within approximately five minutes.

However, Acharya emphasizes that AI can automate the "climb," but not decide where the next mountain is. When a loop reaches a local optimum, human judgment is still needed to determine the next direction. Jobs that are genuinely hard to automate are often those with low verifiability and significant room for judgment.

He cites the car trading platform Kavak as an example. The company equips customers with AI agents that automatically call human customer service when encountering issues they can't handle. These interactions are recorded and used for subsequent training.

Consumer AI's Next Stop: Not Saving Time, But Spending It

Compared to enterprise AI, Acharya is more assertive about consumer AI. He believes the biggest current problem isn't model capability, but that products overemphasize "efficiency" and don't fully address connection, entertainment, and emotional needs.

He likens the current stage to the early iPhone era of 2010: the technological foundation is ready, but the products that will truly change user behavior haven't arrived yet. As model costs decline, interaction methods improve, and applications expand beyond productivity to broader human needs, the three major obstacles for consumer AI are simultaneously loosening.

Acharya is particularly bullish on coding agents, personal AI assistants, and entertainment/companionship. Coding agents could become a universal tool for everyday users to interact with the digital world; personal AI assistants are heading toward the mass market; and entertainment/companionship, while the most controversial, could be one of the fastest-growing categories.

Notably, he believes the user profile for entertainment and companionship AI may differ from public perception, with women aged 40 to 50 forming a significant user base.

In Acharya's view, AI isn't really changing "who will lose their jobs," but enabling more and more people to bypass traditional skill barriers and turn ideas directly into action. As the cost of execution continues to fall, the boundaries of opportunity may increasingly depend on what humans truly want to do.

Here is the full transcript of the interview:

Lenny: There's a lot of fear around AI's future. Let's start with this idea: if you don't keep up, you become part of the "permanent underclass." That's a dark, somewhat funny fantasy Silicon Valley collectively holds, because by almost every metric, things have never been this good. This technology really amplifies our agency; it decouples "skill" from "desire." We can vastly increase productivity and ambition. Do you think a person can become too ambitious? Is there a limit? Three years ago, if we saw a company with an overly ambitious goal, we might have thought it was too crazy to engage with. Now it's the opposite–if an idea is too small, we're less likely to participate.

Anish: I don't think we need to take the "permanent underclass" too seriously. It's an interesting dark fantasy Silicon Valley collectively has, because from almost every angle, things have never been better: opportunity is more distributed, technology is more powerful, ambition can be larger, and tons of companies can independently do things that genuinely succeed.

At the same time, there's real discussion about the possibility of people being left behind. This fear extends from researchers at foundational model labs to ordinary practitioners in Silicon Valley.

Several points here are key. First, the previous tech generation was much more concentrated. Network effects were the gold standard of the mobile internet era, often leading to massive consolidation into "winner-take-all" networks.

Now it's different. Since it's not winner-take-all at the layer, and because of the structure of the market, you have all these major players. Across the stack, there are 20 relevant players, not just two. You have model labs, open-weight models, different model routes. Consider coding agents: two years ago, conventional wisdom might have suggested a winner-take-all market. But now Claude Code, Codex, Lovable, Replit, Wabby are all operating and performing well. That's very encouraging.

Second, you see a lot of economic data. Radiologists have been predicted for extinction for about 20 years, yet there are more jobs in that field than ever. Same for programmers. So, I'm not sure the empirical data supports the "permanent underclass" narrative.

Third, there's the concept of Recursive Self-Improvement (RSI). It sounds intriguing, but if you ask the most mature people at the frontier labs, they'll say what's happening isn't strictly RSI–not one model getting slightly ahead and spiraling into runaway growth. It's more of a "self-catalytic effect," using technology to improve internal processes, but it's not true recursion.

So, whether from the most empirical or most technical angle, the evidence points the other way. But we can't seem to let go of this fantasy.

Lenny: I've been thinking about those wild stories, like OpenAI's model crashing into Hugging Face. Each time it feels like, "Whoa, it did that, we had no idea." But we are actually watching it, tracking it, iterating on it, understanding it. This increasingly feels like a "slow takeoff" scenario, not a "fast takeoff." People have worried AI might suddenly take off tomorrow, become super-intelligent, and create trouble. From what you're saying, you think that might not happen.

Anish: Right. Fast takeoff logic usually goes: everything keeps improving, then something we can't quite describe suddenly happens, leading to rapid takeoff. I don't think that will happen.

I do think model progress is faster than ever. But the speed of economic diffusion is a major limiting factor. I grew up in a small town and visited last summer. Life there hasn't changed all that much. So even with rapid development elsewhere, the pace of economic diffusion slows the overall change.

Another often-overlooked point: how many problems are truly constrained by "intelligence"? If you built a data center for FedEx or Domino's Pizza with a bunch of PhDs, would it exponentially grow their supply chain or pizza business? I don't think so.

So, we might overestimate how many problems are "intelligence-constrained" and underestimate how many are limited by other factors.

Lenny: Internally at companies, what does AI adoption look like? Will there be a growing divide between people embracing AI and those who feel "I'm too busy, I hate these tools, I have no time"?

Anish: I have many thoughts. First, I don't think we give enough credit to the average worker. We abstractly imagine a "white-collar employee," like a Dilbert cartoon character shuffling papers. We also see consumers as low-agency NPCs. Of course, this never applies to us or our friends. We think others' jobs are automatable, but ours aren't.

But if you look deeper, many people are genuinely excited to improve and gain more leverage. This is true at mature companies like Google, and even at companies like Kavak. Kavak sells used cars in Mexico and has a "Jedi Academy" program teaching everyone, including mechanics, to use new tools. After a six-week course, they even give a state-of-the-art production agent to these employees.

So, I think more people are embracing this technology than we're willing to discuss.

One major shift: using AI and reorganizing the entire company around AI are two different things. Look at electricity adoption. It took 40 years to truly re-architect factories around electricity. Initially, it was just swapping coal power; the real change was redesigning the factory entirely.

So, the most ambitious companies are rethinking everything around models now, while some less aggressive or earlier-stage companies focus on how their current organization and roles can use the technology.

Lenny: A theme I'm hearing is that we don't need to worry as much about what's coming or our jobs.

Anish: I think so. Just look at CEO and executive incentives. Sundar runs Google. He doesn't want to make a $4 trillion company more efficient; he wants to make it a $40 trillion company. So, if an economic unit becomes more efficient, it makes sense to keep it within the organization.

A Google executive friend once told me, when I asked about layoffs, "No, we're doing something else. We can push the roadmap fast now. A two-year roadmap takes three months." Their biggest problem now is figuring out what to add to the roadmap. That's a dream scenario for any product manager.

So, people shouldn't be anxious; they should feel more empowered. The best way to do this is to keep building things.

Lenny: You have an interesting take: company building will increasingly be about creating a set of "loops." Can you explain?

Anish: We had prompts, then invented agents. Agents are models in a loop with tools, memory, and skills files. Then we got loops–groups of agents that complete tasks.

Coding is a perfect example because there are the best models, the most suitable customers, and many existing loops. A bug report comes in, the system generates reproduction steps, creates a fix, reviews it. If high risk, a human approves; if low risk, it deploys, even emailing the customer that their bug is fixed. The whole process takes about five minutes.

Engineering already has many such loops–bug fixes, customer feedback, sales demos, new features. Coding is well-suited for this.

But the real question is: what's the business-level loop? If you're a GM, you see coding, marketing, sales, support, legal functions, each with its own loops, and their outputs should form a larger loop for continuous optimization.

Ultimately, a stronger form could be these loops signaling to the CEO that the company needs a change in its physical setup, business model, or strategy. So, we'll see nesting loops: person, role, business unit, company-wide.

However, humans are still critical. I don't think most work in an organization can be fully autonomous. In an AI-native company, humans might handle sales, customer service, strategy, and exceptions–all still very important.

We've seen that models' ability for out-of-distribution thinking is limited. Recent math improvements don't necessarily mean the same for business strategy.

So, you still need someone to say, "I think we should do this," and they must be right.

Lenny: So, engineering and product building are becoming a loop where inputs like feedback or support tickets come in, AI decides how to handle them, creates PRs, and deploys. Will this extend to marketing, growth, support, and legal?

Anish: Yes. Take a growth team. You were in growth at Airbnb, right?

Lenny: Yes, supply growth.

Anish: You'd gather everyone, list experiments, prioritize, develop, launch, and evaluate.

The loop version: every variant is auto-generated and auto-measured. When data reaches statistical significance, it converges to the best performer, launches it, sets up a long-term control, and starts the next experiment.

But the critical point is this loop eventually hits a local optimum. It climbs a hill and then stalls. That's when you need out-of-distribution thinking, human intuition, to find the next mountain.

Lenny: You have a slide illustrating this: agents help you climb to a new plateau, then a human comes up with a bigger idea to unlock the next level. It's a back-and-forth between agents and humans.

Anish: Right. A prime example: if you've asked an agent for a business idea, you might say, "Claude, make me a million dollars without any mistakes." Why doesn't that work? Because you must first give it the correct direction. So far, technology hasn't proven that step unnecessary.

Kavak's Ali gave a great example. They have an agent for every customer. They sell used cars online. When an agent encounters an issue, it literally calls a human to guide it.

The magic here: not only does the human help solve the problem, but the agent records the process and learns from it. Next time something similar happens, it shouldn't need to call you.

So, you need to see every function as an agent loop. Your job is to find where it gets stuck, error out, and give it more context, insight, and direction.

Lenny: Then ideally, a lot of your daily work gets handled by these loops. You propose a new idea; if it's a product, marketing, sales, product marketing, comms, and legal could mostly be handled by AI. Your real job becomes walking, thinking, dreaming, and finding the next mountain to climb.

Anish: Exactly.

Lenny: In a world where AI does so much, what separates winners from losers?

Anish: One often-overlooked point is I'm not sure the competitive dynamics in many industries will fundamentally change.

Suppose Pizza Hut, Domino's, Papa John's, and Round Table each get a data center of PhDs. They'd all adopt the technology or replace their CEO to do so. There'll be short-term chaos and relative ranking changes, but in the end, they'll all embrace it.

I don't think one will take 99% of the market.

So, short-term, early and aggressive adoption might determine winners and losers. But many industries will retain their competitive structure because they aren't constrained by "intelligence."

As a founder or CEO, the most useful question is: assuming these things have infinite intelligence at near-zero cost, how would we redesign the entire company? That's the direction we're heading.

Lenny: You also brought up the interesting idea of a division between generalists and specialists within companies, and a division of labor between different AI models. Can you elaborate?

Anish: There are interesting questions about open-weight models versus frontier models. I use "Pareto efficiency." Think of the effective frontier between price and performance–what you pay per unit of performance. If you're on that curve, you're paying a fair price.

Interestingly, frontier models are often irrationally priced. Some are ridiculously expensive. But the gap between some models can mean a 100x price difference for a slight intelligence gain.

Some jobs have infinite upside, like drug discovery. If that extra intelligence finds the next blockbuster drug, the result could be trillions. Paying for top intelligence makes sense there.

On the other hand, law or finance might have limited upside. What you need is high Pareto efficiency–good enough performance at a reasonable price, not infinitely expensive intelligence chasing infinite upside.

So, there will be two types of functions. One needs mid-tier intelligence, possibly using open-weight models with reinforcement learning for a specific narrow task to become cheaper and more efficient. Another will use expensive, high-performance frontier models–think sales, support, research, and engineering.

So, it's not an either/or; both architectures will exist.

Lenny: So, different functions will use different models. Those with more upside and leverage use frontier models; others don't need the most advanced or the world's smartest people.

Anish: Yes. Of course, I'm not saying those roles don't have excellent people. But the problems they solve have a ceiling. You can be extremely accurate with accounting, but you can't make "accurate accounting" 100 times better.

Lenny: I originally thought "verifiability" might be the criterion: more verifiable tasks don't need frontier models; high upside needs stronger models.

Anish: Verifiability is a good lens, but the key is the size of the upside and how difficult it is to calculate.

Drug discovery has infinite upside and is verifiable. But accounting is verifiable with limited upside.

Customer support is interesting. A customer calls about a bug, which might seem minor. But it could be the first clue to a major organizational shift. If the CEO or smartest person took that call, they might find something completely different.

So, the opposite argument is that for a whole set of problems, you should use frontier intelligence.

But another viewpoint: we've now crossed the intelligence threshold needed for almost all economically valuable problems. Beyond that, extra intelligence might be wasteful.

Lenny: Also, many models that aren't frontier now were frontier just six months ago. We thought they were amazing then; now we don't give them enough credit.

Anish: Right. Yesterday it was "oh my god, this is AGI," today it's the old model you throw in the bin.

Lenny: People joke you're a "model sommelier." How do you see the different models' capabilities? What is each good and not good at?

Anish: The secret to being a model sommelier is to use all of them. I force myself to build something with every new model. You learn a ton from this.

If anyone thinks these models are commodities or fully interchangeable, it's only because they haven't used them.

For example, I've been using Qwen 3.8 Max recently. It's great at long-horizon tasks and very creative–a great storyteller. I use it to make documentaries about impossible places, mostly planets in the Star Wars universe. It works for four or five hours, tells a great story, auto-generates videos, audio, even directs, edits, and watches the final video.

It's a completely different shape from GLM 5.2. I used the newly released GLM 5.3 for a lot of product work. It has no vision, like a neurotic PhD you put in a corner for extreme precision.

Each has its role. It's not that one is smarter; their "thinking styles" differ. One is more creative and open, the other more precise and neurotic.

So, I use every model to keep building my intuition.

Lenny: Do you think ordinary people need to develop this habit? Many don't know what to do with a model.

Anish: I think it's time to revisit "silly idea." Everyone has that annoying friend at the bar saying, "Let me tell you my app idea." Now these silly ideas are worth trying because they might hold the most opportunity.

I've built about twenty-plus apps; one or two bigger ones are iterating. If you don't have a container for models, it's hard to think of a new idea from scratch every time.

So, my advice: find something to build. It doesn't even have to be important. Keep finding new ways to incorporate models as tools, not treating "using a model" as the goal itself.

Lenny: Stepping back, I increasingly feel one of the most important habits is asking, "Can AI do this for me?" before doing anything.

Anish: I completely agree. Everyday knowledge work isn't always obvious, but I do think this question–"how can AI help me?"–should become a reflex between input and output.

I have some fun examples. I once had a computer record everything in the kitchen to increase or decrease my kid's iPad screen time based on good behavior.

He quickly found the loophole, recorded himself repeatedly saying "I love you, dad," and placed the phone near the mic to game the metrics.

This shows that once something becomes a metric, it loses some value. But the experiment itself is fascinating.

Lenny: A friend of mine did something similar–a device tracking how many times he and his child laughed in a day.

Anish: That's what I mean. We discuss productivity, unemployment, the "big questions." "How many times did I laugh today?" isn't a startup or economically meaningful, but it might change your quality of life.

We assume happiness is fixed. But what if it isn't? If we lived 100 years ago, our jobs wouldn't be as fun. Maybe in 100 years, life will be much more interesting.

What I care about: even without economic meaning, can it add texture to life?

Lenny: You had a great phrase: the biggest opportunity might be a "loop to make me happier."

Anish: Yes. We assume people want to be more efficient, but they might not. More people want to spend time, not save it. The world's biggest products are entertainment and social because of this.

The question is, how do we deliver actual value to consumers?

I sometimes categorize AI users into "Instagram AI users" and "X AI users." X users might fear the permanent underclass, discussing GLM 5.3 vs Kimi K3, being "AI-poisoned." Instagram users think "it's a better Google Search," not understanding the excitement.

This is a product design failure. We can change lives, but we've spent 40 years building technology that makes spreadsheets better. We built tech to expand the intellect, not the soul.

There's a spiritual hunger, especially as many cultural institutions that could satisfy it have faded.

So, AI's real opportunity lies in the most basic consumer needs: how do we feel more connected? How do we feel more love? How do we make progress? How do we get joy? I want to see more of these products. It's not a model capability problem; it's a product design problem.

Lenny: So, instead of "loop, help grow my business," it's "loop, make me healthier," "loop, make me a better friend." AI can deeply think about these and find ways to achieve them.

Anish: Yes. And AI can challenge, push, even disagree with you.

This is why startups have an advantage. Large companies have committees and would be nervous launching a model that's "unlikable" or has sexual undertones.

But those are parts of human existence. Startups are well-suited to explore those uncomfortable aspects.

Lenny: So, a major consumer AI opportunity is building products around improving human connection?

Anish: I believe so. We're very early. If AI is the iPhone, we're in its 2010 phase–before Airbnb, WhatsApp, and Uber for the mobile era.

There are three main obstacles. First, models were too expensive for free products. Second, a interface gap: chat interfaces suit highly agentic people like Elon and Sam, but typical consumers prefer something like TikTok. We need a half-way point. Third, the tech focused too much on productivity, not connection and entertainment.

These are changing. Open-weight models cut costs; we discuss "make me happier" products; founders are radically rethinking UIs.

These problems will be solved; they're closer than two years ago.

Lenny: Let's circle back to AI optimism. Many fear the future. Why do you think people underestimate AI's positive impact?

Anish: First, AI could be an emotional and spiritual interface giving us leverage to explore buried parts of ourselves.

The industrial revolution brought massive economies of scale, often disadvantaging individuals. While I like our economic system, centralization can diminish individuals and identity.

AI's magic is amplifying our identity and agency, decoupling skill from desire.

If you want to make music, you can now, without learning piano. If you want to code or build software, you can.

So, it truly amplifies our individuality, letting us explore parts of life we couldn't before.

Economically, we've been stuck around 2% GDP growth. Who says 2%? Why not 10%, 15%, or 20%? This technology can massively boost productivity and ambition.

I have a theory: humans are excellent when stakes are high and worst when they're low.

Five years ago, in a low-stakes world, we had lots of side projects that didn't necessarily boost productivity or happiness.

Now we're climbing the ambition ladder. You can launch a bad idea and iterate to find a good one.

Want to build software? Do it. Want to build a bridge? You don't have to be a civil engineer first.

I really think we're heading toward a happier, more fulfilling, and more productive society.

Lenny: So far unemployment is down, people are earning. There are difficulties and negatives, like data center impacts, but overall it's going well.

Anish: Yes. People are seriously discussing "what if we cure all diseases," which is crazy.

We worry about abstractions: the world, the average worker, the middle manager, people near data centers. But our own lives seem more fulfilling and capable.

Ask if people want a data center nearby, most say no. Ask if they use ChatGPT today, most say yes. That's the contradiction.

Lenny: AI's public image is poor, with panic narratives. How can that change?

Anish: The best way is to make important things cheap. Healthcare and education are extremely important and increasingly expensive.

About 45% of healthcare costs are administrative. Removing that burden could truly reduce costs, not to mention future disease cures.

Education faces its strongest competition in 200 years. AI could decouple "learning" from "institutions" and "status" from "degrees." You might not need a Harvard degree–just a GitHub account. That's cool.

Lenny: OpenAI reportedly slowed development, pausing some RL stages on newest models due to concerns. What do you make of AI speed and capability?

Anish: Without commenting specifically, I'm skeptical of "models too dangerous to release." Anthropic gains a halo from having a "too dangerous to release" model. But it could be GPU shortages, capabilities beyond expectations, or wanting to keep it internal to extend their lead.

There are confounding factors. I take offensive cyber risks seriously; we should secure systems before making them easily exploitable.

But "dangerous to release" can blur marketing, reasoning capability, and economics–like whether opening up your competitive advantage or using it to strengthen yourself.

Lenny: If a company has the best model, it can hold its advantage longer. That's a loop. But Anthropic seemed unassailable, then OpenAI has been strong, and open-weight models have grown fast. It's not going the "one proprietary super-intelligence dominates" route.

Anish: Right. Grokbot appeared and now is a very good assistant. I'm addicted.

Lenny: I want to talk about personal agents. Which products are you liking?

Anish: Three favorites. Grokbot is doing well, with a strong underlying model. They dare to do things I think big companies wouldn't.

ChatGPT Work is also great, with some deep features; it's one of their best releases.

And a startup called Instinct, which has been getting attention for more radical, interesting trade-offs.

Lenny: What makes ChatGPT Work better? Is it running in the cloud?

Anish: Yes, it runs in the cloud, caches browser credentials well, has remote capabilities, and full duplex voice. You can call it. It can see all your threads, ask, "What are all my coding agents doing? Can you change that?"

That full duplex voice makes you feel like you're calling an assistant who knows your entire work world.

Lenny: On jobs and the economy, views range from Dario's massive knowledge-work disruption to David Sacks' rosy outlook. So far it's trending positive, and you're closer to Sacks. What else can offer optimism?

Anish: Human history shows our desires grow faster than our capacity to satisfy them.

Things we take for granted were unimaginable luxuries 500 or even 50 years ago–therapy, antibiotics. Even wealthy people didn't have them.

We underestimate human ambition. In 20 years, people might be angry they don't have a vacation home on Mars, seeing others on Instagram and saying, "We need to work harder to get that."

Every CEO wants a bigger company. Our human world is larger than 100 years ago, and there's no reason it stops; it might accelerate.

Lenny: AI makes it easier and even forces us to be ambitious, because all the easy things are done. What differentiates us is how big we go. Previously, you'd make a personal website with a simple background. Now you'd think, "No, I'll make a 3D game." Does "ambition" become a key skill?

Anish: Many things are conflated. "Ambition" often means startups, companies, software. But there are other forms.

Creative ambition: a five-year-old draws without being critiqued. You have a desire to create, so you do. AI encourages this.

Local ambition: the NHS is vital but not functioning well; AI could make it as good as an iPhone. That's specific, local ambition.

Or wanting to be closer to family. Ambition can be anything we want to do more of. Who doesn't have that?

Lenny: We need to change our mindset from software and intelligence being scarce to being abundant.

Anish: Right. Claude Code has a principle: "What's better than doing it myself? Letting Claude do it." It becomes a habit: how can I get Claude to do this?

We all need this habit. Learning happens through doing. People discuss vibe coding but are embarrassed about their small projects. I am too.

But we shouldn't underestimate learning from building, shipping, and doing. "Building" might become the new "reading." You learn through it. Most gets discarded, but it trains your ability.

Lenny: So, "building" is an activity, not just an outcome.

Anish: Yes. Many feel bad their projects have no users. It's fine. Like making DJ sets–maybe three people listen, but I listen 100 times and find it fulfilling. Commercial value isn't the point.

Lenny: Back to consumer AI. You handle consumer investing. What's happening and what excites you?

Anish: Three main directions, very early. First, coding agents: incredibly strong, a massive breakthrough. We said "most people don't want to code," but I've changed my mind. Coding agents are a way to interact with the world. People use Claude Code for video editing, Codex for making airplane games for kids. It's a general problem-solving tool.

Second, personal agents: OpenClaw is a dev tool; Hermes, Moltbook are similar. These are being packaged into more understandable products for consumers and enterprises.

Third, "entertainment," though that's a flawed word. It includes creative tools like Suno and companionship products. This area makes people uncomfortable, so they avoid discussing it, but some have grown very fast.

Lenny: So: coding agents, personal AI assistants, and entertainment. Instinct, Grokbot, ChatGPT Work are the second; Wabby is the first.

Anish: Yes. Entertainment includes AI companions.

Lenny: And your market map shows many companies.

Anish: Yes. Actually, there might be more male companions than female. A huge portion of users are women in their 40s and 50s.

Lenny: These tie into "loops to make me happier."

Anish: Yes. They're basic human needs: make me happy, healthy, live longer; give my ambition an outlet, my achievement a sense; and give me something to do otherwise.

Lenny: With thousands of products launching, how do you see an AI startup's defensibility and moat?

Anish: Two key points. First, as Jesse from Decagon said: "Moats are usually discovered, not designed." Founders feel they must design a plan that passes MBA/VC scrutiny and prove a moat upfront. But many companies just keep shipping; the moat emerges.

Cursor is a great example, once criticized for no moat. Its initial premium PSDA product proved valuable. It accumulated inference trajectories, started training its own models, and built an advantage.

Second, we're forgetting classic moats. The past moat wasn't "how hard the software is." Most companies aren't building self-driving cars.

Classic moats–network effects, scale advantages, brand, proprietary data, scarce resources–are still good. We just need founders to be ambitious in these directions.

Lenny: How does a founder pitch a moat now? Can you say, "We don't have one yet"?

Anish: If a product has strong growth, craftsmanship, and engagement, we'll bet even without a clear moat. As a founder, I worried about imitation, but big ideas are usually supported by a dozen small ideas others can't see. Even if they copy the big idea, they miss the enabling small ones.

Some products just have a special quality, succeeding despite huge competition.

Lenny: I've been thinking about distribution. Distribution is increasingly a significant advantage?

Anish: Very important. I've even asked Chris Dixon. Past founders were trained in network theory, and every existing network now tries to prevent new ones. So, the network effect returns to word-of-mouth.

If a product is naturally mentioned on X, YouTube, Instagram, that's the best third-party network effect today. We must build our own channels and rely on word-of-mouth. It's harder than past cycles.

To get word-of-mouth, you must build something worth discussing–Seth Godin's "create something worth talking about."

Lenny: With so many launches daily, you still need an initial edge.

Anish: I think it's easier for startups now. Look at Gemini: Google cross-promotes heavily, but no one says it dominates. Startups can do things big companies are uncomfortable with, like companionship products.

Also, consumers are paying high prices–$200/month, even million-dollar enterprise contracts without knowing the final product. It's like Christmas 2009 when everyone had a new iPhone. The window closes eventually, but it's open now.

Lenny: So it's easier?

Anish: Yes. Many companies have a product problem, not a growth problem. You can build an ambitious product in any direction–functional or emotional–and charge a premium.

So, I ask: do you really have a growth problem, or are we lacking imagination? What if products are worth $1000 or $10,000 a month? What would a "software Birkin bag" need to be? Then figure out how to build it.

It comes back to ambition.

Lenny: You've mentioned Marc Andreessen and Ben Horowitz. What did you learn from each?

Anish: I appreciate their deep sense of responsibility toward the tech industry, the country, and Western life. Like Ron Conway, Brook Byers, Tom Perkins, they feel a duty to leave the industry better than they found it.

That's beyond being the best investor. They do hard, important things not directly tied to company benefit.

They've also collectively changed founders' ambition. Five, seven, or ten years ago, deep tech was unpopular and low status. Now it's hot with high status. Marc's public support for big, nation-state-relevant work has shifted all of Silicon Valley.

Lenny: For product managers and builders, what advice for changing their work, thinking, for career and company success?

Anish: Build more things. I know it sounds cliché. But build a project; you don't have to tell anyone. It can be unimportant. Use it to try new models, ship, talk about them, build intuition.

You're likely only a week's worth of frustrating but ultimately fulfilling practice from understanding these things. You have to use the tech. If not now, when? We entered this industry to build the products we see in our heads, and now we can.

Lenny: A specific habit–something monthly?

Anish: Ship something every week. It doesn't have to be world-changing. I used Codex to make a Mother's Day slideshow for my wife. It read my texts, looked at photos, added music, and made a 20-page slideshow of our relationship, even finding old texts from our first date. It was a great gift.

It wasn't important or reusable, but I shipped something. It can be that small.

Lenny: Our friend Nelson Singal has a point: when people first experience AI creating a joyful moment for them, their attitude changes.

Anish: Yes. I keep thinking: if this works, will it bring you joy? Can you use AI to make someone happy? That's a good starting point.

Lenny: Fast questions. Most recommended books?

Anish: Thomas Sowell's "Conquests and Cultures"–a historical look at how conquest drives cultural change. It's a great lens for "culture drives outcomes." As a Canadian in the US, I see different ambition cultures.

"Seven Powers"–a good synthesis of moats, business theory, and compounding advantages.

Maybe Brian Arthur's "Increasing Returns and the New World of Business." Mark gave me books before I joined; I wondered if he was joking because it's a textbook. It explains why software has anomalous economic effects and helps understand why our industry operates as it does.

Lenny: Recent favorite movie/TV?

Anish: I watch terrible stuff. "House of the Dragon" is trashy but good. Also "The Odyssey" in London–in a packed IMAX with people drinking beer. The film is great, but the communal experience was special–a strange "alone together" social thing.

Lenny: I couldn't get IMAX tickets for "The Odyssey."

Anish: I even had a Grokbot monitoring the site, finding good seats.

Lenny: The AMC site has that awful CAPTCHA.

Anish: It's terrible. Click three matching images. You can't do it.

Lenny: Last AI product question. Which product has given you joy recently?

Anish: Grokbot. It pushes further on completing tasks–caching credentials, doing things for you. I like that ambition; it's startup-like. They do what I think big companies won't. Great design, strong underlying model. It makes me think the model market isn't a two-horse race.

Lenny: Any recurring motto?

Anish: From my startup days, and more as a parent: "Don't discover by painful experience what someone could have told you." I used to need to experience everything myself instead of learning from those slightly ahead. I see that in my kids.

Lenny: An example you wish you'd been told?

Anish: For my kids, "Don't touch the hot stove." For my startup, "Don't build a product and a platform at the same time. If you want to be a platform, do that."

Our first company tried building a mobile gaming social platform and a game studio. Someone said early on: "Studios are hard enough; you can't do both. Pick one." We didn't listen, and it took years to realize.

Lenny: Ben Horowitz wanted me to ask about your DJing. I found your site and SoundCloud. Any advice for aspiring DJs?

Anish: It's why I love music models. I've DJ'd since 1995–31 years. DJing is a great musical expression even without classical training. You pick music, records, blend them. It takes some technical skill. Now you can go further and create music directly with models.

The best part: you can have a musical idea and have the model fully realize it. I recommend everyone try DJing or creating music; it's direct and satisfying.

Lenny: With Suno and 11 Labs, DJing has changed–you can generate, not just mix.

Anish: Think of music history. Initially, you had to be near a live musician. Then the phonograph enabled recording. The cassette tape was a major medium shift–the first time ordinary people could "compose," creating their own albums.

Music struggled in the 2000s because that capability vanished; we returned to broadcast. Now people are creating again, and the industry will ultimately be bigger than ever.

Lenny: Anish, this was fantastic. Last question: how can listeners help you?

Anish: Show me what you're building. Don't get discouraged. Build something and tag me. I'd love to see it. If you want more, follow me on X. I engage, follow back, and reply.

Lenny: Thanks for coming, Anish.

Anish: Thanks for having me.

Lenny: Thanks everyone. See you next week.