Why Companies Are Becoming a Series of Loops | Anish Acharya on Lenny’s Podcast

1. Key Themes

Companies Are Becoming Cascading Systems of "Loops"

Acharya's central framework: work is evolving from prompts → agents → loops (sets of agents completing recurring tasks) → cascading loops across entire business functions. "We're going to see this sort of cascading set of everything from a loop per person to loops that can run large parts of the company." [00:00:45] He describes engineering as the model case: "There are many loops like that in engineering, everything from bug fixes to customer feedback to sales demos to new feature development." [00:12:05] He expects this to spread to growth, sales, legal, and support functions, with humans remaining essential for the parts loops can't do: "Humans are a critical ingredient... The loop will help you climb to the local maxima, but then it plateaus. You need human intuition." [00:00:45]

The "Permanent Underclass" Fear Is Overblown and Empirically Unsupported

Acharya directly challenges Silicon Valley's dominant anxiety narrative. "Things have never been better, really by almost every measure... And yet, there's this sort of discussion of permanent underclass, being outside of the light cone." [00:03:49] He cites radiologists as proof: "You've heard all the kind of economic data, everything from radiologists who are supposed to be cooked every year for, I think, about 20 years now. And of course, job postings are higher than they've ever been." [00:05:08]

Most Problems Aren't "Intelligence Bound"

A key reframe: intelligence may not be the bottleneck people assume. "How many problems are truly intelligence bound? Like if you had a data center of PhDs working at FedEx or Domino's Pizza, are they going to be exponentially dominating supply chain and pizzas? I don't think so." [00:07:13] This implies competitive dynamics in many industries won't radically change even with universal AI adoption: "I don't know that one of them is going to have 99% of the market... a lot of industries... will maintain their current competitive dynamics because they're not intelligence bound." [00:20:39]

Consumer AI's Real Opportunity Is Emotional, Not Productivity

Acharya argues the biggest untapped opportunity in consumer AI is emotional fulfillment, not efficiency. "We believe that people want to be more productive, but they don't. I think more people want to spend time than save time." [00:01:07] "We spent 40 years building a technology that enables better spreadsheets... but nothing to extend our soul." [00:32:17] The goal: "loop, make me happier" — products addressing "how do we feel more connected, more loved? Do we make progress? How do we have fun?" [00:01:07]

Ambition Itself Is Becoming the Scarce Resource and Differentiator

As AI commoditizes execution, ambition (not skill) becomes the gating factor for success. "In the old days, three years ago, we would see a company and if what they were trying to do was too ambitious, we would not engage. Today, we're almost seeing the opposite problem. An idea that's too small is not something that we want to engage with." [00:00:09] "This is a technology with which not only can we dramatically drive productivity, we can dramatically drive ambition." [00:37:26]

Moats Are Discovered, Not Designed

Traditional strategic-moat thinking is being replaced by an emergent, ship-first philosophy. "Moats are most often discovered, not designed... For that team, they just started shipping and it developed over time." [00:53:06] Cursor is the case study: "They were criticized a lot for not having a moat, but it turned out that initially being a high-end PSDAU product was really good. And over time, they captured all the reasoning traces. They trained their own models." [00:53:34]

Distribution Has Reverted to Grassroots Word-of-Mouth

The mobile-era distribution playbook (app stores, growth hacking, network-effect platforms) no longer applies the same way. "Every network that exists today is hyper trained to ensure no one else builds a network on their network. So I actually think that the sort of network effect has gone back to this grassroots, like true word of mouth." [00:58:04] But he reframes this as a product issue, not a growth issue: "Nobody has a growth problem these days. They have a product problem." [00:59:25]

Model Specialization: Frontier vs. Open-Weight Split by Job Function

Acharya predicts organizations will bifurcate model usage based on the "upside boundedness" of a task. "Frontier models are actually irrationally priced... for one IQ, conceptually one IQ of extra intelligence, you're paying 100x more than Opus... but there's a lot of jobs in which you have unbounded upside, like drug discovery." [00:21:51] Conversely: "There's perhaps only so much upside to be had in legal or finance... for those jobs, you actually do want to be very Pareto efficient." [00:22:20]

Slow Economic Diffusion Will Blunt Fears of Sudden Disruption

Even fast model progress won't translate to fast societal change because diffusion is inherently slow. "I grew up in a small town. I went back home last summer. Like people's lives haven't changed that much. So if nothing else, the sort of slow rate of economic diffusion will catch it." [00:06:44] He draws the electricity analogy: "It took 40 years for us to get from the inception of electricity to reorganizing factories." [00:08:47]

Expensive, Premium Consumer Software Is a New Frontier

Rather than assuming consumer products must be free/ad-supported, Acharya argues high-price consumer software is under-explored. "Let's think about consumer products that are extraordinarily expensive... a really useful product exercise is what is the Birkin bag $10,000 a month, $1,000 a month version of our product." [01:04:47]

2. Contrarian Perspectives

AI Labs' "Safety Pauses" May Be Partly Strategic Theater

Acharya casts doubt on the narrative that frontier labs are slowing releases purely for safety reasons. "I think the aura that Anthropic got from having a model that was too dangerous to release was extraordinary. And maybe they had a GPU shortage... maybe they actually wanted to keep that proprietary model internal to extend their own lead." [00:41:22] He adds: "The sort of concept of the model that's too dangerous to release, it kind of conflates marketing, inference capacity, and then also economic considerations." [00:41:48]

$100M Seed Rounds Aren't Crazy Anymore — Being Too Small Is the Real Risk

Contrary to conventional VC caution around capital efficiency, Acharya argues under-ambition is now the bigger red flag than over-ambition. "You can talk about how you put a hundred million dollars to work in the seed productively... there's this sort of no ceiling on ambition is also showing up in how we're picking companies." [01:03:20]

Startups Have an Edge Precisely Because They'll Build What Incumbents Are Too Afraid To

He argues moral/reputational caution from Big Tech creates whitespace for founders willing to go into "uncomfortable" territory. "Those a thousand Google committees who would roll in their graves at the idea that they're going to release a model that's disagreeable... but guess what? Those are all parts of human existence. So I think exploring the kind of uncomfortable parts of our social existence are things that startups are uniquely set up to do." [00:33:41]

Most PMs Aren't Actually Good at Zero-to-One — And That's Fine

A subtly heretical claim about product management: the AI-enabled ability to "try everything" will reveal that most self-styled zero-to-one thinkers are mediocre at it, and that's a healthy outcome. "I feel like every PM at every company feels like they're a true zero to one thinker, but they're held back by the heavy hand of management... I think a lot of PMs are going to realize they're actually not that good at zero to one. And it's much more fulfilling to work on someone else's good idea than your own bad idea." [00:16:52]

Winner-Take-All Dynamics From the Mobile Era May Not Repeat

Against the standard VC pattern-match toward network effects and centralization, Acharya notes the current AI stack is unusually fragmented at every layer. "If you look at coding agents... your mental model for two years ago should have been winner take all. And yet, Claude Code, Codex, Lovable, Replit, Wabi... they're all sort of working." [00:04:44]

3. Companies Identified

Kavak — Used-car e-commerce company in Mexico. Cited as a model for AI adoption across all job levels, including non-technical staff. "They've got this concept of a Jedi Academy where they're teaching everybody at the company, including the mechanics, how to use the new tools and technologies. And kind of at the end of the six-week course, they ship a cutting edge in-production agent." [00:08:20] Also praised for its agent-per-customer architecture with human-in-the-loop escalation: "When the agent gets stuck, it actually calls a human and the human will coach the agent through... the agent then captures all the traces and learns from it." [00:18:49]

Cursor — AI coding assistant. Cited as proof that moats can be discovered rather than pre-designed. "They were criticized a lot for not having a moat, but it turned out that initially being a high-end PSDAU product was really good. And over time, they captured all the reasoning traces. They trained their own models. The Composer 1, 2 models." [00:53:34]

Wabi — a16z portfolio company; a platform for creating, consuming, and sharing many small apps via coding agents. "Probably the best example of a company that we've invested in is Wabi, where it's sort of a platform for many apps. People can create them, consume them, share them." [00:50:22]

Instinct — Emerging personal-agent startup gaining buzz for aggressive product trade-offs. "There's a startup that's really getting some buzz called Instinct, which has made some more aggressive and interesting trade-offs, but all in the same domain." [00:43:01]

GrokBot (xAI) — Praised repeatedly as the standout personal-agent product. "GrokBot's totally nailed it. And also the model underneath it is awesome... They came out of left field." [00:43:01] Named his favorite recent AI product: "It's just so unhinged for it to be so ambitious about sort of caching credentials and getting work done on your behalf... they actually are doing things that I think no other sort of big company would do." [01:12:55]

ChatGPT Work (OpenAI/Codex team) — Praised for full-duplex voice and cross-thread visibility. "It runs in the cloud... it's got the full duplex voice mode. So you can actually just call it and... say, hey, what's happening across all my coding agents... so you really feel like you're calling your assistant who knows everything that's happening in your world." [00:43:55]

Granola — Note-taking/meeting product cited as proof that craft and momentum can substitute for a classic "moat." "Two years ago it was really criticized and I don't know that they have a super strong durability story today. And yet it is like beloved and dominant." [00:55:58]

Decagon — Referenced via founder Jesse for the "moats are discovered, not designed" insight. [00:53:06]

Suno — AI music generation tool. "Suna has done such an amazing job" in the creative-tools/entertainment bucket. [00:50:49]

Mischief (historical fintech/creative studio) — Cited for the viral "Card vs. Card" experiment as inspiration for socially engaging fintech products. "They created this very funny product called Card vs. Card where they shipped, I think, 100,000 people a debit card... it was just this hilarious social experiment that went hyperviral." [01:01:36]

Moderna — Referenced as evidence of tangible medical breakthroughs happening now. "They just did... I mean the fact that Dario can write a blog post and say, what happens when we cure every disease? And then we debate it as a serious topic." [00:39:04]

Google — Cited as an example of a sophisticated large company embracing AI without layoffs, instead compressing roadmap timelines. "Have you laid anyone off? And he said, no, we didn't. What we instead do is now rip through our roadmap. So two years of roadmap happens in three months." [00:09:40]

Adaptive/Adams (likely referencing an ambitious frontier company, possibly misheard) — Cited as an example of previously "insurmountable" ambition now being pursued. "How crazy is Adams... what an incredible hero's journey for all of us collectively. But also what they're trying to do is something that... five or seven or ten years ago would have felt insurmountable." [01:04:23]

4. People Identified

Claire Vo — Product leader repeatedly cited as an exemplar of embracing AI through constant shipping. "Claire is my muse. She's so awesome because she's always shipping... She's not afraid to be a little embarrassed by it. And if you just see her whole kind of affect, she feels like the best version of herself that she's ever been." [00:10:09] Lenny adds: "I want to be Claire Vo when I grow up." [00:10:53]

Ale (Kavak) — Described as one of the most sophisticated operational thinkers on AI agents Acharya knows. "He's probably the most sophisticated thinker on this stuff that I get to hang out with." [00:18:20]

Eugenia — Founder recommended as an ambitious thinker on new consumer UI paradigms. "Founders like Eugenia, who you should have on the show, she's tremendous, are thinking really ambitiously about user interfaces." [00:35:23]

Brian Chesky (Airbnb) — Noted for starting a research effort on next-gen user interfaces. "Brian Chesky from Airbnb, I think he started a foundation lab focused on next gen user interfaces." [00:35:23]

Mark Andreessen — Cited for his "AI came just in time" thesis regarding demographic and productivity decline, and for shaping the firm's philosophy and founder ambition broadly. "The way we sort of told our story when we were raising our first fund was... we were going to the moon or we were going to leave a moon sized crater in the ground and there was no other option." [01:03:49]

Ben Horowitz — Praised for authoring "the first emotionally honest book about business" (Hard Things) and for embodying industry stewardship. "That's why founders love it. That's why I love it. Because you read it and you're like, wow, I'm not the only one that's anxious." [00:54:48]

Hamilton Helmer — Author of "Seven Powers," cited as one of only a handful of legitimate business books. "I feel like there's only five real business books in the world... his is on my list of five." [00:54:36]

Chris Dixon — Referenced for originating "come for the tool, stay for the network," which Acharya says needs updating for the current era. [00:58:04]

Nikhil Singal — Former Credit Karma colleague, credited with the insight that people's AI attitudes flip once they experience a moment of AI-created joy. "He finds that people flip on AI and how they feel about it once they find some moment of joy that it had created for them." [01:09:19]

Sundar Pichai — Referenced regarding CEO incentives to grow rather than merely optimize. "Sundar running Google, he doesn't want to run a more efficient $4 trillion company. He wants to build a $40 trillion company." [00:09:40]

Tara — Named as the OpenAI PM behind ChatGPT's "work" product, referenced as a recent guest on Lenny's podcast. [00:43:29]

Dario Amodei — Referenced regarding his public writing on curing all diseases and representing the more cautious end of the jobs-disruption spectrum. [00:39:04], [00:44:38]

David Sacks — Referenced as representing the optimistic end of the spectrum on AI and jobs, which Acharya says he's closer to. [00:44:38]

Thomas Sowell — Author of "Conquest and Cultures," Acharya's favorite book, on how culture drives outcomes across societies. [01:10:14]

Brian Arthur — Author of "Increasing Returns to Scale," recommended by Mark Andreessen, on the outlier economics of software. [01:10:42]

Ron Conway, Brooke Byers, Tom Perkins — Referenced as earlier generation investors who embodied a sense of stewardship for the tech industry, akin to Andreessen and Horowitz. [01:05:33]

5. Operating Insights

Build a "Chassis" Project Rather Than Searching for a New Idea Each Time

Acharya's tactical advice for staying current with models: maintain one or two ongoing side projects that serve as a testbed for every new model release, rather than trying to invent a new idea from scratch each time. "I'd say like, work on something. It's actually better if it's not important with a capital I. And then keep finding new ways to invest in it and add to it with a model as a kind of tool rather than as a goal." [00:28:37]

Diagnose Agent Failures as Either a Knowledge Gap or a Data Gap

When an agent underperforms, the operating discipline is to ask a specific diagnostic question rather than generically "improving prompts." "Anytime the model's making a mistake or doing something you wouldn't do. What do you know that it doesn't know?... It's either a knowledge gap or a data gap that you have to give the agent. And the next time, it shouldn't have to call you." [00:18:20]

Never Build a Platform and a Studio Simultaneously

From his own founder failure at his first company: "Don't build a product in a platform at the same time. You know, if you're going to be a platform company, build one. Our first company would try to build a sort of a social platform for mobile games and be a gaming studio... it took us years to figure out we were wrong." [01:14:04]

Reframe Every Function as an "Input-to-Impact" Loop and Locate the Blockage Point

Practical management heuristic for reorganizing teams around AI: identify where the loop stalls and feed it more context, rather than assuming full automation is possible everywhere. "Basically, the takeaway here is you need to start thinking about every function as an agent loop. And the job is to figure out where it gets blocked, where it goes wrong, and give it more context, more insight, more direction." [00:19:13] (Lenny's synthesis, confirmed by Anish: "That's right.")

Test Product Ambition by Pricing at the Extreme

A generative exercise for product teams facing a "growth problem": ask what the product would need to do to justify a $1,000/month or $10,000/month price point, which forces genuine differentiation rather than incremental growth tactics. "If we imagined our product cost a thousand dollars a month, ten thousand dollars a month, like what if our product was a software Birkenbag? What would it have to do to justify that? Okay, let's figure out how we build that." [00:59:25]

6. Overlooked Insights

The Majority of AI Companion Users Are Women in Their 40s and 50s, Seeking Boyfriends, Not Girlfriends

This was mentioned almost in passing but overturns the default cultural assumption (usually framed around lonely young men and AI girlfriends) about who is actually using AI companionship products and why. "Though actually, to be accurate, it's more boyfriends than girlfriends, actually, you know? The majority of people using companion products are women that are in their 40s and 50s, actually." [00:51:59] This has major implications for consumer AI product design, targeting, and the broader narrative that AI companionship is primarily a "male loneliness" phenomenon — the actual market may be entirely different demographically, pointing to unmet emotional/social needs among middle-aged women that current products aren't explicitly designed around.

GPU/Compute Shortage May Be a Hidden Driver Behind Publicized "AI Safety" Pauses

Buried in a broader answer about OpenAI slowing RL training, Acharya floats an economic explanation that reframes a widely-reported "safety" story as possibly a capacity or competitive-strategy story: "Maybe I'm a little skeptical on some of these things where I think the aura that Anthropic got from having a model that was too dangerous to release was extraordinary. And maybe they had a GPU shortage. Now, maybe it was the capabilities were more advanced than they actually wanted... maybe they actually wanted to keep that proprietary model internal to extend their own lead." [00:41:22] This is a significant reframing for investors evaluating lab valuations and compute demand — it suggests publicized "pauses" may be as much about compute scarcity and competitive lead-extension as genuine safety concerns, which has direct implications for how to interpret lab PR and compute capex cycles.