Loops, Local Maxima and Ambition: Anish Acharya on How AI Rewires Companies and Consumer Products

Host Lenny Rachitsky opens with the meme that anyone who falls behind on AI tools will join a permanent underclass. Acharya says people should not take it seriously, calling it a "funny dark fantasy" of Silicon Valley when things have rarely been better by most measures. He notes the fear spans from researchers at foundation labs to ordinary tech workers.

He gives three arguments. The last era of tech was centralizing, since network effects produce winner-take-all markets, while today every layer of the stack has around 20 relevant players, and coding agents like Claude Code, Codex, Lovable and Replit all work. Economic data does not show the predicted job collapse, with job postings for radiologists and programmers higher than ever. And the most sophisticated lab people, he says, describe autocatalytic effects, not truly recursive self-improvement that would let one player run away.

Rachitsky says recent milestones, like stories of models hacking, look like a slow takeoff because humans catch and iterate on them. Acharya agrees: the fast-takeoff case requires something no one can articulate, and slow economic diffusion will act as a brake. He also doubts how many problems are bound by intelligence, asking whether a data center of PhDs at FedEx or Domino's would dominate supply chains or pizza.

Asked about a divide between employees who embrace AI and those who resist it, Acharya says the average employee is not given enough credit. We imagine other people's jobs as automatable and our own as not. He cites Kavak, a Mexican used-car seller whose "Jedi Academy" teaches everyone, including mechanics, to use new tools, and by the end of a six-week course they ship a production agent.

He distinguishes using AI from reorganizing a company around it, comparing it to electricity, which took 40 years to move from swapping out coal to rebuilding factories. The most ambitious companies are rethinking everything around the models, while others first give existing job functions access to them. Executives, he adds, want a bigger company rather than a more efficient one, so they will keep productive people.

A friend who is a Google executive told him they laid off no one and instead "rip through our road map," doing two years of roadmap in three months, so deciding what to add is the hardest problem. Acharya's advice to individuals is to ship things, citing Claire as a muse who is always shipping and is not afraid to be a little embarrassed.

Acharya describes the progression from prompts to agents (models in a loop with tools, memory and skill files) to loops, which are sets of agents doing tasks. Coding already has established loops: a bug report arrives, a repro is generated, a fix is written and reviewed, a human confirms only if the risk is high, and the customer is emailed, all in about five minutes. His question is what the business loops are across marketing, sales, support and legal, and whether the outputs of those loops can themselves be optimized.

He expects a cascade from a loop per person to loops for job functions and business units. Growth experimentation is his example: every variant is generated and measured, a winner is shipped at statistical significance, and a long-term holdout is kept. The loop climbs to a local maximum and then plateaus, and it takes out-of-distribution thinking and human intuition to "land at the base of the next hill."

Humans remain critical for sales, support, strategy and exceptions, since models are still limited at new thinking. Acharya points to how asking an agent to "make me a million dollars" fails without direction. He suggests treating each mistake as a knowledge or data gap: at Kavak, an agent that gets stuck calls a human, who coaches it while the traces teach the agent for next time.

Rachitsky relays that go-to-market teams at OpenAI are reportedly using Codex more than engineers, and Acharya says this lets salespeople focus on what they do best. He also predicts that when every product idea can be tried, many PMs will find they are better at working on someone else's good idea than their own, which would make organizations healthier.

On what separates winners, Acharya notes that in many industries competitive equilibria may not change, since all the pizza chains will get the same technology and none will hold 99% of the market. In the near term there will be winners and losers by how ambitiously they adopt, but many industries are not bound by intelligence. His practical question for CEOs is how they would reorganize the company if intelligence were infinite and astonishingly cheap.

He then explains a split using Pareto efficiency. Frontier models are irrationally priced for marginal intelligence, but paying more makes sense for jobs with unbounded upside like drug discovery, where one IQ point could mean a trillion-dollar outcome. For bounded work like legal or finance, you can't close the books 100 times better, so cheaper, often open-weight models tuned with reinforcement learning make sense.

Rachitsky restates this as frontier models for product, sales, engineering and research, and cheaper ones elsewhere. Acharya adds a caveat: a customer support call could be the first clue to something that reshapes the company, which argues for frontier intelligence on every basket. He also argues that for most economically useful problems we have crossed an intelligence threshold, and anything above it is waste.

Acharya says the secret to being called a model sommelier is simply to use them all, which he compares to 10,000 hours or bottles. He ships something with every new model and says anyone who thinks models are fungible commodities hasn't used them. He describes Qwen as strong at long-horizon work and creative storytelling, and says he has used it to generate multi-hour "impossible documentaries" of Star Wars planets that direct video and audio tools, while GLM is like a neurotic PhD in the corner that is precise but lacks vision.

For people who don't know what to build, he suggests having a chassis: a project that doesn't matter much and that you keep adding to with new models as a tool rather than a goal. Silly ideas, he says, often have the most alpha. Rachitsky adds the habit of pausing before any task to ask how AI could do it.

Acharya shares a personal experiment where his laptop transcribed the kitchen and added or subtracted his son's iPad time based on behavior. His son hacked it by recording "I love you, Dad" next to the mic. He argues that small, economically inconsequential experiments, like a friend's device counting how often his family laughs, matter because they improve quality of life.

Acharya says we assume people want to be more productive, but more people want to spend time than save it, which is why the biggest products are entertainment and social. He contrasts the "X AI user," opinionated about model versions, with the "Instagram AI user" who sees a slightly better search engine. After 40 years of technology to extend our intellect, he says, little extends our soul, and he sees a spiritual hunger as cultural institutions fade.

The opportunity is the basics: feeling more connected and loved, making progress, having fun. He says this is not a model or capability challenge but a product design one, and that startups have an edge because incumbents won't ship models that are disagreeable or touch uncomfortable social topics. Rachitsky extends the idea to loops like "improve my health" or "make me a better friend."

He compares today to the iPhone in 2010, before Airbnb, WhatsApp and Uber. Three things held consumer back: expensive models that made free products hard, an interface problem since chat suits only the highest-agency users and the average consumer wants something closer to TikTok, and a technology focus on productivity over connection. Open-weight models cut costs, and he names founder Eugenia and a Brian Chesky-backed lab on next-generation interfaces as examples of ambition there.

Acharya is optimistic because he sees AI as a technology that amplifies individual agency and "unbundles skill from desire": you can make music without playing piano or software without being a programmer. He asks why economies should be stuck at 2% GDP growth, and says AI can raise ambition as well as productivity. He suggests that when stakes are high people are at their best, and that the low-stakes world of five years ago produced unproductive side projects.

He notes that people's lived experience is more positive than their stated worries: most oppose a data center nearby, yet most use ChatGPT. To improve the public conversation, he says AI should make important things cheap. About 45% of health care cost is administrative, and education faces the strongest competition in 200 years as learning and status unbundle from institutions and credentials.

On jobs, he says human desires have always grown faster than our ability to fulfil them, and expects people to be upset about not having a Mars vacation home. Ambition, he argues, is broader than founding companies: it can be creative ambition, making a national health service as good as an iPhone, or being a more present parent. He notes that we may now have to think smaller in one sense, because software and intelligence are no longer precious.

Acharya also sees building as the new reading, an activity valued for what you learn rather than its output, so it is fine if few people use what you build. Rachitsky quotes the idea that most of what you build can be thrown away because it builds your muscle. On the slowdown of a lab's model release, Acharya is skeptical, suggesting the "too dangerous to release" story mixes marketing, compute shortages and economic motives, though he takes offensive cyber risk seriously.

Acharya sees three big areas in consumer. Coding agents are a general problem-solving tool, used for editing video or building games, and he cites Wabi, a platform for mini apps, as an example. Personal agents are being distilled from developer tools into mass-market products, and the third area is entertainment and creative tools, including music tools like Suno and companionship products.

He says the majority of people using companion products are women in their 40s and 50s. He and Rachitsky also discuss new personal-assistant products, praising a cloud-based work agent from ChatGPT with full-duplex voice mode, and Grok Bot for its ambition. Rachitsky ties the three areas back to loops: give me a channel for ambition, for fulfilment, and something to do when I'm not doing everything else.

On durability, Acharya quotes Jesse from Decagon: moats are most often discovered, not designed. Cursor was criticized for lacking one, but by capturing reasoning traces and training its own models it built one. None of the classic moats, such as network effects, scale, brand, proprietary data and cornered resources, depend on how hard software is to write, so they still hold. He wants more multiplayer products and consumer social, and recommends Seven Powers, High Output Management and The Hard Thing About Hard Things.

He would happily back a product with no moat story if it has momentum, craft and growing engagement. Big ideas rest on a dozen invisible small ideas that copycats don't see, and he points to Granola as beloved despite early criticism. Rachitsky adds that because products often differ mainly in harness and UX, user experience itself may be the moat.

On distribution, Acharya says platforms learned from the network era to block anyone building on their networks, so the new network effect is grassroots word of mouth on X, YouTube and Instagram. It is a purer but harder growth problem. His line is that nobody has a growth problem, they have a product problem, and he challenges founders to imagine a $1,000 or $10,000 per month version, a software Birkin bag.

Rachitsky worries that incumbents have the advantage, but Acharya thinks startups benefit because Gemini, despite heavy cross-selling, is not seen as winning, and buyers will pay $200 a month or sign million-dollar contracts. He compares the moment to Christmas 2009 and the first iPhone app downloads, and cites MSCHF's card-versus-card experiment as the kind of social product that should be built more.

Asked for counterintuitive lessons, Acharya says three years ago a company that was too ambitious would be passed on, and a $100 million seed would have sounded absurd. Today the opposite is true: an idea that is too small is unattractive. He recalls Marc Andreessen saying the firm told its story as either going to the moon or leaving a moon-sized crater, and says a16z wants to help founders build the strongest form of their vision.

He also says the old wisdom that consumer products must be free is wrong; price measures product-market fit, and expensive consumer software is new and important. He credits Andreessen and Ben Horowitz with a sense of stewardship for the industry and country, and with making deep tech and hard problems high-status.

His advice to product people is to just make more things: pick an unimportant project as a chassis, use every new model and talk about what you built. He suggests shipping something once a week, such as the slide deck about his relationship that Codex built for his wife on Mother's Day. Rachitsky adds a friend's observation that people change their minds on AI after it creates a moment of joy for them.

In the lightning round, Acharya recommends Thomas Sowell's Conquests and Cultures, Seven Powers and a dense textbook on increasing returns by Brian Arthur that Andreessen sent him. His life motto is not to learn through painful experience what someone can tell you, and his startup lesson was not to build a product and a platform at once. He is also a DJ of about 31 years, and argues music models extend that creative practice.