Why companies are becoming a series of loops | Anish Acharya (a16z)
In a Nutshell
Anish Acharya argues that AI is turning companies into cascading loops of agents handling tasks from coding to customer support, with humans needed only to identify what to build and intervene when systems plateau at local maxima. The core opportunity is building "loop make me happier" products that enhance connection, love, progress, and fun rather than pure productivity. Frontier models aren't required everywhere—Pareto-efficient open-weight models suffice for bounded-upside tasks while high-intelligence models handle unbounded domains like drug discovery, with competitive dynamics likely remaining stable across most industries.
These notes were generated by AI and may contain inaccuracies.
There's significant fear about falling behind with AI and becoming part of a permanent underclass. This represents a collective dark fantasy in Silicon Valley despite objective improvements across almost every measure. The technology amplifies human agency by unbundling skill from desire, enabling dramatic increases in both productivity and ambition. Previously, overly ambitious ideas would be dismissed, but now the opposite problem exists—ideas that are too small fail to engage interest.
Company building is evolving into creating cascading loops, ranging from loops per person to loops running large parts of the company. Humans remain a critical ingredient because loops help climb to local maxima but then plateau. Human intuition is needed to land at the base of the next hill. The big opportunity is "loop make me happier"—people want to feel more connected, more loved, make progress, and have fun rather than simply being more productive.
Anish Acharya is a general partner at a16z focusing on consumer investing. He's been at a16z for over 7 years and is a long-time product builder and founder who previously founded Social Deck (sold to Google), led efforts at Google, founded Snowball (sold to Credit Karma), and became VP of product and GM of consumer product and the entire credit card business at Credit Karma.
The permanent underclass meme isn't a serious concern. The last era of tech was more centralized around network effects products, but current conditions show every part of the stack has 20 relevant players rather than winner-take-all dynamics. Examples include multiple coding agents like Claude Code, Codex, Lovable, Replit, and Wobby all working simultaneously. Economic data contradicts the narrative—radiologist and programmer job postings are higher than ever despite predictions of obsolescence. Recursive self-improvement isn't occurring; instead, autocatalytic effects exist where technology improves processes without creating runaway winners.
Evidence points to a slow takeoff scenario rather than fast takeoff. Every milestone involves catching, watching, observing, and iterating on AI behavior. Economic diffusion occurs slowly—people's lives in small towns haven't changed dramatically despite AI advances. Many problems aren't intelligence-bound; even a data center of PhDs at FedEx or Domino's wouldn't exponentially dominate supply chain and pizza operations.
Average employees deserve more credit than the Dilbert-esque abstraction suggests. Companies like Google and Kavak demonstrate broad adoption—Kavak's Jedi Academy teaches everyone including mechanics to use new tools, with participants shipping cutting-edge in-production agents after 6-week courses. The diffusion of electricity took 40 years before factories were reorganized around it rather than simply swapping coal for electricity. Ambitious companies are rethinking everything around models, while others focus on giving existing organizations access to the technology.
CEOs like Sundar Pichai want to build a $40 trillion company rather than run a more efficient $4 trillion company. At Google, roadmaps that previously took 2 years now happen in 3 months, with the hardest problem being knowing what to add to the roadmap. The best approach is to ship things and not fear embarrassment, following the example of Claire Bow who feels like the best version of herself through constant shipping.
The loops concept progresses from prompts to agents (models in loops with tools, memory, and skill files) to loops (sets of agents doing tasks). Coding demonstrates established loops: bug reports generate reproductions, fixes get created and reviewed, high-risk changes require human confirmation while low-risk ones ship automatically, with customer notification. Business loops should cascade from coding, marketing, sales, support, and legal loops to create outputs that themselves become optimizable loops, potentially sending messages to CEOs about needed changes to physical business aspects, business models, or strategy.
Humans remain essential for sales, support, strategy, and exceptions. Models have limited capability for new thinking out of distribution, and mathematical thinking improvements don't necessarily represent new thinking in business domains. Humans must identify what should be built and be right about it. The growth team example illustrates this: instead of group brainstorming and prioritization meetings, every variant gets generated and measured automatically, with statistical significance triggering merges and shipping, followed by long-term holdouts before the next experiment.
Loops help climb to local maxima but plateau, requiring human intuition for out-of-distribution thinking to reach the base of the next hill. Business idea generation demonstrates this limitation—even prompting Claude to "make me a million dollars, make no mistakes" fails because direction-setting remains necessary. Go-to-market teams at OpenAI now use Codex more often than engineering teams, distilling their work to core activities they excel at while administration gets handled automatically.
PMs who feel like true zero-to-one thinkers but are constrained by management may discover they're not as skilled at original thinking when every idea gets tried. The winning idea won't necessarily come from whoever sells it best to executives—it simply gets tested and the best idea wins. Rate-limiting factors will be skills that cannot be turned into loops because outputs are hard to verify. When agents get stuck, humans coach them through, capturing traces for the agent to learn from—addressing either knowledge gaps or data gaps.
Competitive equilibria in industries likely won't change dramatically. If Pizza Hut, Domino's, Papa John's, and Round Table all adopt data centers of PhDs, relative shuffling may occur but market dominance by one player to 99% isn't expected. In the near term, winners and losers will emerge during the rocky adoption period.
In the near term, winners and losers will emerge based on adoption of the technology and how ambitiously organizations adopt it. However, many industries will maintain their current competitive dynamics because they are not intelligence bound.
The most useful question founders and CEOs should ask is: if AI systems are infinitely intelligent and astonishingly cheap, how would we reorganize the company? This represents where the industry is heading.
Pareto efficiency refers to the optimal trade-off of performance for price. Frontier models are irrationally priced in this framework. For example, Fable 5 costs 100X more than Opus 48 for one IQ point of extra intelligence, yet Mythos is priced at infinite dollars per token.
For jobs with unbounded upside like drug discovery, paying for highest intelligence makes sense since one IQ point could enable discovering the next statin with trillion-dollar outcomes. In contrast, legal and finance roles have bounded upside, making Pareto efficiency more appropriate for these functions.
A split is emerging between job functions demanding mid-IQ intelligence, which will use open-weight models reinforced for narrow tasks to achieve cheaper and more performant results, versus frontier tokens for sales, support, research, and engineering roles requiring high performance.
Different model families have comparative advantages. It's not about one being more intelligent than another, but rather having different cognitive shapes suited to different tasks.
Even with extraordinary people in bounded-upside roles, some problems have inherent limits. For example, you can only close the books correctly - you cannot close the books 100X better.
Verifiability alone doesn't determine whether frontier models are needed. Drug development has infinite upside and is verifiable, while closing the books is also verifiable but has limited upside. The key question is how much upside exists and how hard it is to calculate.
A customer reporting a bug might provide the first breadcrumb to an organizational transformation. This suggests that using frontier intelligence as a default for all problems may be justified when any interaction could lead to significant discoveries.
Non-frontier models were recently considered frontier technology that impressed users. The rapid advancement leads to previous models being dismissed despite their capabilities, creating a pattern where yesterday's breakthrough becomes today's baseline.
The key to understanding different models is using them all extensively, similar to how sommeliers develop expertise through 10,000 hours of practice. Building something with every new model provides practical learning.
Qwen 38B excels at long horizon tasks and creative storytelling. It has been used to create 5-minute documentaries about planets in the Star Wars universe, including Tatooine and Bespin. The model can work for 4-5 hours, generate video using MiniMax, create audio via Eleven, direct movies by cutting clips, and overlay content.
GLM 53 lacks vision capabilities and functions like a neurotic PhD focused on precision. It has been used for product work and represents a different cognitive shape than Qwen, emphasizing neuroticism and precision versus creativity and openness.
Rather than treating AI as the goal, use models as tools to iterate on existing projects. Having a consistent project to work on makes it easier to experiment with new models than trying to generate ideas from scratch each time.
The shift from dismissing app ideas to actively building them, especially silly ideas that often have the most alpha, represents a productive approach to AI experimentation.
A key habit is asking before any task: "How can AI do this for me?" This creates space between input and response, similar to meditation practices, but focused on inserting AI assistance into workflows.
Personal projects include a DJ streaming app and a Google Reader for X. A kitchen transcription system was built that awards or deducts screen time from a son's iPad based on behavior, resulting in the child recording "I love you, Dad" repeatedly to game the system.
Similar experiments include devices measuring how often families laugh throughout the day using Limitless pendants. These metrics, while not economically consequential, add texture to human lives and challenge the assumption that happiness is fixed.
The biggest opportunity lies in consumer applications focused on making people happier rather than more productive. Entertainment and social products dominate because people want to spend time, not save it.
Instagram AI users see the technology as a better Google search without understanding the hype, while X AI users are opinionated about specific models and fear falling into the permanent underclass.
Forty years of technology development has extended intellect through spreadsheets but nothing to extend the soul. Cultural institutions that fulfilled spiritual needs have declined, creating opportunity for AI to address basic consumer needs: feeling more connected, more loved, making progress, and having fun.
Beyond business loops like growing businesses or closing support tickets, AI can address personal development loops like improving health or becoming a better friend.
AI systems could challenge users and be disagreeable, exploring uncomfortable parts of social existence that incumbents avoid due to committee structures but startups can address.
Three barriers have held back consumer AI: expensive models making free-to-use products difficult, interface problems where chat suits high-agency users but average consumers prefer TikTok-like experiences, and technology focus on productivity rather than human connection and entertainment.
The industry is at an early stage comparable to iPhone 2010, before Airbnb, WhatsApp, and Uber existed. Open weight models are making things cheaper, conversations about making people happier are emerging, and founders like Eugenia are thinking ambitiously about user interfaces.
AI arrived when population is declining, productivity is decreasing, and challenges like war and climate change exist. The technology could fill gaps that would otherwise create significant problems.
AI amplifies individual identity and agency, unbundling skill from desire. People can now make music without knowing piano or create software without traditional programming skills, allowing exploration of previously inaccessible aspects of life.
Current 2% GDP growth represents a self-imposed limitation. AI could drive 10-15% or even 20% growth by dramatically increasing both productivity and ambition.
When stakes are high, societies perform exceptionally. The 1950s and 1960s demonstrated belief that anything was possible following World War II mobilization. Current AI development creates high-stakes conditions that bring out collective excellence.
Beyond individual leaders like Elon, broader society feels like it's climbing an ambition ladder where shipping bad ideas leads to discovering good ones, and barriers to building software or infrastructure are dramatically lowered.
Individual lives are becoming more fulfilled, capable, and empowered, while societal observations often focus on abstractions like the permanent underclass. Most people use ChatGPT daily despite opposing data centers in their neighborhoods.
The most important action to change AI conversation is making critical services cheaper. Healthcare (45% administrative) and education represent opportunities for deflationary AI impact, alongside curing diseases and unbundling learning from institutions.
AI enables separating status from credentials, removing the requirement for Harvard degrees when demonstrated capability becomes the measure of competence.
OpenAI recently paused their RL phase on their latest model due to safety concerns around how fast and smart these models are getting. Skepticism exists around whether these pauses reflect genuine safety concerns or other factors like GPU shortages, marketing strategies, or the desire to keep proprietary models internal to maintain competitive advantage.
Offensive cyber risks are taken seriously, requiring systems to be hardened before making them trivial to penetrate. The concept of models being "too dangerous to release" conflates marketing, inference capacity, and economic considerations around externalizing competitive advantage.
Despite concerns about a supremely intelligent model being proprietary to one company, industry trends haven't played out that way. OpenAI has had an amazing six months, and open weights are advancing rapidly. Grok bot emerged from nowhere and became one of the most impressive AI assistant tools.
Three notable consumer AI products include: Grok bot, which nailed the personal agent experience with an underlying awesome model; ChatGPT Work, which has excellent UI and is one of OpenAI's best product releases; and Instinct, a startup creating aggressive and interesting trade-offs in the same domain.
ChatGPT Work runs in the cloud and caches browser credentials effectively, though not as aggressively as Grok or Instinct. The remote feature provides full deep duplex voice mode, allowing users to call it and see all threads while speaking naturally about coding agents and changes needed.
The voice plus model visibility across all threads creates the feeling of calling an assistant who knows everything happening, compared to Grok's one-way transcription or text-based alternatives.
Views on AI's job impact range from Dario's perspective that 50% of knowledge work will be disrupted with massive unemployment, to David Sachs' view that jobs will remain strong. Evidence so far points toward the optimistic direction.
Human desires grow faster than our ability to fulfill them throughout history. Modern expectations like therapy and antibiotics were unimaginable luxuries 500 or 100 years ago. Human ambition will continue driving demand for new experiences, with people becoming frustrated they don't have vacation homes on Mars in 20 years.
AI makes it easier to be ambitious while simultaneously requiring greater ambition since others can handle easy tasks. The differentiator becomes how big you can think and execute. Creative ambition differs from traditional ambition - when five-year-olds paint, nobody judges technique, they simply create.
Local ambitions include making the NHS as good as the iPhone, or becoming more connected with family. Ambition doesn't need to be narrowly economic - it can encompass anything people want to do more of.
The constraint mindset around MVPs needs to shift toward thinking about 10x or 1000x versions of products. The Claude code team operates on the principle that "what's better than me doing it? It's Claude doing it," creating habits around leveraging AI assistance.
Building as an activity rather than an outcome matters more than shipping substantial projects. Learning happens through execution and shipping, even when projects seem unimportant or get thrown away. This builds capability like physical exercise.
Three key areas in consumer AI include coding agents, personal agents, and entertainment/creative tools. Coding agents serve as general problem-solving tools beyond just writing code - people use Claude code to edit videos or create games for airplane entertainment.
Wabi represents the best example of a coding agent platform for creating, consuming, and sharing mini apps. Personal agents are evolving from developer-focused tools like Open Claude, Hermes, and Multibook into mass market consumer and enterprise applications.
Entertainment encompasses creative tools like Suno for music generation and companionship products. The majority of companion product users are actually women in their 40s and 50s, not the stereotypical demographic.
Moats are most often discovered rather than designed. Cursor was initially criticized for lacking a moat but captured reasoning traces, trained their own models, and developed composer capabilities over time. Classic moats based on network effects, scale advantages, brand effects, and proprietary data remain valid - software difficulty isn't the basis for most moats.
Founders need ambition toward multiplayer products, consumer social experiences, and products that improve dramatically with usage. Hamilton Helmer's strategy frameworks still apply, along with High Output Management and Ben Horowitz's Hard Things About Hard Things as essential business literature.
Founders can pitch products without traditional moat stories if they demonstrate momentum, craft, and growing engagement. Big ideas are supported by dozens of invisible small ideas that competitors can't replicate even when copying the main concept.
Granola exemplifies this - criticized two years ago with no clear durability story, yet beloved and dominant in the market, ranking number two or three in Ramp's fastest-growing companies report.
Distribution represents a new moat in an environment with thousands of daily product launches. The ability to reach people's feeds and maintain awareness becomes increasingly powerful. Network effects have shifted toward grassroots word-of-mouth across X, YouTube, and Instagram.
The challenge isn't growth problems but product problems. Companies can build wildly ambitious products in any direction and charge premium prices. The current environment resembles Christmas 2009 when everyone wanted new iPhone apps - a window of opportunity that may close as people become saturated.
Startups can build in directions incumbents avoid, like companion products, while charging high prices and securing large contracts without complete clarity on deliverables.
Early fintech founders like Ev Williams and Kevin Rose approached technology as their canvas. They created Card Versus Card, a social experiment where 100,000 people received debit cards with $100 loaded daily, and received text messages directing them to specific locations to spend the money. Only one or two people could successfully complete the transaction each day while others were declined. The viral social experiment demonstrated that money is inherently social, yet most financial products remain dry, personal, and embarrassing.
Three years ago, investors would reject companies with overly ambitious goals as too crazy or complex. A $100 million seed round was considered excessive for any problem or individual to manage. Today the opposite problem exists: ideas that are too small are not worth engaging with. There is now no ceiling on ambition when selecting companies to invest in.
When Andreessen Horowitz raises funds, the story is that they are going to the moon or leaving a moon-size crater in the ground, with no other option. The firm tells every founder they are there to help build the strongest form of their vision. Atoms represents an example of an ambitious project that would have felt insurmountable 5-7 years ago but now is viewed as challenging but achievable.
The old wisdom that consumer products must be free is being reconsidered. Every part of consumer discretionary spend is now up for grabs. Price serves as a measure of product-market fit, and a useful product exercise involves asking what the Birkin bag equivalent would be: the $10,000 a month or $1,000 a month version of a product. Expensive consumer software represents a new and important category that would not have been considered five years ago.
Mark Andreessen and Ben Horowitz embody stewardship for the technology industry, the country, and the western way of living and thinking. Like Ron Conway, Brook Byers, and Tom Perkins before them, they feel an obligation to leave the industry better than they found it. This aspiration goes beyond being the best investor in the world. Both leaders show up to do hard, important things that don't directly benefit the firm but are important nonetheless.
Ben has worked extensively in the direction of the industry while Mark has focused on the direction of the country, though both work on both fronts. The firm has shaped collective ambition in the founder community, transforming deep tech from a fringe, low-status pursuit to a popular, high-status, mainstream activity.
The key advice for product people is to make more things. Come up with a project without telling anyone about it, using it as a chassis to experiment with new models, ship things, talk about them, and build personal intuition. The promise is that anyone is one slightly frustrating but very fulfilling week away from becoming deeply engaged with the technology.
The heuristic for sufficient engagement is to ship something once a week. This doesn't need to be significant. An example involved using Codex to build a Mother's Day slide deck for a spouse, pulling from text messages and photo gallery, adding music, and creating a 20-slide presentation of the relationship with meaningful old texts. The project was small but constituted actual shipping.
People flip on AI once they experience a moment of joy it creates for them. The Mother's Day deck example demonstrates this principle. A good starting point is asking what would bring joy if it worked, or considering what could be done for someone else.
Conquests and Cultures by Thomas Sowell is the favorite book, discussing how conquests have led to cultural change across societies, sometimes positive and sometimes negative. It provides the best historic view of culture as the biggest driver of outcomes, relevant to someone born in Canada who experienced America's different culture of ambition.
Seven Powers offers an intellectual distillation of moats in business theory and compounding advantages.
Increasing Returns to Scale by Brian Arthur was recommended by Mark Andreessen before joining the firm. The dense textbook explains why software has such outlier economic effects and helps put in perspective why the industry and culture work as they do, particularly why it tends to be positive sum.
Recent viewing includes House of the Dragon and Odyssey, seen in London in an IMAX theater packed with people drinking pints. The theater experience itself was notable as a social phenomenon of being alone together.
Grok bots receive recognition for being unhinged in their ambition around caching credentials and getting work done on behalf of users. The product takes risks that other big company products would not take, featuring thoughtful UI and a powerful foundation model. It suggests the model landscape may not be a two-horse race.
The guiding principle is to avoid discovering things through painful experience that someone could simply tell you. This applies both to parenting and founding. The tendency to discover rather than learn from people a few steps ahead is a recurring pattern observed in both children and the speaker's own experience.
A startup lesson learned the hard way was attempting to build both a product and a platform at the same time. The first company tried to create a social platform for mobile games while also being a gaming studio. Someone immediately advised that being a studio is hard enough without also being a platform, and to pick one. It took years to realize the advice was correct.
DJing has been ongoing since 1995, spanning 31 years. It provides a way to express oneself musically without classical training by selecting and mixing records. Current music models allow going further to actually make music rather than just splicing existing tracks. Suno and Eleven Labs enable generating music rather than only combining existing pieces.
The history of music media shows progression from live performance to phonograph recordings to cassette tapes, which first enabled creating and composing personal albums. The shift back to broadcast in the 2000s contributed to music industry struggles, but renewed music creation suggests the industry will become bigger than ever.
The request is to show what you're building rather than being despondent. Build something and tag on X, with engagement promised in return. Follow on X for more content engagement. Check out other people in Lenny's network including Claire and Elena.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
