Companies Are Becoming a Series of Nested Loops, but a Loop Only Climbs to the Local Optimum
After prompt comes agent, after agent comes loop. A loop will carry you up to the local maximum and then plateau, and at that point you need human intuition to land you at the foot of the next mountain.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Nobody funds an idea that is too small anymore
Anish says that three years ago, if a company wanted to do something too grand, a16z would not invest; today it is the reverse — ideas that are too small, too local, they do not want to touch. The reason is that model capability has torn down the "can it be done" threshold, and the remaining constraint is ambition itself. He calls this technology "unbundling skill from desire": not only can productivity be amplified, the scale of what you want to do can be amplified too.
— Anish AcharyaIt is not recursive self-improvement, it is autocatalysis
Anish pushes back on the "fast takeoff" narrative: the people in the labs who understand it best do not say RSI (recursive self-improvement), they say autocatalytic effects — using new technology to improve your own process, but not truly recursively. He gives two counterexamples: first, economic diffusion is extremely slow, and when he went back to his hometown he found people's lives had barely changed; second, many problems are simply not intelligence bound — give FedEx or Domino's a data center full of PhDs and it still will not let supply chains and pizza crush everything exponentially.
— Anish AcharyaUsing AI and reorganising around AI are two different things
Anish uses the diffusion of electricity as an analogy: from the invention of electricity to factories actually reorganising took 40 years, and it meant burning down the old plants and rebuilding, not swapping coal for electricity. Today the most aggressive companies redesign everything around models; slower companies just grant permissions to existing roles and functions. He also cites the Mexican used-car company Kavak's Jedi Academy: a six-week course that teaches everyone, including mechanics, to use the new tools, and at graduation each person delivers an agent in the production environment.
— Anish AcharyaCompanies will become a series of nested loops
Anish's steel man: after prompt comes agent (a loop of model plus tools, memory, skill files), and after agent comes loop (a group of agents doing a task). Coding got there first because the models are best, the users are the most technical, and a mature loop already existed — a bug report comes in, a reproduction is generated, it is fixed, high-risk changes get human confirmation, low-risk ones ship directly, and the customer may even get an automatic reply. Above that there is loop per person, loop per job function, loops across business units, up to a loop that can run most of the company.
— Anish AcharyaA loop only climbs to the local optimum and stops
The growth team's version of the loop: every variant is generated automatically, measured automatically, and once it reaches statistical significance it converges and ships, with a long-term holdout left behind, and then the next experiment begins. But Anish stresses that a loop hits the local maximum and then plateaus, and at that point you need out of distribution thinking and human intuition to land you at the foot of the next mountain. His example is telling Claude "make me a million dollars, no mistakes allowed" — it cannot do it, because the direction has to be set by a person.
— Anish AcharyaFrontier model pricing is actually irrational
Anish explains model choice through Pareto efficiency: frontier models are "irrationally priced" on the price-performance curve — one more IQ point can cost 100 times more than Opus. But for industries with unbounded upside like drug discovery, that one extra IQ point may be worth a trillion-dollar outcome, so paying that price is rational; for work with bounded upside like law and finance, you just want Pareto efficiency, using a cheap open-weight model that is good enough. The conclusion is that the two architectures will coexist, not one or the other.
— Anish AcharyaPeople want to spend time, not save it
Anish says "loop, make me happier" is the big opportunity: people say they want to be more efficient, but what they actually want is to spend time — the biggest products in the world are entertainment and social, not productivity tools. He distinguishes two kinds of users: the AI users on X fear falling into a permanent underclass and argue endlessly over whether GLM or Kimi is stronger; the Instagram user just thinks "this is a better Google search". He thinks this is not a model or capability problem, it is a product design problem: for the past 40 years we built technology that extends the intellect, and nothing that extends the soul.
— Anish AcharyaConsumer AI is still stuck in iPhone 2010
Anish likens the present to iPhone 2010 — Airbnb, WhatsApp and Uber had not appeared yet. Three things hold consumer back: models are too expensive, so free products are hard; the interface problem, since chat only suits "the highest agency people in the world" like Elon and Sam, while the ideal interface for an ordinary consumer is TikTok, so you need to find something between chat and TikTok; and the technology has been too skewed toward productivity rather than human connection and entertainment. He thinks all three are now loosening: open-weight models have brought costs down sharply, founders like Eugenia are thinking seriously about interface, and Brian Chesky has also set up a lab focused on the next generation of interfaces.
— Anish AcharyaPeople are at their best when the stakes are high, worst when they are low
One of Anish's theories: when the stakes are high we are great, when the stakes are low we are in our absolute worst state. The world five years ago felt like a low-stakes world, so society collectively took on a lot of side projects that neither produced anything nor made people happier. Now everyone feels like they are climbing the ladder of ambition. He cites GDP growth being stuck in a 2% rut for a long time, and asks who said we have to stay there, why not 10%, 15%, 20%. He also says AI unbundles skill from desire: you do not have to play piano to want to make music, you do not have to be a programmer to want to make software.
— Anish AcharyaMaking important things cheap is AI's best PR
Anish thinks the most important thing for changing AI's public image is making important things cheap. America has two extremely important things that keep getting more expensive: healthcare and education. In healthcare 45% is administrative cost, and removing the administrative burden would show healthcare cost deflation; education faces the strongest competition in 200 years, which will unbundle learning from institutions and also unbundle identity from credentials — you do not need a Harvard degree. He also offers an observation: revealed preferences show individual experience is getting better, stated preferences show people think society is getting worse. Ask whether to build a data center near your home and most say no; ask whether you use ChatGPT today and most say yes.
— Anish Acharya"Too dangerous to release" mixes marketing and compute
On OpenAI slowing RL training, and on the claim that "the model is too dangerous so we will not release it", Anish is skeptical. He thinks the aura Anthropic gains from "the model is too dangerous to release" is enormous, but behind it there may be confounding factors like a GPU shortage, capability not actually being there, or simply wanting to keep this proprietary model in-house to extend the lead. He says the concept mixes marketing, inference compute and economic considerations — are you trying to externalise your competitive advantage, or use it to make yourself stronger. He takes offensive cyberattacks seriously: before making systems easy to break into, you should harden all the systems first.
— Anish AcharyaMoats are discovered, not designed
Anish quotes Jesse of Decagon: most moats are discovered, not designed. He made this mistake himself as a founder, insisting on forcing out a business plan that could withstand scrutiny from MBAs and BCG. Cursor is the example: early on it was criticised for having no moat, but a high-NPS DAU product is itself good, and later they accumulated all the reasoning traces and trained their own Composer model. He also says not one classic moat is built on "how hard the software is to make" — network effects, scale advantages, brand, proprietary data, the thing history calls a cornered resource; the moats of five years ago are still moats today.
— Anish AcharyaThere is no growth problem, only a product problem
Anish says nobody has a growth problem now, only a product problem. The reason is that you can build a product that is extremely ambitious in any direction, functional or emotional, and you can charge a lot for it. His challenge is: is it really a growth problem, or a collective failure of our imagination? If you imagine your product charging a thousand dollars a month, ten thousand dollars a month, if it is the Birkin bag of software, what would it have to do to deserve that price? He also says it is now easier rather than harder for startups: despite Google's heavy cross-selling, nobody would say Gemini won; startups can build in directions incumbents are uncomfortable with, like companions; prices can be high, people will pay 200 dollars a month, and enterprises will sign million-dollar ACV contracts without being quite sure what they are getting.
— Anish AcharyaThe switch that makes AI change you is joy
Anish says he has observed that people's attitude toward AI flips, and the trigger is often not an efficiency gain but some moment of joy AI created for them. His example is a Mother's Day idea — first ask yourself "if this worked, what would make me happy", or "what can I do for someone else". He treats this as an actionable starting point: do not start from technical capability and look for a use case, start from the concrete thing that brings joy.
— Anish AcharyaThe three books he recommends are all about the same thing
Anish's three most-recommended books: Thomas Sowell's Conquest and Cultures, on how conquest changed cultures everywhere, which he considers the best historical lens on "culture is the biggest driver of outcomes", and which also maps onto the ambition-culture gap he felt after moving from Canada to the US; Seven Powers, on moats and compounding advantages as business theory; and the textbook Mark recommended to him, Increasing Returns to Scale (by Brian Arthur), a slightly dense study explaining why things like software have extreme economic effects. He says these three helped him understand "why our industry works the way it does".
— Anish AcharyaHe used a Grokbot to grab IMAX tickets and got stuck on a captcha
Anish says he kept failing to get IMAX tickets for Odyssey, so he had a Grokbot keep watching the ticketing site to find him good seats. But the captcha on the AMC site blocked it — you have to click three matching shapes. Lenny asks whether it is the pick-the-fruit kind, and Anish says yes, that kind, "come on, you can't do this?" This is the most concrete agent deployment case in the whole episode: a real task, a real point of failure.
— Anish AcharyaGrokbot made him doubt that models are a two-horse race
Asked for his favourite AI product, Anish picks Grokbot, because it "so unguardedly" caches credentials and completes tasks for you — he says he expects startups to do this, but did not expect a big company to do it. He thinks it has a thoughtful UI and a strong foundation model, and from that draws a judgment: maybe the model side is not a two-horse race.
— Anish AcharyaDo not build a product and a platform at the same time
Anish's life creed comes from his founder experience: do not discover through painful firsthand experience what someone could have told you in one sentence. The version for his kids is "do not touch the hot stove", and the version for founders is "do not build a product and a platform at the same time". His first company wanted to be both a mobile-game social platform and a game studio, and someone told him to his face: building a studio is already hard enough, and you are doing a platform at the same time, pick one. They did not listen, and it took them several years to figure out they were wrong.
— Anish AcharyaTape was the real watershed in music history
Anish has DJed for 31 years (since 95), and he thinks DJ plus music models is a new form of expression: you supply the idea, the model does the heavy lifting. He gives a history of music media — at first you could only hear live performance, then there were recordings on the phonograph, and the real turning point was tape, because that was the first time a person could "create" on their own and sequence their own album. He thinks the music industry's troubles in the 2000s came precisely from that disappearing and reverting to a broadcast model; now that people are making music again, the music industry will be bigger than ever.
— Anish AcharyaIn their own words · checked verbatim
It kind of unbundles skill from desire. Not only can we dramatically drive productivity, we can dramatically drive ambition.
Anish Acharya0:00
if you ask the most sophisticated individuals at the labs, it's not actually RSI that's occurring, which could lead to some sort of runaway winner because they were an epsilon ahead of the others. It's autocatalytic effects, which just means you're using the new technology to improve your process, but it's not truly recursive.
Anish Acharya6:10
the loop will help you climb to the local maxima, but then it plateaus. And you need some sort of out of distribution thinking. You need human intuition. You need somebody to actually help you land at the base of the next hill.
Anish Acharya14:18
And one of my theories, Lenny, is that like when the stakes are high, we are awesome. When the stakes are low, we are at our absolute worst.
Anish Acharya37:58
moats are most often discovered, not designed
Anish Acharya53:29
I always say that nobody has a growth problem these days. They have a product problem.
Anish Acharya59:49
I think it also is like, okay, wait, maybe this is not a two horse race on the sort of model side.
Anish Acharya1:13:30
don't discover things through painful experience that somebody can just tell you
Anish Acharya1:14:31
The big change in music actually from a medium perspective was the cassette tape because the cassette tape was really the first time you could create, right?
Anish Acharya1:15:32
Figures
| Time from the invention of electricity to factory reorganisation | 40 years | 8:16 |
| Length of the Kavak Jedi Academy course | six weeks | 8:16 |
| Price multiple for each extra IQ point on a frontier model | 100x (relative to Opus) | 21:32 |
| Share of healthcare spending that is administrative cost | 45% | 40:00 |
| Long-run GDP growth | 2% | 37:58 |
| Consumer subscription price | $200/month | 59:49 |
| Years DJing | 31 years (since 95) | 1:15:32 |
Glossary
- autocatalytic effects
- Using new technology to improve your own process, but not truly recursive self-improvement.
- RSI
- The hypothesis that a model can improve itself, producing a runaway winner.
- out of distribution thinking
- Thinking that steps outside the existing data distribution to find the next mountain.
- cornered resource
- A proprietary resource that history counts as one of the moats.
- revealed preferences
- Preferences inferred from actual behaviour rather than from what people say.
- stated preferences
- The preferences people voice, which may not match their actual behaviour.
How to listen
Founders and investors watching agent deployment, model pricing and consumer AI, especially anyone designing a product loop or thinking about moats.
The book recommendations at 1:11:20 and the music history at 1:15:32 can be skipped.