Why jobs are becoming a series of loops | Anish Acharya (a16z)

Why does it matter? Because every AI loop you deploy will eventually stall — and what you do at that plateau defines the next decade

Anish Acharya, general partner at a16z and one of the sharpest consumer-product thinkers in venture, thinks most of us are asking the wrong questions about AI. Not whether jobs survive, but where the loops break. Not how to save time, but why the biggest markets are built on wasting it.

• AI agent loops optimize to a local maximum and then plateau — humans are the only mechanism for pointing at the next hill
• The permanent lower class fear is empirically false; programmer job vacancies are at all-time highs, and recursive self-improvement is not what is actually happening at frontier labs
• Consumer AI's mass market is not productivity — entertainment and social connection represent the larger, nearly unbuilt opportunity
• The VC rejection reason has inverted: three years ago, too-ambitious ideas did not get funded; today, ideas that are too small do not get meetings

Every AI loop climbs to a local maximum — then stalls, and only human intuition can find the next hill

Agent loops, run without a human steering layer, produce a very efficient dead end.

Acharya describes the emerging architecture of AI-native companies as a cascade — from a loop per person, to loops running individual business functions, to loops managing large portions of the company. The programming function is already there: incoming error report, generated reproduction, fix created, fix checked, low-risk changes pushed automatically, high-risk ones flagged for human sign-off. The same structure will reach marketing, sales, legal, and eventually the whole-company optimization cycle.

But every one of those loops eventually flatlines. An A/B testing loop generates every variant, measures every result, reaches statistical significance — and then the metrics plateau. The loop cannot look at the landscape and decide the next mountain is worth climbing. "The cycle will help you to climb to local maximum, but then it goes out onto the plateau. You need human intuition. You need someone who will actually help you land at the foot of the next hill."

Kavak, the Mexican used-car platform, has already built this model precisely. Each customer has a dedicated agent; when the agent gets stuck, it calls a human. That human resolves the stall, the agent records the interaction, and the system learns. The human is not the fallback — they are the teacher. The stall point is not a flaw in the loop; it is the exact location of human leverage.

AI separates skill from desire — and want is now the only prerequisite to build

The constraint on what any person can build has shifted from capability to imagination.

"She seems to separate skills from desires," Acharya says. "If you want to create music, you can do it now. You do not have to be able to play the piano." Programming, bridge design, filmmaking — the same logic applies to anything that previously required years of craft before an idea could even be tested. Desire was always there. Skill was the filter. That filter is dissolving.

The GDP implication follows directly. Acharya sees no structural reason the U.S. economy should be stuck in the 2% growth range. "We can not only significantly increase productivity, but also significantly strengthen ambitions." He draws the analogy to the postwar era — a moment when collective stakes were high enough that whole economies mobilized in ways no one had thought possible. AI recreates those conditions, not by making people smarter, but by closing the gap between what they can envision and what they can execute.

The filter to apply before scoping any project has changed. Stop asking whether you have the skills to execute. Ask whether the desire is strong enough to direct the tools.

The biggest consumer AI market is not productivity — most people want to waste time, and the largest products in history are built on exactly that

Forty years of computing built better spreadsheets. Every tool that followed — email, calendars, task management — treated human attention as something to be optimized. The actual evidence of what people do with their time points somewhere else entirely.

"The largest products in the world are entertainment and social networks," Acharya says flatly. Not productivity tools. Not time-savers. The applications people spend the most hours inside are the ones that make them feel something — connected, entertained, less alone.

The consumer AI market that excites him is the loop that asks: how do I feel more related, more loved? How do I develop? How do I have fun? He frames it as a design problem, not a model problem. The technology is capable. The products do not exist yet because founders keep building for the wrong user. The X AI user is fixated on model benchmarks and afraid of falling behind. The Instagram AI user just wants something that improves on their current scroll.

The structural analogy is 2010 for the iPhone. The infrastructure arrived with the original device; Airbnb, WhatsApp, and Uber did not exist yet. "We spent 40 years creating technologies which expand our intelligence, but nothing which would expand our soul." That gap is where the mass market is.

The permanent lower class is a strangely gloomy Silicon Valley fantasy — every empirical indicator contradicts it

Radiologists were supposed to be automated away twenty years ago. "The number of job vacancies for programmers is higher than ever," Acharya notes. The empirical record does not support the displacement narrative, and he is not gentle about saying so: "It is strangely gloomy, a fantasy which seems to prevail in Silicon Valley collectively. In fact, things have never gone better on almost every indicator."

His technical clarification cuts deeper than the economic point. What frontier lab researchers actually describe is not recursive self-improvement — the scenario where a model vaults irreversibly ahead of human control. "It is autocatalytic effects, which simply means using new technologies for process improvement, but it is not real recursion." Autocatalysis compounds quickly; it does not produce a runaway gap.

He also challenges an unexamined assumption embedded in the fear: that most organizational problems are intelligence-constrained. "If you had an entire data center filled with PhDs working at FedEx or Domino's Pizza, would they become exponentially better at logistics or pizza?" He doubts it. A lot of organizational work is blocked by factors that more intelligence does not change. Redirect the anxiety from whether you will survive AI to how you use it to do things you previously could not — the data says the threat is largely imagined.

Frontier model pricing is only rational when upside is genuinely uncapped — most organizational functions should run on something far cheaper

One conceptual IQ point costs a hundred times more on a frontier model than on something like Opus 4.8. That price is irrational for most tasks — but not all.

Drug discovery has uncapped upside. If a single additional unit of intelligence allows a model to identify the next statin, the economics of frontier pricing are sound. Closing the quarterly books does not carry that property. "We crossed the threshold of intelligence for almost each economically useful task. And everything that comes out of the limits of this threshold is simply wastefulness."

Acharya predicts a structural split inside AI-native companies. Functions with bounded upside — accounting closes, routine legal review — will route to Pareto-efficient open-weight models tuned with reinforcement learning. Functions with genuinely uncapped potential — a support call that might surface a company-changing insight, sales conversations, fundamental research, engineering — justify frontier token costs. The practical question before picking a model for any function is mapping its actual upside ceiling. Only assign frontier tokens where a single extra unit of intelligence could change the magnitude of the outcome.

Three years ago, too-ambitious founders could not get funded — today, ideas that are too small do not get meetings

The VC filter has fully inverted.

"In the old days, three years ago, we saw a company, and if what they tried to do was too ambitious, we did not get involved." The funding logic was consistent: a $100M seed round made no sense because no team could responsibly manage that capital, and no problem benefited from it. Conservative scope and a clear wedge were the entry ticket.

Today Acharya describes "almost the opposite problem." An idea that is too small does not produce the desire to engage. A16Z signals this in what they fund and in what they say to founders at the moment of investment: we are here to help you build the strongest version of your vision, full stop.

The implication for founders is specific and uncomfortable. Pressure-test your pitch in the opposite direction from habit. The question is no longer whether your wedge is defensible enough at the beachhead. It is whether your vision is large enough to justify the meeting.

Competitive moats are discovered through use, not architected upfront — and the classic ones still work; founders just need the ambition to pursue them

Cursor was criticized early for lacking a moat. It now has deep workflow integration, daily active use, and compounding retention — none of which appeared in a pitch deck. It emerged from building.

Acharya quotes Decagon's Jesse directly: "Competitive advantages are most often discovered, not designed." Founders who invest energy in moat storytelling at the pitch stage are spending it on the wrong thing. Founders who underbuild the actual moat mechanics — multiplayer features, data flywheels, social loops — pay for it later.

The underlying framework has not changed: network effects, scale benefits, brand, unique data, proprietary access. Hamilton Helmer's seven powers remain the map. "We just need founders who have ambitions in these directions. We need more multiplayer products. We need consumer social networks. We need products that become significantly better the more you use them." The moat is real — it only reveals itself in usage data. Ship fast enough to find it, then double down on reinforcing it. Acharya's tell: he now listens more to what customers say than to what founders write in business plans.

The skill floor just collapsed — what survives is the ambition ceiling

The consistent thread across loops, model tiers, VC filters, and consumer products is that the constraints have moved. Skill is no longer the bottleneck; desire is. The companies that win will not be the ones with the most technical horsepower. They will be the ones that asked a bigger question and refused to shrink it. The loop will optimize everything it can see. Someone has to decide what is worth seeing.

---

Topics: AI agents, agentic loops, consumer AI, model selection, startup strategy, ambition, competitive moats, productivity, happiness, VC investing, future of work, Anish Acharya, a16z