Stop Worrying About the AI Underclass - 5 Surprising Truths About the Future of Product

Analysis of AI Evolution: From Productivity to Human-Centricity and Autonomous Loops

Executive Summary

The current technological landscape, driven by Artificial Intelligence, is characterized not by a “fast takeoff” toward super-intelligence, but by a “slow takeoff” that emphasizes human agency and the unbundling of skill from desire. Anish Acharya, General Partner at Andreessen Horowitz, argues that the prevailing fear of a “permanent underclass” is a “dark fantasy” that contradicts empirical data and the decentralized nature of the current tech era.

The central shift in company building is the transition toward a series of loops. In this model, agents perform repetitive, verifiable tasks in autonomous cycles, while humans provide the “out-of-distribution” thinking necessary to move from a local maximum to the next opportunity. For product builders, the opportunity lies in moving beyond productivity-centric tools to address “spiritual hunger”—using AI to enhance human connection, happiness, and creative ambition. Success in this era is defined by extreme ambition, the abandonment of the “preciousness” of intelligence, and a commitment to frequent, iterative shipping.

Debunking the “Permanent Underclass” and the Fast Takeoff Myth

Despite Silicon Valley’s collective anxiety regarding AI-driven displacement, empirical evidence suggests a more optimistic trajectory.

Decentralization vs. Centralization: Unlike the mobile era, which was dominated by N-of-1 network effects and centralization, the AI stack is highly fragmented. There are dozens of relevant players at every level, from foundation model labs to open-weight variations and coding agents.
Empirical Job Data: Historically “threatened” professions, such as radiologists and programmers, currently see higher job postings than ever before.
Autocatalytic vs. Recursive Self-Improvement (RSI): Acharya distinguishes between RSI (runaway intelligence) and autocatalytic effects. The industry is currently seeing the latter: using technology to improve processes without achieving true, autonomous recursive growth.
Intelligence-Bound Problems: Many industries (e.g., logistics like FedEx or food service like Domino’s) are not primarily bound by intelligence but by physical or supply-chain constraints. Consequently, a “data center of PhDs” would not necessarily lead to exponential market domination.

The Architectural Shift: Companies as a Series of Loops

Company building is evolving into a cascading set of autonomous loops. This architecture moves from individual tasks to entire business units.

The Loop Hierarchy

Agent Loop: A model combined with tools, memory, and skill files to perform a specific task.
Functional Loop: A set of agents handling a domain (e.g., the coding loop: bug report -> reproduction -> fix -> review -> deployment).
Business Loop: The output of various functional loops (marketing, sales, legal) synthesized to inform strategy or business model changes.

The Growth Team Example

In a traditional growth team, humans prioritize and build experiments. In a loop-native company:

Every variant is automatically generated and measured.
Systems converge on variants once statistical significance is reached.
The system maintains long-term holdouts and moves to the next experiment autonomously.

The Local Maxima and Human Intuition

AI agents are exceptional at “hill climbing”—optimizing a process until it reaches a plateau (local maximum). However, agents lack “out-of-distribution” thinking.

The Human Role: Humans are required to identify the “base of the next hill.”
Exception Handling: Humans manage strategy, sales, and complex exceptions where models fail due to data or knowledge gaps.

Model Strategic Categorization

The industry is seeing a split between frontier models and open-weight models based on the “upside” of the problem being solved.

Problem Category	Model Type Recommended	Examples
Unbounded Upside	Frontier / High-Intelligence	Drug discovery, high-stakes research, engineering.
Bounded Upside	Open-Weight / Mid-IQ	Closing financial books, legal compliance, routine admin.

Acharya suggests that for “verifiable” problems with limited upside, paying for frontier intelligence is “wasteful.” However, for problems where a single insight can lead to a trillion-dollar outcome (e.g., drug discovery), the cost of frontier tokens is rational.

The “Model Sommelier” Perspective

Models are no longer commodities; they have distinct “shapes” of intelligence:

Quen 2.5 Max: Creative, excellent at long-horizon tasks and storytelling.
GLM-4: Described as a “neurotic PhD”—precise and focused on technical/product work.

The Consumer Opportunity: “Make Me Happier”

Acharya posits that the industry has spent 40 years building technology to enable better spreadsheets, neglecting “the soul.”

Productivity vs. Time-Spending: While VCs often fund productivity (saving time), most consumers want ways to spend time meaningfully (entertainment, social connection).
The Spiritual Hunger: With the decline of traditional cultural institutions, there is an opportunity for AI to address the “basics of consumer need”: feeling connected, loved, and fulfilled.
Interface Evolution: The current “chat” interface is ideal for high-agency individuals (e.g., Sam Altman, Elon Musk), but the average consumer requires something between Chat and TikTok.

Strategic Lessons for Founders and Product Builders

Ambition and Idea Scale

In the current environment, an idea that is “too small” is a greater risk than an idea that is “too ambitious.”

The “Birkin Bag” Exercise: Builders should ask, “What would the $10,000/month version of this product look like?” to push the boundaries of their imagination.
Unbundling Skill from Desire: AI allows individuals to be “creatively ambitious” without needing technical mastery (e.g., making music without playing piano).

Moats and Durability

Moats are Discovered, Not Designed: Many successful companies (e.g., Cursor, Decagon) find their moats through shipping and capturing reasoning traces rather than through initial business plans.
Classic Moats Still Apply: Network effects, brand, and scale advantages remain the gold standard.
The Craft as a Moat: High-quality user experience and “craft” (as seen in products like Granola) provide a temporary moat that allows for the discovery of more durable advantages.

The Habit of Shipping

Product people are encouraged to “build to learn” rather than just to produce an outcome.

The Chassis Method: Maintain a “chassis” (a core app or project) to test every new model that is released.
The “Joy Moment”: Successful AI adoption often follows a moment of personal joy (e.g., using AI to create a Mother’s Day presentation from old texts).
Frequency: Aim to ship something—however small—once a week to build intuition.

Notable Insights and Quotes

On Automation: “The white-collar manager… has this abstraction that everybody else’s job is super automatable by AI, but of course, ours is not.”
On Economic Progress: “We have been in this morass of 2% GDP growth… why can’t we be 10 or 15 or 20%? This is a technology with which we can dramatically drive productivity and ambition.”
On Human Nature: “The entire trend of human existence has been that our desires grow faster than our ability to fulfill them.”
On the Current Era: “It feels like Christmas 2009 where everybody got their iPhone and wants to download new apps… the windows are open for now.”
On the Future of Software: “Intelligence is no longer precious. That will be a hard change for product people to accept.”

Sources

[Why jobs are becoming a series of loops Anish Acharya (a16z)](https://www.youtube.com/watch?v=LdIyXiq2DTY)