The Human Architecture of AI

Every AI model in existence was built on human labor. Annotators label images, rate model outputs, moderate harmful content, transcribe audio, and provide the feedback that teaches AI systems what "correct" looks like. Without this work, no model exists. The people doing it are nearly invisible to the public that benefits from it.

• Estimated 1–4 million contract workers globally — Kenya, Philippines, India, Venezuela, and elsewhere
• Typical pay: $1–$6/hour, often with no benefits, no job security, and no advancement path
• Many label graphic content — violence, abuse, self-harm — to train content moderation systems, with no psychological support
• Work is tracked to the second; productivity monitored continuously
• Kenyan workers at Sama who trained ChatGPT's safety filters organized and struck in 2023 — one of the first labor actions in AI
• The global data annotation market is projected to grow from $2.3B in 2025 to nearly $10B by 2030

Data engineers build and maintain the pipelines that collect, clean, and format the training data that annotators label. They are the plumbing of AI — rarely celebrated, essential to everything.

• Design systems to collect data at scale from web scraping, APIs, and proprietary sources
• Clean and standardize data — removing duplicates, fixing formatting, filtering unusable content
• Build the storage and retrieval systems that make training datasets usable
• Maintain data quality as datasets grow to billions or trillions of tokens

Machine learning engineers implement, train, and deploy the models that researchers design. Software engineers build the products that end users interact with. This is the largest professionally-compensated tier — tens of thousands of people at major tech companies and AI startups.

• Write the code that implements model architectures, training loops, and inference systems
• Optimize models for speed and cost at deployment scale
• Build the applications — chatbots, APIs, consumer products — that bring AI to users
• Total compensation includes significant equity at frontier labs — actual cash salary is a minority of earnings

The rarest tier. Roughly 5,000–10,000 people globally have genuine frontier impact — designing new architectures, developing training methods, and advancing the state of what AI can do. This is the group that produced the Transformer, AlphaFold, and RLHF. Their work sets the direction for everything above and below them.

• Publish research that shapes the entire field — a single paper (Attention Is All You Need, 2017) changed everything
• Heavily concentrated at a handful of labs: Google DeepMind, OpenAI, Anthropic, Meta AI, and a few universities
• Compensation reflects scarcity — bidding wars between labs are common; signing bonuses in the millions
• Also includes AI safety researchers — a small but growing subset focused on preventing harm from advanced AI

The people who decide what gets built, how it's deployed, and how the organization engages with governments and the public. Safety tradeoffs happen here in practice — not in papers. Policy positions that shape regulation originate here. This tier has enormous influence over how AI affects the world.

• Product leadership decides which capabilities ship and which are held back
• Policy teams engage with Congress, the EU, and international bodies — shaping the rules the industry will operate under
• Safety teams nominally oversee risk — but report to leadership with commercial incentives
• This is where the tension between "move fast" and "be careful" is resolved in practice, every day

The founders and CEOs of frontier AI labs set vision, culture, and risk tolerance for the entire enterprise. Their personal beliefs about AI's potential and its risks shape decisions affecting billions of people. A small number of individuals hold disproportionate influence over the direction of one of the most consequential technologies in human history.

• Sam Altman — CEO, OpenAI. Primary public face of the generative AI era.
• Dario & Daniela Amodei — Co-founders, Anthropic. Former OpenAI leadership who left over safety concerns.
• Demis Hassabis — CEO, Google DeepMind. AlphaFold and AlphaGo originator. Nobel Prize in Chemistry 2024.
• Yann LeCun — Chief AI Scientist, Meta. Turing Award winner. Skeptic of current path to AGI.
• Their compensation is primarily equity — if the company succeeds, their wealth is generational.

At the apex sit the entities whose capital makes the infrastructure possible. They do not write code, label data, or run experiments. They fund the operations that require billions of dollars in compute before any revenue exists. In return, they own the upside. Their influence shapes which labs survive and which directions the field takes.

• Microsoft — ~$13B invested in OpenAI. Azure cloud infrastructure runs many frontier models.
• Google/Alphabet — Owns DeepMind outright. Invested ~$2B in Anthropic. Runs Gemini internally.
• Amazon/AWS — Committed up to $4B to Anthropic. Cloud infrastructure for much of the industry.
• Meta — Owns FAIR (Fundamental AI Research) outright. Released LLaMA open-source to compete differently.
• Venture capital — Sequoia, a16z, Khosla Ventures among the largest backers of AI startups.
• Sovereign wealth — Saudi Arabia (PIF) and UAE (MGX) have committed tens of billions to AI infrastructure globally.
• NVIDIA — Structurally most powerful of all. Supplies GPU chips to all frontier labs simultaneously and profits regardless of which model wins the race. Market cap exceeded $3 trillion in 2024.