The Latent Space Essay

The Rise of the AI Engineer

A new engineering discipline is forming — distinct from ML research, trained on APIs not PhDs, and poised to outnumber traditional ML engineers within five years.

swyx · latent.space · June 30, 2023 Career + Industry
"When it comes to shipping AI products, you want engineers, not researchers."
The core claim — Latent.Space
5K+
LLM researchers globally
50M+
Software engineers globally
1–10K×
Faster prototyping with APIs vs. training
Definition

What Is an AI Engineer?

An AI Engineer applies existing foundation models — GPT-4, Claude, open-source LLMs — to build practical products. They don't train models from scratch. They don't need a PhD. What they need is software engineering fundamentals, product intuition, and the ability to rapidly iterate using existing AI capabilities via APIs and toolchains.

Dimension AI Engineer ML Engineer / Researcher
Core work Build products with foundation models Train and fine-tune models
Credential No PhD required PhD typical for research roles
Primary tool APIs, orchestration, prompts PyTorch, CUDA, data pipelines
Iteration speed Days to weeks (fire, ready, aim) Months to years (data-first)
Focus Generative: writing, coding, learning tools Classification: fraud, recommendation
Supply Emerging — fast-growing Scarce — ~5,000 LLM researchers
Context

Why This Role Is Emerging Now

Six forces have converged to make the AI Engineer role both viable and necessary — none of which existed three years ago.

1
Foundation models have surprising generalization
Models exhibit unexpected zero-shot transfer and in-context learning that even their creators don't fully understand. This creates an entire class of applications that can be discovered by curious engineers — no training required.
▾
2
Research talent is scarce by design
~5,000 LLM researchers worldwide versus 50 million software engineers. This supply constraint means AI product-building must be accessible to engineers who didn't train models. APIs make this possible.
▾
3
GPU economics favor API access
Hardware is concentrated in large labs. Startups and individuals cannot afford to train frontier models. The inevitable result: build with APIs or don't build. AI Engineers are the practitioners who work within these economics.
▾
4
Agile beats waterfall for LLM products
"Fire, ready, aim" prototyping with LLMs enables 1,000–10,000× faster validation than traditional ML's laborious data collection → training → evaluation cycle. Speed-to-signal is a structural advantage.
▾
5
JavaScript expands the addressable talent pool
LangChain.js, Transformers.js, and Vercel AI SDK expanded the ecosystem beyond Python-only. This doubled+ the TAM of practitioners who can build AI products — from Python-native to the full web development community.
▾
6
Generative products are a new category
AI Engineers build writing assistants, coding tools, and personalized learning products — distinct from traditional ML's fraud detection and recommendation systems. New category, new practitioners.
▾
Skills

What You Actually Need

Conventional wisdom says start with Ng's ML courses, PyTorch, and transformer architecture. Successful AI Engineers like Simon Willison and Riley Goodside shipped products without any of this foundation.

✓
Software engineering fundamentals
APIs, debugging, data structures, version control — core craft transfers completely
✓
API integration experience
If you can call a REST API and handle responses, you can build AI products today
✓
Product thinking
Understanding user needs and what makes AI useful — often more valuable than technical depth
✓
Rapid prototyping ability
The ability to go from idea to working demo in days, not months
✓
Understanding LLM capabilities and limits
Practical intuition about what models can and can't reliably do — from spending time with them
✕
PhD in any field
Top AI Engineer compensation ($300K–$900K) is achieved without academic credentials
✕
PyTorch / CUDA knowledge
You're building on top of models, not training them — training frameworks are irrelevant
✕
Deep learning fundamentals
Backpropagation, gradient descent, and architecture math rarely encountered in product work
✕
"Attention Is All You Need" comprehension
You use the outputs of this research, not the mathematics — a user doesn't need to understand the engine
✕
Data warehouse architecture
Traditional ML data infrastructure skills don't transfer to API-first product development
Compensation

The Market Has Decided

Sheer demand-and-supply economics have already set the price. The market doesn't wait for credentials.

$300K
Prompt Engineering
at Anthropic
$900K
AI Software Building
at OpenAI
Job Title Trajectory — Indeed.com (2023 baseline)
ML Engineer 10× more jobs (currently)
AI Engineer Growing faster — predicted to flip within 5 years
Andrej Karpathy: "There's probably going to be significantly more AI Engineers than ML Engineers. One can be quite successful in this role without ever training anything."
Framing

Software 1.0 → 2.0 → 3.0

Extending Andrej Karpathy's original Software 2.0 framing — English is becoming the hottest new programming language.

1.0
Classical Code
Explicit logic hand-written by engineers. Every behavior is deterministic and manually specified. The dominant paradigm for 60+ years of software.
if / else / while
2.0
Neural Networks
Karpathy's original framing: instead of writing logic, you write training data and let gradient descent find the weights. Behavior is learned, not programmed.
PyTorch · TensorFlow · data pipelines
3.0
English as Code
AI Engineers orchestrate foundation models with prompts, context, and tool use. The "code" is natural language. The intelligence is pre-trained. The engineer's job: compose capabilities into products.
LangChain · vector DBs · agent loops
"English is the hottest new programming language."
— Latent.Space, The Rise of the AI Engineer
Stack

The AI Engineer's Toolbox

Orchestration & Chaining
LangChain ($200M+ funded) LlamaIndex Guardrails Vercel AI SDK
Foundation Models
GPT-4 Claude LLaMA (open source) HuggingFace
Agents
Auto-GPT BabyAGI Voyager (Nvidia)
Infrastructure
Pinecone (vector DB) Transformers.js LangChain.js
Career Advice

How to Become One

Don't start with academic papers. Don't take ML prerequisites. The practitioners who ship products follow a different path entirely.

  1. Build with existing APIs immediately
    Take a problem you care about and make it work with GPT-4 or Claude. Don't study first — ship something, however rough.
  2. Spend time with models to discover their capabilities
    Zero-shot and few-shot capabilities are often surprising even to researchers. Exploration reveals what's possible — documentation doesn't capture it fully.
  3. Learn tools through application, not tutorials
    Pick up LangChain, a vector database, or an agent framework because your project needs it. Context-driven learning is faster than sequential study.
  4. Iterate rapidly — the "fire, ready, aim" workflow
    LLMs reward fast experimentation. Validate ideas in days, not months. Traditional ML's requirement for large labeled datasets before any validation is inverted.
  5. Join the community and show your work
    The AI Engineer Summit (ai.engineer) and Latent.Space represent a community forming around this identity — the connective tissue for the emerging discipline.