How to Build AI Yourself

Building AI sounds like something only a handful of billion-dollar labs can do. Most of what actually separates them from one motivated person is five ordinary steps — not a secret only they know.

Prepared by Richstorm.co

Key Takeaways

• Building an AI-powered app or tool without writing a model yourself is realistic today, using existing AI as a component.

• Companies like Cursor and Harvey are built the same way — calling Claude or GPT through an API and adding a workflow around it, not inventing new AI.

• Fine-tuning an existing open AI model on your own data is achievable with basic coding skills, often for a few hundred to a few thousand dollars.

• Training a narrow, task-specific model from scratch is a real, achievable project for a well-resourced team — training a general-purpose model from scratch is the one door that stays closed to almost everyone.

• The hardest skill in building with AI isn't writing code — it's judging whether the AI's output is actually correct.

Four Levels, Very Different Doors

"Build an AI" gets used to describe wildly different projects. Some of them are open to anyone with a laptop and patience. Others require a research team and hundreds of millions of dollars. The confusion mostly comes from treating these as one thing.

The first two levels are genuinely open to a motivated non-specialist. The third is within reach with real effort. "Training from scratch" isn't one category — a narrow, task-specific model is a real project for a well-funded lab, while a general-purpose foundation model needs resources only a handful of frontier labs have, as the scorecard below breaks down.

The Skill That Matters

Here's the part that's easy to miss: even where building is realistic — fine-tuning, or wiring together an AI-powered tool — writing the code is no longer the hard part. Coding agents can now write a working training script or app in minutes. The harder, and more valuable, skill is judging whether the AI's output is actually correct.

A broken app is obvious — a button doesn't work, a page doesn't load. A poorly fine-tuned model is not obvious. It can run cleanly, produce fluent answers, and still have quietly learned the wrong thing — memorizing instead of generalizing, or picking up a bias buried in the training data. Catching that requires domain judgment, not programming skill.

This is genuinely good news for a non-specialist with deep knowledge in another field. The scarce skill in AI-assisted building is increasingly evaluation and judgment, not typing code — and that's a skill built from real-world expertise, not a computer science degree.

Real Companies at Each Level

These levels aren't just a theoretical ladder — real, named companies sit at each one right now, and the pattern holds up: almost everyone, including companies worth billions, operates at Level 2.

One nuance worth flagging: OpenAI's own life-sciences model, GPT-Rosalind, released in April 2026 and now used by Amgen and Moderna, sits between Levels 3 and 4 — it specializes an existing frontier model rather than building one from zero, but it's still a project only a frontier lab like OpenAI has the resources to undertake, not something an individual pharma company does in-house.

Roughly 78% of AI startups launched in 2024 are, by CB Insights' count, essentially API wrappers sitting at Level 2 — and the two companies that actually operate at Level 4, OpenAI and Anthropic, together capture about 89% of all AI startup revenue. Most of the AI industry, including companies worth billions, is built exactly the way the next section walks through.

How to Build a Cursor, Harvey, or Jasper

None of this requires inventing new AI. It requires wiring an existing model into one specific, painful workflow — which is exactly what Cursor and Harvey did above, and it's a project one motivated person can start this week.

There is no hidden AI breakthrough inside Cursor or Harvey. Both call the same kind of model anyone can access with an API key — the entire company is steps 1, 3, 4, and 5, built around step 2. Step 4 is the one Jasper skipped, and it's the one that decides whether a foundation model can absorb your product next quarter or can't touch what you've built.

The Scorecard

The honest version of "can a layman build AI" is: yes, for using AI, building with it, and fine-tuning it, and even for training a narrow model from scratch with the right team. The one door that stays shut is training a general-purpose foundation model — and that was never really about being a specialist. It's about being an organization with hundreds of millions of dollars and a research team, which is a different kind of barrier entirely.

Sources: Stanford AI Index Report 2025. Epoch AI, "The Rising Costs of Training Frontier AI Models" (May 2024).

CB Insights, AI startup wrapper data (2024). TechCrunch, "Google VP Warns That Two Types of AI Startups May Not Survive" (February 2026). Sanofi, Formation Bio, and OpenAI, joint press release (May 2024). CNBC, "Novo Nordisk Partners With OpenAI as AI Drug Discovery Hopes Mount" (April 2026). Bloomberg, "OpenAI Takes on Google With New AI Model Aimed at Drug Discovery" (April 2026). Oxford Academic, Briefings in Bioinformatics, "ADCNet: A Unified Framework for Predicting the Activity of Antibody-Drug Conjugates" (2025). Frontiers in Drug Discovery, "AI-Driven Innovation in Antibody-Drug Conjugate Design" (2025). Microsoft AI, "Building a Hill-Climbing Machine: Launching Seven New MAI Models" (June 2026). Anthropic customer case study, "Perplexity Delivers Factual and Relevant Answers With Claude."