Making money in the software 3.0 era

The old playbook — courses, agencies, apps, services — is being dismantled by AI. Here is an honest look at what's changing, what still works, and where the real opportunities are.

What works in 2026

Find real problems. Validate demand. Build useful solutions. Distribute them where people already spend time. Learn from real users. Refine quickly. Build trust.

The winners in 2026 will not necessarily be the people producing the most content or building the most features. They will be the people who understand their audience deeply, test ideas early, and become a trusted source of solutions for a specific group of people.

The Karpathy framework: the evolution of software

1.0

Explicit rules

You write every step. If the code doesn't cover it, it breaks.

2.0

Learned patterns

Machine learning improves on data and feedback over time.

3.0 — now

Prompted agents

Give a goal in plain language. An agent figures out how to achieve it.

What this means in practice

Installer

1.0

A bash script with every step hardcoded. Hits an unknown error — it stops.

2.0

Reports errors back. ML improves success rate over time.

3.0

Paste one line into your agent. It reads your environment and installs itself, debugging anything it hits.

Restaurant menu app (Karpathy's own)

1.0

Database of pre-matched food photos. No photo on file = no image shown.

2.0

Download the app. AI renders images in real time from menu text.

3.0

No app needed. Photo + one prompt to Gemini. The 2.0 app is instantly obsolete.

Online courses

1.0

Fixed lecture order. Follow steps 1–10 in sequence.

2.0

AI trims what students skip. Course adapts over time.

3.0

A personal agent coaches you live as you work — the expert in your ear while you act.

Video editing

1.0

Premiere Pro. Manual cuts, transitions, sound by a human editor.

2.0

Descript. AI spots silences and mistakes. You still operate the tool.

3.0

"Edit this in a viral Mr Beast style, under 8 minutes." Agent delivers a finished cut.

Where the opportunity is — four moats

Your data and knowledge

The information only you have — your experience, your style, your domain expertise. That's what nobody else can replicate.

Prompt engineering

The instructions and context you give the model determine the quality of output. Getting this right is a genuine skill that most people won't bother to develop properly.

The system you build around the engine

AI is the engine. You design the car. The wrapper, the workflow, and the interface can be the entire product.

Trust and audience

The relationship and credibility you've built is increasingly the most defensible thing you have. Anyone can spin up an AI tool — not anyone can be you.

The one-person AI business

The idea is simple: use your existing skills and knowledge as the offer, use AI to do most of the delivery, and spend your time building relationships rather than doing grunt work. High margins, no staff overhead, fully remote.

The biggest mistake people make is chasing whatever business model a YouTuber is promoting this week. Start from what you already know — your industry, your experience, your contacts. AI makes it possible to deliver on services you couldn't have scaled alone before.

Build an offer from existing skills

Don't pick a niche at random. What do you already know that solves a real problem for a specific type of business? That's your starting point.

Find prospects using signals, not broad categories

Don't target "all vitamin companies." Find signals — they're hiring for a role you can replace, they've just raised funding, their ads haven't changed in months. Specific signals mean warmer outreach.

Get on the call and ask questions

No fancy sales scripts needed. Ask about their problem, listen carefully, then show how you solve exactly what they described. The pain they tell you is your pitch.

Deliver with AI, charge for the outcome

AI handles 80–90% of the execution. You handle quality control and the relationship. The client pays for results, not for your hours.

Affiliate marketing — still works, but watch this space

Affiliate marketing — earning a commission by recommending products — remains one of the more accessible ways to build passive income online. The model is proven: create useful content, attract an audience, recommend relevant products, earn when they buy. More on affiliate marketing ↗

The LLM problem. A growing number of people now ask ChatGPT or Claude instead of Googling. When someone asks an AI "what's the best VPN for streaming?", they get an answer directly — no search results, no affiliate links, no commission for anyone. As AI assistants become the default starting point for product research, traditional SEO-driven affiliate pages will see declining traffic. This is already happening and will accelerate.

How to stay relevant:

Publish original experience, not summaries. AI can summarise specs. It can't replicate hands-on use. First-person testing, photos, and honest opinions are harder to replace.

Keep content genuinely current. Stale reviews get filtered out by both Google and AI citations. Update pricing, availability, and your verdict regularly.

Build an audience, not just pages. An email list or community means you're not entirely dependent on search traffic. People who follow you will come directly. Build one without the monthly SaaS bill ↗

Go niche and specific. Broad "best laptop" articles are dead. Highly specific content — "best laptop for architects running Revit on a budget" — still has a place because it serves intent that's hard to satisfy with a generic AI answer. The AI Niche Marketing Book

Diversify beyond SEO. YouTube, newsletters, and social content can all carry affiliate links and aren't as exposed to the search traffic squeeze.

About the author

Dave Dand

Software engineer and web developer building things on the internet since 1995. Projects include World Town Guide — a fully automated travel guide covering 127,000 destinations — and a UK business directory of 2.5 million pages that reached a top-50 Alexa ranking. Also called Bitcoin at dollar parity in 2011, which is either prescient or just good timing depending on who you ask.

Full profile ↗