Darwin AI · · 424 words · 2 min
Build for a world where every customer gets their own software
If AI keeps improving exponentially, writing software gets so cheap that we stop reading code, then we stop reading the prompts that produce code, and we end up working in something higher up — natural language. Karpathy said English will be the dominant programming language. Juan, our head of data at Darwin, repeated that to me, and I think he’s right.
Once you accept that, you have to think hard about what your B2B software business is.
Try this exercise: take all the source code of any SaaS company — Calendly, Zendesk, Salesforce — and feed it to an AI with one instruction: describe in plain text what this software does. Don’t tell me about the code or the infrastructure. Just describe the goals, the features, the user flows. Now give that text to a new AI and ask it to build the whole thing from scratch. Ezequiel, my Darwin co-founder, and I do this regularly. The result isn’t perfect today, but it gets closer every month.
In a world where that works, your code is worth nothing. The intermediate description layer — the prompt, the spec — is the actual artifact. And it can be regenerated daily, hourly, per customer, even per session.
That means every customer can get their own software. Why give a generic Salesforce to every company when each company only uses 10% of the features and is overwhelmed by the rest? Generate the version they need. Eventually generate the version each user needs, in real time.
This sounds wasteful — like running a calculator app on an iPhone instead of a $5 calculator. It is. But computation got so cheap we do it anyway. AI inference will get there too.
So if you’re building B2B software now: don’t fight this. Build for it. Your moat won’t be the code, because the code is regenerable. Your moat will be:
- Owning the customer relationship and the responsibility for the outcome.
- Owning the dynamic software layer that regenerates per customer.
- The data and the QA layer that makes the regenerated software actually work in production.
The customer you sell to — say, a pharma company — doesn’t want to become an expert in dynamic AI software. They want it to work. So someone runs that. That someone is you, if you build for it.
The transition will be brutal. Most current SaaS will be repriced or rebuilt. But the new wave is going to produce companies bigger than the ones it replaces. Build for the wave, not against it.