Durable Advantages in the AI Era
I've always been a tech guy, and I thought one day I'd build and run a tech company – ideally in something cool and futuristic. We're finally getting the futuristic stuff, but I'm not sure the idea of a tech startup works anymore, all for kind of the same reason: AI.
This probably only applies to software. Hard-tech startups face different dynamics. I once thought I'd end up in space tech – I did a degree in Space Engineering – but ITAR and Australia's small industry made that impractical. That's a story for another time. The point is: AI seems like it's going to eat SaaS, and it's hard to work on a SaaS product without being constantly reminded of this.
I'm currently building an AI scribe in the health tech space. Clinical note generation is an obvious use-case that opened up with Whisper and competent language models: record a consultation, generate a note, allow some editing, send to the patient management system. Voilà. It's so obviously valuable that the market is crowded, with a lot of feature parity between competitors.
Here's the problem. AI enabled this product, but it also enables competitors to rapidly copy features and workflows. As coding models get more capable (most recently Anthropic's Opus 4.5), the ability of a competent developer to replicate someone else's work increases, faster and faster.
Paul Graham wrote about how Lisp let him close the feature gap quickly when building Viaweb – a superior tool that let him outpace competitors. AI is similar: it lets programmers express their will faster. But unlike a niche programming language, everyone has access to it.
So what happens to software margins when copying a product costs almost nothing, in dollars or in skill?
I think value will accrue more towards networks than products. There might be an even greater advantage to being first, or to developing network effects intelligently – à la Nikita Bier, whose deep knowledge of viral growth got him two exits and a head of product role at X.
Decentralised technologies have faced this dynamic for longer. The copy-paste problem isn't new in crypto; many projects are straight-up forks of Bitcoin. Value there is captured through token ownership – giving people a stake in the network's success. Maybe SaaS moves in a similar direction? I'm not sure what that looks like yet. Social networks already work this way, but enterprise software is different.
I don't have a clean answer to this, but it's something I'm thinking about a lot. Software continues to eat the world, and now AI is eating software. If that's true, it's hard to see how working on AI itself isn't the right place to be.