We’re in the midst of a revolution in which free software can be generated faster than the cost of paying for the software it is replacing. I mentioned replacing a third party analytics tool in Slab (writeup) and shared it with a friend who had a very similar story about him vibing up his own frontend for features which were pro on top of a third party service. Meanwhile reading today’s HN “what are you working on” thread I see multiple instances of people creating cheap or free alternatives to paid services in a matter of hours.
The shared pattern I see here I’m calling Spec Mining. It’s where you rip out the pattern of exactly what you want to have happen and have an LLM vibe it up.
Defining a specification takes work, modeling a problem, figuring out the kinds of moving parts and how they need to fit together. Usually that happens in concert with building, taking the learnings along the way, and then continually modifying software to adapt. Most software ends up with a bunch of edges that make it harder to work with as a result of this path dependence. Once that spec is created, though, as a fleshed out usable thing, it tends to be simpler to understand and easier to implement.
I think we’re going to see patterns emerge around spec mining which are even more efficient to the point where much software can be generated this way. (Setting aside software with hard safety or security requirements, though I think there’s a coming market of insurance for AI-generated outputs where the insurers specialize in automated security audits and the contracts allow startups to serve paying customers with things that might break and the insurance can pay out.)
A thought experiment:
Take your favorite SaaS. Now imagine you capture all the network traffic and UI modifications, and have an LLM generate a schema from these changes, followed by a sqlite db that can back that schema. Then have it generate a local server which can perform the mutations to the db, download the UI created by the company and reroute all the network traffic to your local instance. If it’s a collaborative cloud tool, throw it in a shared instance and reroute everyone’s frontend to use it. No need to pay for the service. You can even use their nice frontend (which it turns out generating nice frontends is actually still somewhat tough because LLMs suck at visual layout).
Now instead of generating the spec by personally clicking the UI in this website, have a web using agent automatically use the site to generate a bunch of interactions, capturing the traffic etc. from there. This is automated spec mining. One wonders how far we are from a workflow of
- Give a URL of a website you want
- Automated spec mining, codegen for the backend
- You have a local version of that website as software
So what next?
I’m excited for the anarchy of it all. I see groups coming together and writing up their needs of what they want software to do for them and producing tools which are custom to their particular niche (e.g. making better AAC apps), leveraging the existing tech to bootstrap a foundation to modify. It’s time for some updated notions of property in this age of abundance, and I personally am excited to practice giving things away for free in a network of mutuality.
(Thanks to ezez.win for sharing the AAC tiktok with me, it got my brain churning on this.)