Headless internet
I've been using AI to explore some of the little brain worms that bobble around in my head. I thought it would be fun to write up one of these conversations as a record of how I used AI in what will seem like relatively early days at some point in the future.
ME: “Do you think there will ever be a headless internet, like a headless CMS but for everything?”
Turns out this is already happening, which I’m going to say I knew. Data has been decoupling from presentation, REST APIs, GraphQL and headless CMS platforms are all examples.
Claude did a good job of explaining why this is becoming a more mainstream question (which is no doubt why I’m asking it). Basically:
CLAUDE: “The HTML, CSS, JavaScript rendering pipeline was built for human eyes. For agents it's archaeology.”
Its prediction is that we’ll end up with a ‘dual layer web’ which has both a human and agent friendly interfaces. That seems reasonable to me, and, although not the perfect analogy, kind of reminds me of the many different ways it was possible to watch TV when I was a kid. Cable, satellite, terrestrial, VHS, DVD, Blueray…
Anyway, I then asked:
ME: “What do you think is interesting or unexplored in this space?”
The whole answer was pretty great but the most interesting bit for me was about how schemas require a shared ontology:
CLAUDE: “Even if every site exposed clean JSON, what does a field called price actually mean? Is it inclusive of tax? In what currency?”
One of my favourite things at this point in a conversation with Claude is to get it to pull in a bunch of philosophy, sociology and anthropology. This has two benefits. Firstly it makes me feel smart. Secondly, there is usually some smart thinking by actual humans which help me get into the theme a bit more.
Some of the references were familiar and sounded plausible, Wittgenstein says meaning isn’t a property of words, but a product of use within practice, Saussure’s idea of the split between signified and signifier. But what was new to me, and felt very resonant was this:
CLAUDE: “Eleanor Rosch's prototype theory is also relevant. Categories don't have hard edges — they have a prototype at the centre and fuzzy membership toward the margins. A "chair" prototypically has four legs, a back, is made of wood. But a bean bag is a chair. A tree stump used for sitting is arguably a chair. Schema.org implicitly assumes classical categories with necessary and sufficient conditions. Natural language doesn't work that way, and neither do the human concepts underlying it.”
It then came up with this punchy little conclusion:
CLAUDE: meaning is irreducibly social and contextual, and any attempt to formalise it into a universal schema is simultaneously an act of power and an act of impoverishment. The vocabulary gets frozen; the practice continues evolving underneath it.
To which the obvious question is, if not schema.org and perfect JSON files then what?
Claude suggested that what AI can potentially do is bridge between two people who use different words to mean same things. For anyone whose tried to design a form with a single question that works for all users this is an exciting prospect. A magical “semantic broker” that bridges the gap between, for example, someone who needs help with a medical problem and the organisation which has a nomenclature indecipherable to the average joe. Wouldn’t that be nice?
Well, turns out Claude also thinks there are some pretty big trade-offs within this future (assuming you prompt a critique into existence):
CLAUDE: A form field that can't capture your situation is frustrating, but it's legibly frustrating. You can see the gap. You can point at the dropdown that doesn't have your option. You can complain, and someone might eventually add it. A system that translates your rich, contextual description into someone else's rich, contextual description — smoothly, fluently, without visible seams — is doing something more dangerous. It's making the loss invisible.
Had to prod Claude a bit more to get to a clear description of why this is bad, got there in the end though I think:
CLAUDE: Over thousands of interactions, those small approximations accumulate into an invisible policy. The broker is making de facto decisions about what counts as urgent, what counts as close enough, what needs get quietly downgraded — and nobody can audit that because it never looked like a decision. It looked like a conversation.
And here ended my pub chat with Claude. I didn’t completely lose my mind at Claude’s tone and it scratched an itch, but to be honest I do still feel a bit like the guy at the end of burn after reading:
Sign in to leave a note.