Language

For thousands of years, language has been the interface between minds. We write, speak, draw, summarise, search, and explain because meaning has to be turned into a representation another person can receive.

AI changes that constraint. If a machine can transform, filter, expand, compress, verify, personalise, and eventually act on, meaning before another person encounters it, then language is no longer necessarily the entire communication process. It becomes one possible representation produced by a larger system.

This book does not argue that words disappear, or that AI captures complete human meaning. Language does not stop mattering. The question is what happens to communication when intelligence moves into the channel itself, and language becomes one representation among those a larger system can produce:

What form does communication take when intelligence becomes part of the medium through which meaning moves?

What the book builds

The book describes one imagined system, a layer that sits between people and AI systems as they communicate, and it grows that layer part by part. Bounded parts now run in a reference implementation; the complete layer, envisaged as a modest OpenCode plugin, remains a proposal. Its final form is still to be narrowed by implementation and experiment.

Part I: A New Language. Intent, concepts, representation, reconstruction. What happens when a source no longer has one fixed representation? The part begins with how a receiver rebuilds meaning from whatever a representation gives them, then follows fixed pages, previews, enhancement, source-linked semantic records, inspectable meaning inventories, sidecars, typed relationships, preservation obligations and the representations that sit between human language and machine state. Together they make a Semantic Browser: the layer working on one source, choosing a representation for a task, comparing proposed meaning at observable boundaries and correcting what failed to survive.

Part II: Too Much Information. Attention, selection, preservation, mediation. Once representations are cheap, what deserves to reach a person at all? Information units, task relevance, prior knowledge, standing semantic subscriptions, levels of resolution with a stop rule, and delivery timing compose into a Personal Information Radar: the same layer attending to a stream, with a decision record for every candidate and no single score. Too Much YouTube puts its outer edge to work on real subscription feeds, with a deterministic screen and no claim yet that the selections help.

Part III: The Personal Experience. Personal communication state and adaptation. When communication adapts to a person, what may the layer keep, who decides, and what must stay explicit? Declared, observed and inferred information kept apart, no stable person type, evidence and authority kept separate, session overlays that shed, user-authored friction, boundaries that travel between services, and beliefs the person can inspect. Optional personal conditioning can be switched off while binding constraints remain.

Part IV: The Second Reality. Prediction, agency, delegation, speech, Taint, discovery, ecosystem. What happens when the layer begins to predict, speak or act for someone? Prediction measured and kept apart from identity, what a person wants to keep, an explicit grant with an action gate and a data-use gate, speech typed by the act, and the hypothesis of a co-produced interaction signature called Taint. Possible affinity remains separate from permission as private discovery proceeds in stages and meets the failures of a proxy ecosystem. The book ends by asking where the word language stops being the right one.

Read in order, the parts follow the layer as it observes communication, keeps track of intent and concepts, preserves meaning, adapts its representation, maintains scoped personal state, predicts carefully, allocates agency, holds bounded authority, speaks and acts within it, and meets other people’s systems.

How the book reasons

Every transformation in this book, whether compression, expansion, personalisation, filtering or delegation, creates an obligation: to measure what survived it. And every mediation must remain inspectable and contestable by the person it serves. A third principle follows: a successful operation in one domain never automatically becomes evidence or permission in another. Similarity is not relationship. Preference is not performance. A model of a person is not the person. Prediction is not authority. Private computation does not legitimate its purpose.

Much of the book’s machinery comes from refusing to collapse concepts that look adjacent but are not interchangeable, and each separation is stated plainly.

What kind of claims this book makes

Three kinds of statement appear. Established background is what prior research supports independently of this book, stated with its populations and limits. Argument is what the book proposes: a mechanism, a distinction, a policy that prior work motivates without settling. Proposal is a design for a test that has not been run. The runnable examples demonstrate bounded mechanisms with a scripted stand-in for a model. The author’s production count and the dated Too Much YouTube scan are observations of real activity, identified where they appear; neither establishes a benefit from mediation. The other people, sessions and numerical examples are constructed unless stated otherwise. Where the book describes a future study, it is a design. The evidence the argument would need, and the ways it could be wrong, are recorded alongside the manuscript. The thesis itself, that language moves up a stack and becomes one rendering of a deeper communicative process, is a hypothesis, and the last chapter says what would count against it.

Why first principles

Rather than starting from today’s chat interfaces, the book starts from the underlying communication problem: one mind holds state another does not have, representations are lossy, attention is finite, recipients differ, and acting for another person requires authority. Representation, selection, adaptation and delegation follow from those constraints, as do the obligations of preservation, provenance and contestability that bind every stage.

You do not need to believe AI will reshape language to use this framework. The distinctions between representation and retrieval, attention and comprehension, preference and benefit, prediction and authority already matter to systems being built now. The book ends at the horizon where language stops being the whole communication system, while remaining one of its most important rendered representations, and says plainly how far that claim has been tested, which is not far yet.