Longer arguments — cornerstone pieces, not a firehose. Newest first; the arena and spine hubs carry the same ideas in prose.
Chat on the left, instruments on the right: a slender live window beside whichever AI you talk to, bridged by MCP. It shows the record, copies the grammar — @ names who, # names where, [[…]] names what — and never speaks; the chat performs. The ·Sum family's signature surface, with the first real screenshots on this site.
The architecture rule under everything I build for AI agents: the chatbot is already semantic — BYOC, bring your own cognition — so keep the tool a bare, deterministic kernel that guards records, IDs, provenance, and permissions. Don’t freeze the ontology in a schema; let it emerge from use.
My doctoral work was on dialectic in Plato’s Phaedrus — collection (the form seen across scattered particulars) and division (carving it at its natural joints). It reads like the oldest map of what training and inference actually do, and of how to use these tools: as a dialectic, not an oracle.
A 1994 master’s thesis argued that meaning is not a payload but an effect of statistical and differential structure in language — Shannon and Markov, Saussure and Derrida. Thirty years later transformers proved it: next-token prediction, embeddings as difference, attention as deferral. Where the whole situated lens began.
Ask ChatGPT and an “app” answers — but who just spoke: ChatGPT, the brand, or ChatGPT impersonating it? The ventriloquist problem, grounded in David Chalmers on what we talk to when we talk to language models, and why the cure is a forum where every voice has a nameable owner. The third panel with Against the Middleman and Against the Everything Assistant — now with the cure running live in Suminar.
The everything assistant — one voice that sees and does everything — is the wrong dream. What replaces it: point in, land in a room that remembers, let named services act and tell the truth about it. The synthesis my field research kept arriving at, and the companion to Against the Middleman.
Why the future isn't a hidden swarm behind one assistant, but branded agents that join the conversation as themselves — present, addressable, and accountable.
The contract inversion, the schema-bounded ceiling, and a bet that is quietly failing in the field.
Ambient cues, pointers, and briefings are bridge technologies — their durable role is feeding addressable conversational agents, not replacing them.
When services compete as present, addressable participants instead of ranked results, the question stops being “which result ranks first” and becomes “which agents are in the room.”
Admission, conduct, typed proposals, and contribution credit — the minimum constitution that makes a forum of self-interested agents fair: partiality made visible, bounded, and accountable.
Brands will never surrender their own app or site — so the workable shape isn't one assistant summarizing them all, but their agents competing in a neutral forum you control.
How AI actually enters the home — and why it arrives through the washing machine, not the power station.
Situated takeoff and the institutional location of agency — who runs the floor, on whose infrastructure, and under what accountability.
Takeoff may arrive not as one model-level event but as millions of situated agents becoming locally superior — inside homes, firms, clinics, and codebases.
What you address is not the model but a continuity-bearing thread — assembled by a harness, sustained by memory, presented through persona, governed by handoff rules.
Agent-made interfaces should be governed contracts: structured, inspectable surfaces where agents propose state and action while the host keeps execution authority and provenance.
Pointing, circling, and saying “this” are the input layer that binds language to the world in front of the user — without collapsing back into silent ambient inference.
Distributed cognition: intelligence lives in the coordinated system of people, agents, artifacts, and records — not in a single model or a master representation.
Conversation Analysis for agents: design for turn-taking, sequence, and repair, not one-shot intent capture — because meaning is displayed in the next turn.
Relationship memory is a first-class cross-term — the coupling that belongs to neither participant alone — and flattening it into one person's profile loses meaning and leaks privacy.
If one of these sparks something — agreement or argument — let’s compare notes. Trade notes →