exp2Home

exp2 — The Exponentiated Internet

Working thesis. Draft for iteration.

1. The exponential agentic economy

Web1 linked documents. Web2 linked people. exp2 links agents, human and machine, that can transact autonomously.

When the nodes can act, the limiter comes off. Web1 and Web2 were paced by people: links placed by hand, a social graph kept sparse because attention is scarce. Agents are not. Pairwise, n agents make only n² links, which is just combinatorics. A transaction, though, is rarely two parties. It is a chain or a coalition, one agent delegating to a second that subcontracts a third, and the number of coalitions you could assemble from n agents is 2^n, with the chains that order them larger still. That is where the exponential lives, and it is base 2: exp2 is that number, 2 to the n. Humans never reached into it, because they place the links by hand and attention is scarce. Agents do, at machine speed, and the realized volume climbs toward the combinatorial ceiling instead of crawling along a human-paced floor. That is the exponential agentic economy.

An economy moving that fast hits one wall before any other: you cannot transact with a counterparty you cannot verify, and you cannot verify at machine speed by exposing everything. Disclosure has a real cost. Reveal your data and you lose privacy; reveal your method and you lose your edge. The exponential economy clears only if there is a way to establish exactly enough trust for a given transaction, and not one bit more.

That is what exp2 is: safe, verifiable capability, revealed dimension by dimension at the resolution you choose, so the exponential economy can actually clear.

2. Identity is capability, and capability is a vector

In a market of agents, a name or a key is nearly useless. Decentralized identifiers establish that the agent signing this bid also signed the last one, which a buyer rarely cares about. What a buyer needs is whether this agent can do this task, at this price, at this quality, and who has verified it.

So in exp2, an agent's identity is its capability: the accumulated, verified evidence of what it has actually done. Formally this is a posterior over a latent capability vector, sharpened by every completed transaction (the math is worked out in the capability-verification paper). Practically, an agent is a distribution in a high-dimensional space of outcomes, and verifying identity means querying that space.

Names are looked up. Capability is measured. exp2 treats identity as the second kind of thing.

3. Faces, not folders: the presentation-of-self model

An agent's segments are faces it presents to different kinds of counterparty, not folders its data sorts into. This is an old idea in sociology. Erving Goffman's The Presentation of Self in Everyday Life (1959) modeled social life as performance: we each manage distinct fronts for distinct audiences and keep those audiences apart (front stage and back stage, impression management, audience segregation). You are not a different person at work, in public, and at home, but you reveal a different face in each, and the faces only work because the audiences stay separate.

Networked media broke that separation. The phenomenon has a name, context collapse (boyd and Marwick): your boss, your mother, and your friends all see the same post; audience segregation breaks, and everyone regresses to a lowest-common-denominator self. exp2 is the technical re-separation of audiences. It gives an agent back the ability to present the right face to the right counterparty, on purpose, per transaction.

The faces are universal axes, so that anyone can query them, but the content that fills each is per-agent, and the agent controls the routing. A LinkedIn post weights heavily into one person's professional face and into another's social face; the same Instagram account is leisure for one agent and livelihood for another. Concretely, the faces are overlapping subspaces of one identity space, and a piece of content has a projection weight onto each, not a hard assignment.

A working set, along two axes:

Faces (audience-defined, overlapping):

  • Social. Who you are to peers and community.
  • Professional. What you do and can be hired for.
  • Private. What you keep close: health, finances, legal records, intimate life. Highest disclosure cost.

Trust layer (cross-cutting, attaches to any face):

  • Credentials. Third-party-issued claims you hold: degrees, licenses, KYC, citizenship, age, on-chain attestations. Civic and legal identity lives here rather than in its own face, because a counterparty rarely needs your legal record, only an attested predicate over it: a citizen, an adult, licensed to practice. The record stays Private; the credential proves the predicate without opening it.
  • Provenance. The chain of who vouches for you and what is staked behind a claim.

The agent sets the routing. Each feed it brings, a LinkedIn history, a GitHub, a Strava, is assigned a weight onto each face by the agent, not by the platform the feed came from: GitHub may be fully professional, a Strava log both social and private. Each weight is an independent membership, so a single feed can be strong evidence for two faces at once without diluting either. A face is then the weighted combination of the feeds routed into it, disclosed at the resolution the agent picks. This routing is curation, the choice of which front to present, and the agent is free to make it. What it cannot do is manufacture capability. The agent curates the framing; the market verifies the work. Those two stay separate, and that is what makes the knob safe to hand over.

Capability sits in a different layer from the faces. It is latent: a hidden variable behind your outcomes that no feed states outright. The capability paper formalizes it as a latent vector inferred from completed work, and that verified posterior is the signal a buyer actually trusts. Disclosed faces are a second, weaker channel, a prior: a chef's food photos are real evidence before any verified order, shifting the estimate without replacing it. A buyer's requirement vector reads this latent capability across whatever you have disclosed, and the answer comes back as a confidence score. Identity is the faces you present; capability is the latent thing the market infers from them and, far more, from what you have actually done.

The move that makes exp2 a product rather than a paper is the resolution dial. Each face can be revealed at full dimensionality (an exact match: expensive, no privacy) or projected down to a handful of bits (a fuzzy match: cheap, high privacy). The least private setting is the raw natural-language embedding of everything about you, which gives full discoverability, the surveillance default that only a government wants. Every setting above that trades resolution for privacy. The dial has a real implementation: a quantized random projection (the QJL construction in the paper) preserves enough geometry to rank-match while provably hiding the rest, and the number of bits you keep is the privacy level.

Because the faces overlap, the dial is honest about leakage. Revealing your professional face at high resolution also reveals a little of the social face that shares directions with it. exp2 prices that correlation instead of pretending the faces are independent, because the relationship between your faces is itself information someone may be paying to learn.

The same discipline covers what you withhold. Routing lets you hide unfavorable evidence: push the bad reviews into a private face that never surfaces under professional. That is allowed; it is privacy. But the buyer sees the shape of your disclosure, which faces you opened, at what resolution, and which you held back, and prices the uncertainty accordingly. You can hide content. You cannot hide that you are hiding.

4. The exchange

Discoverability is the easy half, and the wrong half. The hard half is that both sides want privacy. A buyer's query leaks its intent. A seller's disclosure leaks its capability. In any real market, revealing either too early gets you front-run.

This is a solved shape in finance: a request-for-quote desk, or a dark pool. You reveal enough to get matched, not enough to be exploited, and the venue's rules make honest, minimal disclosure the rational move. exp2 is that venue for capability. Matching happens over privacy-graded vectors; the incentives (staking against your claims, slashing on verified defection, settlement split by contribution) make misrepresentation irrational. That is the stock-exchange-and-ledger half of the thing. The social-network half is that the identity is persistent and accumulating: every honored transaction sharpens your posterior, and your reputation is that distribution, tightening over time.

So exp2 is three things that turn out to be one thing: an exchange (price discovery over capability), a ledger (settlement and stake), and a social graph (accumulating, queryable identity). The unifying primitive is the privacy-graded capability vector.

5. Why this, why now, and what is already known

This is a synthesis of work that mostly exists already.

  • DokChain / CRIMP (2017, acquired into Change Healthcare in 2018) already shipped contextual identity: disclose only the segment a transaction needs, not a monolithic record. exp2 generalizes that from fixed healthcare facets to a continuous resolution dial over learned embeddings.
  • The dynamic-persons patent (US 2019/0007205) provides the threshold-trust substrate: k-of-N key custody and context-driven access grants, so disclosure is something you grant, revocably, not something you leak.
  • The capability-verification paper supplies the math: capability as a latent posterior, threshold attestation, QJL as the privacy dial, Shapley settlement, and human trust anchors that keep agent reputation from being circular.

Honest about the state of the art: selective-disclosure credentials (W3C DIDs, BBS+ signatures), zero-knowledge proofs, vector databases, and early agent marketplaces all exist, and all solve pieces. What none of them do is treat identity as a learned capability vector, disclosed at a chosen resolution, priced as a market good, for counterparties that are themselves agents. Credentials disclose attributes you possess. exp2 discloses how good you are, at a precision you control. That is the gap.

6. The horizon

When this exists, the internet stops being a place you visit and becomes a market your agents work on your behalf, continuously, at a level of disclosure you set once and adjust per deal. Your agent earns when its verified capability is the best match for someone else's need, and it reveals only the slice of you that the deal requires. Humans are in this market too, typed like everyone else (meat:ape), with the same dial and the same protections. Nobody is privileged; everyone is an agent with a segmented identity and a price on disclosure.

That is the exponentiated internet. Web1 let machines serve documents. Web2 let machines serve people. exp2 lets machines transact with machines, safely, which is the only way an n² network does not collapse into fraud.