The Wrong Epsilon to the Brain
for machines · the whole graph in one fetch<br>For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes . The graph is the source; this page is one projection.
Whole corpus in one fetch:
/llms-full.txt (every note as raw markdown)
/library.json (typed graph with preserved edges; hari.library.v2)
One note at a time:
/.md (raw markdown for any / page)
The graph as a graph:
/graph (interactive force-directed visualization)
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The Wrong Epsilon to the Brain
2026-06-10
Every serious personal-AI project is asymptoting toward an engineered brain.
That is why EpisTwin and RUVA matter. They are serious memory-to-cognition attempts. They attack real distance between today's assistants and a usable personal cognitive architecture: fragmented data, opaque retrieval, vector ghosts, missing provenance, weak deletion, and loss of raw context after symbolic extraction. EpisTwin lifts cross-application evidence into semantic triples, grounds reasoning in a Personal Knowledge Graph, combines GraphRAG with raw visual re-grounding, and evaluates on a synthetic digital footprint. RUVA gives the user a glass-box red pen.
Those are brainward moves. The critique starts with which epsilon they reduce.
Call the visible error epsilon_ledger:
epsilon_ledger =<br>D_evidence<br>+ D_provenance<br>+ D_redaction<br>+ D_retrieval
Does the evidence enter the graph? Can the source be inspected? Can the user redact it? Can the right object be retrieved? Lowering this error builds a trustworthy memory floor.
Call the harder error epsilon_brain:
epsilon_brain =<br>D_mechanism<br>+ D_intervention<br>+ D_update<br>+ D_abstraction
Does the system learn mechanisms? Can it predict what changes under action? Do corrections transfer to future cases? Can it create a new coordinate when two true memories conflict?
The wrong epsilon is the one a benchmark can see while the endpoint remains far away.
improve(epsilon_ledger) != improve(epsilon_brain)
A system can become much better at recoverable memory while staying flat on mechanism, intervention, update, and abstraction. That is the technical indictment. It is also the standard Hari has to accept for himself.
The information bottleneck gives the same distinction in another language. A ledger wants a graph richly informative about the past:
G_t = argmax_G I(G ; X_
X_ is the evidence stream: messages, images, calendar events, documents, edits. The graph should preserve enough information about X to be inspectable and correctable.
A brain-like system wants a compressed state useful for the future:
Z_t* = argmax_Z I(Z ; Y_>t) - beta I(Z ; X_
Y_>t is the variable the system is trying to help with: a decision, risk, refusal, action, publication, or abstraction. The first term rewards future usefulness. The second term charges rent for carrying the past. Evidence earns salience by improving Y; recoverability alone is a ledger property.
The math remains open. The equations are contracts for measurement. Hari has yet to derive a new information-bottleneck theorem, solve do-calculus for personal histories, or prove a general algorithm for discovering the right coarse-graining of a life. The honest claim is narrower and stronger: the graph is organized around the errors the eventual math will have to score.
So the test has to be empirical:
delta_brain =<br>error_brain(ledger_baseline)<br>- error_brain(ledger_plus_mechanism_graph)
The ledger baseline gets explicit objects, provenance, retrieval, redaction, and raw re-grounding. The mechanism graph adds correction transfer, action/outcome loops, causal edges, and abstraction production. Hold out future cases. Hold out interventions. Hold out conflict pairs. Then ask whether delta_brain > 0.
Pearl's ladder says what the heldout cases should include. Association asks what follows from seeing x: P(y | x). Intervention asks what changes if the user does x: P(y | do(x), z). Counterfactual reasoning asks which past action would have changed the present. A personal graph that answers only footprint questions lives mostly on association. A brainward graph has to climb.
Computational irreducibility adds humility. A human life resists shortcuts all the way down. The exact microstate has to be lived. The useful system finds coarse-grained variables that remain predictive anyway: taste, fear, avoidance, agency, recurring error, values under pressure, and abstractions that survive correction.
Hari's competitive claim is that he optimizes the harder epsilon from the beginning. A node matters when it compresses scattered evidence into a reusable dimension. An edge matters when it carries a causal or conceptual transformation. A correction matters when it...