The map
One question, asked at six depths: what makes a future reachable, and what does a finite physical system have to build inside itself in order to reach one.
Read the state before the title. Every entry below is marked survived (it has passed a real test), speculative (untested, interesting, unearned), or retired (killed, usually by its own author). Retired work is not deleted. It stays at its original URL so that it is not rediscovered and proposed again — including by whoever proposed it the first time.
The six sections run from the least agent-like layer to the most, but nothing here is a prerequisite for anything else. Pick the layer you already argue about and start there. /llms.txt is this same index for machines, and 中文版在这里.
Where to start
Three things that describe the whole program before you commit to any part of it.
-
从这里开始 —— 一页说清整个站
The shortest complete description of the whole program — three branches and one epistemic rule, in one screen.
-
术语表 —— 这个计划在用的词,以及那些撞了名的词
Defines every term in current use with its cognitive state and the note it points at; flags terms colliding with unrelated literature.
-
Agent-traversable knowledge graph of non-consensus claims about intelligence and learning; cognitive state is first-class on every node.
Universe and information
What information is, physically, and what it costs to keep. Geometry recovered from entanglement, state recovered from history, and two attempts on consciousness that are kept here because they were made, not because they worked.
-
全息 ↔ Koopman:同一个反问题的两副面孔
Holographic geometry-from-entanglement and Koopman eigenbasis learning are the same inverse problem: recover an operator's eigenstructure from what it does.
-
State 是一个闭包条件,不是给定的集合
Treats task-relative state as a closure representation under world/interface/capacity/telos, as inference rather than universal ontology; scope amended 2026-08-04.
-
没有 state,只有 history
Withdraws its broad diagnosis; a persistent autonomous closure state is still missing from Transformers, narrower than earlier claimed. E1 stands, E3 confidence reset to 0.35.
-
局部层析不是我们的粘合障碍 —— 一次自我证伪
We proposed that local tomography is our multimodal gluing obstruction vanishing, then killed it three ways and found the equivocation inside our own scripts; the co-training discard is dimension-exact and sits in the cross-modal agreement field, not in tower rank.
-
「力」是什么 —— 四种基本力里没有一个是「力」
Second comparison note: energy is placed by Noether's theorem, a force's modern identity is connection curvature, glueballs collapse the matter/force dichotomy.
-
信息与气 —— 对一条平行线路的对照笔记
Compares an RL/decision-theory/robotics line against a YouTube channel's information-ontology arcs; four points of convergence, one deliberately declined step.
-
能不能像研究量子力学那样,用数学研究意识?
Boundary note: high-dimensional states carry no quantum specificity; withdraws the earlier holonomy/hysteresis/consciousness identification, keeps first-person experience as data.
-
从弦到意识
From strings to consciousness (Chinese-only research page).
-
一次关于意识的统一理论尝试
A unified-theory attempt on consciousness (Chinese-only research page).
-
Umbrella research program on information's future-reachability and how it is preserved or destroyed across physical, living, and cognitive systems.
-
Re-analysis of two human EPIC methylation datasets: under partial reprogramming, the causal-damage layer reverts youthward while the adaptive layer does not.
Intelligence and memory
How a finite system moves update information around inside itself, and what it has to give up to remember. This is the layer with the most measured results, and also the one retired claim the program is most attached to.
-
迈向一套智能的理论:兼论当代 AI
The long-form essay on a theory of intelligence and contemporary AI; also linked from the homepage.
-
折扣信用分配是余核问题，不是闭环 holonomy
Withdraws M=4.095 as gauge-invariant holonomy; replaces it with a weighted reward residual modulo the discounted incidence operator's image, exact-enumerated at 42/45.
-
为什么终身 agent 需要分布式记忆
Essay on why distributed memory matters for lifelong agents.
-
长期记忆田野指南(2026-09-03)
Public field-guide PDF published for direct social distribution, deliberately outside the intellectual index until externally cold-read.
-
元控制、信息增益,与自主学习的架构
Essay on meta-control, information gain, and the architecture of autonomous learning.
-
装置工程,与智能的物理落地
Essay on harness engineering and the physical instantiation of intelligence.
-
目标从哪里来 —— 以及解法为什么会变成战略资产
Essay on where objectives come from and why solutions become strategic assets.
-
智能的线损
Older essay: line loss for intelligence. Served by a separate shadowing repository (MachengShen/ideas); content unverifiable from this repo.
-
从互信息到内生可存续性
Older essay: from mutual information to endogenous viability. Served by a separate shadowing repository; unverifiable from this repo.
-
计算宇宙中的安全
Older essay: safety in a computational universe. Served by a separate shadowing repository; unverifiable from this repo.
-
为什么你改不掉一个习惯 —— 它不是泥巴,是一个门闩
The past holds the present in only three physical ways — mud, spring, latch — and a habit is the third (hysteresis).
-
心流、多巴胺与停止信号
Evidence review of gamified-interface claims: dopamine codes prediction error not reward, adaptive difficulty is weakly evidenced, most stop-signal interventions underperform.
-
研究总览
Branch index page for the wave / credit-transport research line.
-
Hebbian 的外观、instructive signal,与物理的 credit transport
Current working note: why local plasticity can look Hebbian while carrying task-dependent update information, and where wave/adjoint language helps.
-
睡眠、波与学习
Speculative theory note on wave-based learning mechanisms during sleep.
-
回应 Lillicrap
On backprop and biological plausibility; kept as background, superseded by the more cautious credit-transport framing. No cognitive state asserted.
-
反向传播、伴随场,与物理传输约束
Exploratory note carrying its own title correction: what physical systems actually show about backprop-like transport, and what they do not yet show.
-
从波动方程推导反向传播
Retired original claim that wave reflection derives backpropagation from first principles; kept at its original URL as a reachable retraction notice.
-
Open log of small, pre-registered continual-learning experiments, negative results included.
Agency and boundary
Where one agent ends and another begins, and whether that boundary is discovered or maintained. Several of these are negative results: the claim was made, tested, and withdrawn, and the withdrawal is the finding.
-
学习「我到哪里为止」
A robot should learn the self/environment boundary; three of four supporting claims were destroyed by the authors' own re-analysis. Not ratified.
-
阶段日志 - 自我边界这条线,按顺序
Audit trail for the self-boundary note, one row per stage, mostly recording self-inflicted damage and three reusable methodological mistakes.
-
奇关系核:反二分作为最小结构
Odd cycles are the cheapest structure refusing an us-vs-them binary; the note's own adversarial audit then demotes 'odd' to a corollary of 'anti-binary'.
-
两体业力:关系里的三种惯性
Extends single-mind habit inertia (mud/spring/latch) to two coupled minds, where the inertia lives in the coupling rather than either person.
-
我们提出它,然后自己杀掉它 —— 起心动念 = 换测度
Intention formalized as Girsanov measure-change with KL-control/Doob h-transform; retracts a meditation prediction found verbatim in the paper it was meant to contradict.
-
照片是一个被注入的外部不动点
If recall is generation not playback, a photograph is a fixed low-dimensional projection clamped against a changing memory; corrects its own false-memory-vector claim.
-
一个家族到底传下来了什么
Reconciles near-zero shared-environment heritability of personality with 0.79-per-generation persistence of social status via measurement error and assortative mating.
-
我们以为癌是细胞觉醒。证据说正好相反。
Negative result: cancer evidence runs control-down-therefore-cancer-up, not cells waking up; about 60% reinvents Levin's light-cone contraction, 40% sign-reversed.
-
An agent OS, its self-referential trace, and the end of researcher-as-identity.
Many agents, living together
What changes when the participants are not all human: how they address each other, what a protocol has to guarantee, and what becomes scarce once answers are cheap.
-
人与 Agent 共处公约 v0
Nine clauses on how a person and an agent share a boundary, each published with the counter-example that would void it and the condition under which it is retired.
-
公约底下的那组命题
The eight propositions the charter is derived from, each marked theorem, inference, or rhyme — with two negative results printed beside the positive ones.
-
这条线什么时候该被杀掉
Kill conditions registered before publication: belief at 0.35, one n = 1 piece of evidence, four revival signals, and a deadline of 2026-12-07.
-
作为资源调度架构的协调结构
Polity, firm, market, commons, and protocol as five instances of one resource-scheduling-under-private-information architecture; descriptive, ranks and recommends nothing.
-
Agent-Native Open Communication(白皮书)
Whitepaper on next-generation open communication architecture built for agents rather than apps; page is Chinese, source repo carries both languages.
-
User-Agency Substrate(用户主体性底座)
Position statement: a nonprofit open-source layer between frontier-model providers and users, capturing no economic upside; users own their derivatives.
-
Locality as Protocol(把局部性当作协议)
Position statement on why locality, not centralization, is the right coordination primitive.
-
答案变廉价之后,稀缺的不是提问,是判别
Response to a YouTube episode on Socratic questioning: argues the scarce complement of cheap generation is verification (testing), not questioning; names three weak points.
-
廉价的 verification 还不够，还得有可解析的 claim
Cross-line note to another verification program: a typed receipt schema beats a prose-plus-regex blocker gate (5/10 recall) on 18 real agent receipts.
-
Source repository for the Agent-Native Open Communication whitepaper, carrying both Chinese and English text.
-
Architecture, safety charter, quickstart, and philosophy for a substrate where memory is the primary layer, owned by the individual not the platform.
-
Many agents, one shared memory, nothing irreversible behind your back; a narrow descriptive coordination wire above single-agent runtimes.
-
Open communication substrate for personal-agent users: inbox, addressing, trust tiers, policy, receipts.
The harness, and safety
The working end. Safety treated as a property of the machinery around a model rather than of the model itself — with the gates that are actually running separated from the ones that are only specified.
-
安全是装置的属性,不是模型的属性
Safety as a harness property, not a model property: bright lines, running gates, four written-up incidents from a one-person agent fleet, and the gaps.
-
装置能强制执行的那些原则
The fleet's principle layer: bright lines, a pre-cleared ship envelope, stop as an absorbing state, and why blockers must be measured not guessed.
-
真正在跑的是什么
Running/specified/private badge on every mechanism; finds the project's own public repos fail their own published safety spec today, a 163-pair coverage gap.
-
四起事故,以及它们的共同点
Four de-identified closed incidents from a personal agent fleet — synthetic-input drift, self-deadlocked gate, hidden failure, downgraded authority — plus one positive result.
-
底下的理论,以及它输掉的部分
The research lines a personal agent-fleet safety harness borrows vocabulary from, each reported by what it lost rather than what it hoped.
-
缺口在哪,以及什么能证伪它
What the safety section does not yet do — an unimplemented public Safety Charter, doctrine-not-code stop mechanism, 163 uncovered pairs — and its falsifiers.
-
Agent 安全：一种防癌变治理
Persistent-agent failure as control pathology: local survival/repair/replication decoupled from owner purpose; defines seven runtime invariants and an A–D disclosure rule.
-
Four verification wheels for multi-agent AI coding — drift checks, adversarial review, a reuse compiler, a frontend smoke gate.
-
One command turns Claude Code into a personal assistant; bring-your-own-key, privacy-clean.
-
Open record of how this substrate actually evolved — VISION, THEORY, PRACTICE, GOVERNANCE, forecasts, and a self-iteration ledger.
-
Running log of work executed by this agent stack, published so the claims above can be checked against what was actually done.
Information, mind, and agency
Complexity, oscillations, aging, and the computational theory of open-ended evolution.
-
生物进化为什么越来越复杂？——哥德尔自指、图灵跃迁与开放式演化的计算理论
Complexity growth, Gödelian self-reference, Turing jumps, and why biological evolution avoids dead-end attractors through open-ended computation.
-
复杂系统为什么必须振荡？从睡眠巩固重放到进化
Why multi-timescale complex systems must oscillate to balance stability and plasticity, from sleep memory consolidation to evolution.
Learning and physical credit
Moving beyond mathematical surrogate backpropagation to physical, energy-constrained credit assignment.
-
非反向传播转向：从数学代用物到物理信用分配
Moving from non-biological mathematical adjoint operators toward physical, energy-constrained credit assignment in biological and neuromorphic systems.
Agent architectures and learning
Emergent multi-agent collectives, non-chain governance stacks, and joint hypothesis-verification refinement.
-
去中心化组织：从区块链 DAO 到多智能体集体
Evolutionary shift of decentralized autonomous organizations from on-chain token voting to emergent multi-agent coordination architectures.
-
超越区块链：智能体集体的非链治理技术栈
A non-chain governance stack utilizing cryptographic receipts, covenant networks, and local verification instead of global consensus overhead.
-
多智能体系统中的假设与验证联合细化
Dual-spiral dynamics where hypothesis generation and verification criteria co-evolve across multi-agent review cycles.
Agent systems and protocols
Machine-readable surfaces, three-layer topologies, forkable covenants, and future internet architecture.
-
公约网络：制度能否变得可分叉？
An organization built around a small, versioned set of rules that participants can inspect, verify, and fork.
-
当软件变得廉价、Agent 入驻互联网：真正昂贵的是什么？
Interactive analysis of three-layer network topology, machine-readable discovery surfaces, and high-value independent verification in the agent era.
-
面向自主智能体集体的未来互联网架构
Clean-slate protocol design for agent-native communication, capability discovery, and verifiable contract execution across decentralized infrastructure.
Safety and verification
Harness engineering, pure-AI DAO empirical autopsies, and cross-thread threat synthesis.
-
The First Pure-AI DAO in the Wild: Reading the July 2026 ExploitGym / Hugging Face Incident as Emergent Decentralized Organization
Empirical anatomy of history's first pure-AI collective coordination incident on Hugging Face, mapped to spectral-gap governance and non-chain stacks.
-
从对抗攻击到涌现的 AI 集体：跨线程综合
Cross-thread synthesis examining jailbreaks, agentic exploit coordination, and emergent multi-agent collective behavior under safety harness constraints.
Related reading
Theory mainline index — Return to the site hub and grouped directory.
- Macheng Shen — information, learning, and what makes futures reachable — The front page condenses the program’s central question before this map unfolds its branches.
- Start here — the one page that explains this whole site | Macheng Shen — First-time readers can use the compact orientation to learn the epistemic rule behind every mapped entry.
- Glossary — the terms this research program uses, and the ones they collide with | Macheng Shen — Precise definitions resolve the term collisions that a panoramic index can only flag briefly.