NOUS OS Research Line · Anchor Atlas · living document

What surrounds this work, and what is empty.

Curated map of academic research, industry voices, top products, and individual essays and podcasts that surround the NOUS OS research line. Each entry has the same four lines — what it is, the core claim, where we share, where we differ and what we add — plus a status flag for our depth of engagement.

note-written · full 1-page inbound note exists scanned · positioning recorded, no full note yet queued · known to exist, awaiting next reading cycle

Six buckets.

Each bucket is an angle the research line connects to. Move sub-line forward when evidence accumulates, not when arguments accumulate.

The lineage we walk in.

Pre-LLM thinkers who already articulated the augmentation vs replacement question. Their ground is firm; our addition is the LLM-era boundary taxonomy and the capability-without-AI instrument.

Vannevar Bush"As We May Think" · 1945 · essay
scanned
Claim
Knowledge work is bottlenecked by retrieval, not generation; the right tool extends memory without replacing thinking.
Share
Augmentation framing — tools extend, do not replace.
Diff
Bush imagined hardware; we instrument the loop between human and AI as the measurable unit.
Doug Engelbart"Augmenting Human Intellect" · 1962 · SRI report
scanned
Claim
Capability is a coupling of human, language, tool, training; the right design loop raises civilization's H-LAM/T capability.
Share
Explicit "human at the center" language; system-level thinking.
Diff
Did not live to confront LLM misalignment; our boundary taxonomy + capability-delta instrument are LLM-era additions.
Seymour PapertMindstorms · 1980 · book
scanned
Claim
Children learn most when they build; computers should enable construction, not deliver answers.
Share
Student-as-builder framing; Sandbox is constructionism re-instantiated with AI.
Diff
Papert used LOGO + minimal scaffolding; we add explicit boundary phases because LLM output is confident in ways LOGO never was.
Lev VygotskyZone of Proximal Development · 1934/1978
scanned
Claim
Learning happens in the zone between what a learner can do alone and what they can do with help; scaffolding withdraws as competence grows.
Share
The Sandbox loop is dynamic AI-as-scaffold in the ZPD.
Diff
We make the withdrawal of scaffolding measurable via the capability-without-AI delta.
J.C.R. Licklider"Man-Computer Symbiosis" · 1960
queued
Claim
Most productive computing will be symbiosis between human and machine, not full automation.
Share
The literal word symbiosis — same lineage.
Diff
Licklider's symbiosis was speculative; ours is operational with measurable boundary integrity.

The reverse front.

Empirical and theoretical work showing offloading cognition onto tools changes what's encoded internally. NOUS OS sits in the literature gap that asks the inverse: under what conditions does offloading make people stronger?

Risko & GilbertCognitive Offloading · 2016 · Trends in Cognitive Sciences
queued
Claim
Humans systematically offload memory, calculation, decision; this changes internal cognition, often costing native capability.
Share
The descriptive framing — offloading happens, does cost.
Diff
The core NOUS OS question is the reverse: under what conditions does offloading make humans stronger?
Sparrow, Liu & Wegner"Google Effects on Memory" · 2011 · Science
queued
Claim
People remember less when they know they can re-look-up; access to external info changes encoding.
Share
Validates that the offloading concern is empirically real.
Diff
They did not test interventions that flip the effect. We do.
Benjamin Storm et al."Saving effect" on memory · 2014+
queued
Claim
Externalizing memory can free or impoverish the internal trace, depending on conditions.
Share
"Depending on conditions" — exactly the conditional we want to characterize.
Diff
We move from file-saving to AI-mediated learning; conditions become design parameters of a scaffold.
Barry ZimmermanSelf-Regulated Learning · 1989+ corpus
queued
Claim
Effective learners cycle forethought → performance → self-reflection; SRL is teachable.
Share
The Sandbox 6-phase loop is an AI-native SRL instantiation; reflection card maps to SRL's self-reflection phase.
Diff
SRL is silent on what happens when a confident AI is in the middle. Our boundary phases are an AI-era addition.
John SwellerCognitive Load Theory · 1988+
queued
Claim
Intrinsic / extraneous / germane load determines what is learned; design that minimizes extraneous load wins.
Share
Our 20-min / 6-phase / 3-4 min-per-phase structure is implicitly CLT-respectful.
Diff
We should explicitly cite CLT going forward; current docs do not.

The contemporary peers.

Same audience, same era, but mostly descriptive taxonomies and policy frameworks. We sit downstream as an instrumented loop with measured outcomes.

Long & Magerko"What is AI Literacy?" · 2020 · CHI
note-written
Claim
16-competency framework for AI literacy across understand / use / evaluate / ethics dimensions.
Share
Same target audience; their critical-evaluation competency maps to Sandbox phase 4.
Diff
Descriptive taxonomy vs prescriptive 20-min protocol with measurable outcomes.
Note
2026-05-17-long-magerko-2020
UNESCOAI Competency Framework for Students · 2024
queued
Claim
International policy framework — four dimensions × twelve competencies students should develop around AI.
Share
Student-facing focus, public-interest framing.
Diff
UNESCO writes standards; we run experiments. They consume evidence; we produce (small-N) evidence.
Stanford HAIHuman-Centered AI Institute · ongoing
queued
Claim
AI should be developed alongside humans, not at them; CRAFT classroom materials.
Share
"Human-centered" framing.
Diff
Publishing institution vs instrumented practice. Worth importing CRAFT lessons.
MIT Media LabLifelong Kindergarten · Personal Robots Group
queued
Claim
Learning compounds best when projects, peers, passion, play are present; AI partners can support these if authorship is preserved.
Share
The "preserve authorship" line is identical to ours.
Diff
They focus on creative-project authorship; we focus on research-loop authorship.
CHI / CSCW / Learning SciencesPrompt scaffolding empirical work · 2024–2026
queued
Claim
Specific prompt designs improve learning outcomes (e.g., "ask AI to ask you questions first").
Share
We are downstream practitioners of these patterns.
Diff
Most measure satisfaction or correctness; we measure capability without AI, which is rare.

Same problem from the lab side.

Labs publishing on alignment, scalable oversight, and model-side boundary design. Their evidence is on AI capability; ours is on human capability when AI is present.

AnthropicConstitutional AI · tool use · interpretability
queued
Claim
Structured human feedback + explicit constitutional rules produce better-aligned model behavior.
Share
The explicit-boundary design philosophy.
Diff
Anthropic's boundaries are model-side; ours are interaction-side. Complementary halves of a symbiosis.
DeepMindScalable oversight · AlphaFold/AlphaProof/AlphaProteo
queued
Claim
AI can augment scientific discovery; scalable oversight is solvable.
Share
The augmentation thesis.
Diff
DeepMind's evidence is on AI capability; ours is on human capability when AI is present.
OpenAIGPTs in the classroom · educator partnerships
queued
Claim
Deployment studies of GPTs as tutors, study helpers, etc.
Share
Practical "AI in education" focus.
Diff
They measure usage + satisfaction; we measure human capability delta.
Stuart RussellHuman Compatible · 2019 · book
queued
Claim
AI safety requires AI that explicitly defers to humans and remains uncertain about goals.
Share
Responsibility-stays-with-human stance.
Diff
Russell is at alignment-policy layer; we are at daily-interaction-design layer. His policy implies our interaction design.
Dario Amodei"Machines of Loving Grace" · 2024 · essay
queued
Claim
If alignment is solved, the next decade could see profound gains in health, science, education, freedom.
Share
The optimistic-but-conditional vision.
Diff
Amodei describes the destination; we propose the per-session protocol that makes a human less likely to lose capability on the way there.

Same-shape attempts in production.

Commercial systems instantiating pieces of our protocol — sometimes one phase, sometimes a sibling philosophy. Useful benchmarks and design references; we are product-agnostic by design.

Khanmigo (Khan Academy)AI tutor product
queued
Claim
Socratic AI tutoring at scale with safety rails for under-18 users.
Share
Socratic / hints-not-answers stance.
Diff
Khanmigo is the AI; Sandbox is a protocol that wraps any AI. Closest commercial cousin to study.
NotebookLM (Google)Source-grounded research notebook
note-written
Claim
Source-grounded AI is more trustworthy and useful; answers cite source passages.
Share
Phase 4 (source check) directly aligned with NotebookLM's source-grounding ethos.
Diff
NotebookLM optimizes one phase; we structure the full 6-phase loop including boundary + reflection.
Note
2026-05-17-notebooklm-product-walk
CursorAI pair programmer · editor
queued
Claim
Developer productivity rises when AI suggestions are explicitly accept-or-reject rather than silently applied.
Share
The explicit accept-or-reject boundary — directly analogous to our boundary phase.
Diff
Cursor proves the pattern in the code domain; we propose the same in research/learning.
Claude Code (Anthropic)Terminal-native agentic coding assistant
scanned
Claim
Agents can do real engineering work safely if boundaries — permissions, plan mode, transparent diffs — are explicit.
Share
Every key design choice — explicit permissions, plan mode, transparent diffs.
Diff
Claude Code is the engineering instantiation of NOUS OS principles; Sandbox is the learning instantiation.
Cognition Devin · Replit Agent · OpenAI OperatorAutonomy-maximizing agent products
queued
Claim
Agents can do longer-horizon tasks unattended.
Share
Shared agentic substrate.
Diff
Autonomy-maximizing vs symbiosis-maximizing. Useful contrast points.
Granola · Mem · similarAI + your notes
queued
Claim
Drafting is cheap; human curation is the value-add.
Share
Human-curates-before-commits pattern.
Diff
They optimize a workflow product; we extract the pattern and codify it as a loop principle.
Perplexity · PhindCited-search AI
queued
Claim
AI search with cited sources beats non-cited AI chat.
Share
Source primacy.
Diff
They optimize answer-with-citations as the whole interaction; we make source-check one explicitly-bounded phase.

The living thinkers.

Essays, podcasts, and contemporary writing that move faster than papers. The L1 capture cron will pull selectively from these.

Andy MatuschakTools for thought · andymatuschak.org
note-written
Claim
Most "tools for thought" are not transformative; transformative ones change the medium of thought itself.
Share
Medium-changes-thought framing.
Diff
Individual cognition vs explicit symbiosis with boundary taxonomy. His vocabulary will inform L3.
Note
2026-05-17-matuschak-nielsen-2019
Michael NielsenReinventing Discovery · Matuschak collab essays
scanned
Claim
New mediums of representation enable new kinds of thought.
Share
Representation-as-cognitive-substrate stance.
Diff
His essays are inspirational and largely ungrounded in trials; we run trials.
Bret Victor"Inventing on Principle" · Dynamicland
queued
Claim
Thought is constrained by representation; better representation enables better thought.
Share
Deep belief in representation-as-cognition.
Diff
Victor's work is largely demonstrative; we add empirical loops.
Ethan MollickCo-Intelligence · 2024 · One Useful Thing Substack
queued
Claim
Treat AI as a co-worker; experiment widely; expect rapid capability shifts.
Share
Practical-empirical attitude.
Diff
Descriptive + adult-facing; ours is prescriptive + instrumented + student-facing.
Cal NewportDeep Work · Slow Productivity · podcast
queued
Claim
Deep cognitive work requires uninterrupted attention; modern tools fight against this.
Share
The concern that AI use without design can shred attention.
Diff
His stance leans minimalist; we propose scaffolded AI use can produce deep work, not destroy it.
Dwarkesh PatelDwarkesh Podcast (formerly Lunar Society)
queued
Claim
Long-form interviews with top AI researchers — N/A; format.
Share
Access to current AI-research thinking before it surfaces in papers.
Diff
Consumer relationship — pick 1-2 episodes per quarter for inbound notes.
Ezra Klein ShowNYT podcast · AI/society episodes
queued
Claim
Mainstream public-interest podcast with frequent AI episodes (Karen Hao, Demis Hassabis, Dario Amodei).
Share
Connection to broader societal AI discourse.
Diff
Mainstream-discourse-shaping; we feed our research-line writing back into that discourse over time.
Tyler CowenMarginal Revolution · Conversations with Tyler
queued
Claim
Economist's takes on AI's distributional + cognitive impact, often contrarian and useful.
Share
Asking "what does this do to people" rather than "what can it do".
Diff
Macro-economic vs individual-cognitive lens.
Ben ThompsonStratechery · paid newsletter
queued
Claim
AI is a platform shift; strategy logic from prior platform shifts applies.
Share
Product-strategy literacy informs our "what is happening in product land" tracking.
Diff
Product-strategy-focused vs research-method-focused.
Latent Space (Swyx)AI engineering podcast
queued
Claim
Practitioner format — N/A.
Share
Stays current on what's deployable.
Diff
Principles-and-evidence layer vs deployable-engineering layer.
Lex Fridman PodcastLong-form AI conversations
queued
Claim
Interview format — N/A.
Share
Broad-public AI discourse.
Diff
Signal-to-noise variable; pick selectively.
晚点 LatePost · 张小珺商业访谈Chinese tech publication + podcast
queued
Claim
Interview / journalism format.
Share
Access to Chinese AI ecosystem thinking, often missing from Western feeds.
Diff
L1 cron cannot reach (no clean RSS); manual L2 promotion for Chinese sources.

Where we actually stand.

After mapping all of the above, the position the literature is most empty at
Empirical, instrumented, public protocols that measure whether a human becomes more capable in AI's absence after AI-assisted work.

Most cognitive-offloading work measures the negative case. Most AI-literacy work writes taxonomies. Most product studies measure usage and satisfaction. Most industry research measures the AI. Most individual essays speculate. The protocol-with-measured-capability-delta position is largely empty. That is where this research line lives.

How this atlas stays alive.

An atlas is not a reading list. It is the cumulative record of what we have engaged with, how engaged we are, and where we differ.

  1. L1 capture (daily). Scheduled remote agent pulls 5–8 narrow high-signal sources, filters by anchor keywords, writes raw daily inbox.
  2. L2 triage (weekly). Scheduled agent + human approval promotes 1–3 candidates per week to full 1-page inbound notes. Each note has an HTML mirror and joins the public corpus.
  3. L3 synthesis (quarterly). Human + Claude write a synthesis: "what we read, what we changed because of it."
  4. Atlas update. New inbound notes are appended to the appropriate bucket above with status flipped to note-written and a link to the inbound markdown.
When you cannot answer "where do we differ?" for an entry, the entry has not really been read.