(31) From Daily Reading to Cumulative Expertise

Reading needs somewhere to accumulate

A morning spent reading can leave behind a surprisingly weak account of what changed. There are new announcements, revised forecasts and several interesting arguments. By Friday, the reader remembers the theme but struggles to explain which assumption moved, which source supplied the evidence, or why two apparently separate stories belong together.

My earlier article on Karpathy’s LLM wiki and the maintenance problem examined a practical response: let an agent maintain a body of knowledge as new material arrives. His April 2026 proposal separates preserved sources, an evolving wiki and instructions governing its upkeep. The follow-on question concerns the reader. Under what conditions does an increasingly capable archive help its owner develop expertise? [1]

My claim is conditional. Daily reading can become cumulative when new evidence updates an existing explanation and the reader repeatedly reconstructs, challenges and applies that explanation. An agent can handle much of the maintenance. Whether the reader becomes better at reasoning remains a separate outcome to establish.

This ambition has a long history. Vannevar Bush imagined navigable trails through recorded knowledge in 1945. Douglas Engelbart later treated the person, tools and working methods as an integrated system. Pirolli and Card’s analysis of intelligence work distinguished gathering information from organising it into an account that answers a question. A recent Obsidian case study likewise found that researchers’ retrieval strategies shaped how they maintained their notes. The arrangement matters because it determines what work becomes convenient. [3, 4, 5, 6]

Secondbrain, my implementation of this approach, provides the case study here. The companion technical paper documents its actual architecture, including the ingestion skill, provenance rules, search and topic graph. It describes a working knowledge system; it does not report a controlled study of learning.

Keep the event and the explanation together

Secondbrain receives material through two collections. News captures current reporting. Library preserves saved reading, including papers, newsletters, podcasts and longer reports. Those labels describe intake channels. A Library item can be speculative, commercially interested or quickly overtaken by events.

Both feed a shared wiki of source pages, concepts and entities. A source page explains an individual item. A concept page develops an explanation across items. An entity page follows an organisation, person or other named subject. The original captured brief remains available beneath those interpretations; it may itself be a summary rather than the complete original publication.

This arrangement gives a reader two useful views of time. Dated updates preserve how an account developed, while a separate “Current understanding” section states the latest supported picture and its evidence horizon. An old forecast retains its original meaning. Later evidence can revise the current account without making the earlier prediction look retrospectively correct.

The distinction resembles a wider principle in knowledge engineering. W3C’s provenance model records how information was produced, while SKOS separates concepts and their relationships from the words used to label them. Secondbrain uses a smaller Markdown-based design, without claiming compliance with either standard. Stable identities and traceable derivations still solve recognisable problems. [31, 32]

The learning rationale also needs precision. The National Academies describes how organised prior knowledge helps people interpret new information. Chi, Feltovich and Glaser found that physics experts categorised problems through underlying principles more than novices did. A page organised around a mechanism could make those principles easier to revisit; automatically generating that page does not install them in the reader’s head. [2, 9]

A connection should change the question

Consider the AI infrastructure material in Secondbrain. A saved credit-market interview discusses limits on investors’ willingness to absorb repeated borrowing. A saved investment-manager outlook considers how apparently different investments can depend on the same AI spending cycle. The compiled infrastructure topic brings those arguments into contact.

The useful connection concerns the relationship between financing capacity and common exposure. More funding channels can increase the capital available to a project while leaving several investors dependent on the same underlying demand. That is an interpretation to examine against the underlying accounts and contracts. The institutions making those arguments also have commercial interests in the markets they describe.

A reader who follows the connection can ask a more precise question of the next announcement: does this structure move risk to a different holder, change the contractual protection, or reduce dependence on the same economic outcome? The answer requires details beyond the headline funding amount. In the paper, this example is treated as a trace through saved evidence, rather than a fresh verification of every market claim in those sources.

This extends the argument in Beyond Retrieval: Recombination as Productivity Gain. Recombination becomes useful when a connection helps explain a new case. Gentner, Loewenstein and Thompson’s negotiation experiments support the value of comparing cases to extract a transferable relationship. Nesbit and Adesope’s meta-analysis found learning benefits from concept and knowledge maps across varied instructional settings. Neither finding establishes that browsing an automatically generated graph produces the same effect. [14, 15]

Secondbrain’s Pathfinder starts from authored links between topic pages. Shared mentions alone cannot create a connection. It presents explanatory passages and distinguishes the sources explicitly cited there from documents that merely mention both topics. Even so, a route through several pages establishes a reading path, not a causal chain.

The skill decides what survives

The ingestion skill is the operational centre of the system. It tells the agent to read selected captures, create their source pages, update relevant concepts, preserve disagreements and review the resulting prose. It also specifies what must remain stable: source identity, original capture, dates and the distinction between reported claims and editorial interpretation.

These rules become consequential during revision. A forecast must remain a forecast after its target date passes. Repeated coverage of one announcement cannot count as independent corroboration. A partial capture must expose its limitations. Corrections must reach the headings and current summaries that repeat the error, because a careful paragraph beneath a misleading title remains misleading.

There is a cognitive reason to invest in those details. Fazio and colleagues found that repetition could increase perceived truth even when participants possessed relevant knowledge. Brashier and Marsh describe how ease of processing enters truth judgments. Reviews by Lewandowsky and by Ecker explain why misinformation can continue to influence reasoning after correction. A readable archive can preserve errors as effectively as sound explanations. [27, 28, 29, 30]

AI adds another maintenance burden. Research on summarisation faithfulness, citation generation and atomic factual evaluation treats fluent language, supported statements and complete citation coverage as different properties. NIST’s generative-AI profile also calls for verifying sources and testing capability claims. A successful file validator cannot establish that a paragraph faithfully represents its evidence. [36, 37, 38, 42]

The same principle underlies briefing an agent with explicit acceptance criteria. The skill defines the work and its review boundary. It cannot make the model’s editorial judgment infallible, and the presence of a review instruction is not proof that every existing page has received an adequate review.

Let the reader do the difficult part

A practical reading routine begins before opening the updated explanation. Write what you currently believe, identify the mechanism, and predict what the new evidence might change. Then inspect the topic page and its sources. Explain the revision in your own words, including the assumption that would make it fail.

This is a proposed way to use Secondbrain, rather than a learning feature already implemented in the app. Its basis comes from several distinct findings. Roediger and Karpicke showed benefits from retrieval on delayed tests. Cepeda and colleagues synthesised evidence for spaced practice. Dunlosky’s review rated practice testing and distributed practice highly. The Institute of Education Sciences recommends deep explanatory questions and connecting concrete examples to abstract concepts. [7, 16, 17, 21]

Generating the explanation matters. Chi’s work on self-explanation examined how students related worked examples to principles. The ICAP framework distinguishes constructive engagement from merely manipulating or receiving information. Freeman’s synthesis of undergraduate STEM studies supports active learning, while Hattie and Timperley show why the content and level of feedback matter. These findings inform the routine; they come from settings different from an adult’s personal research archive. [12, 18, 22, 23]

A graph can support that routine by proposing comparisons and exposing missing steps. Yet Karpicke and Blunt found retrieval practice superior to elaborative concept mapping in their science-text experiments. An attractive map is therefore an invitation to think, with no entitlement to count as thinking completed. [8]

The current Ask feature makes the division of labour unusually visible. It ranks matching compiled pages and returns attributed passages without calling a language model. AI performs editorial work during ingestion; query-time retrieval is deterministic. The reader still has to decide whether the selected passages answer the question and whether the supporting evidence warrants their use.

More assistance can mean less practice

The strongest objection is that this machinery may help its user avoid the effort through which expertise develops. Fisher, Goddu and Keil found that searching online could inflate people’s assessments of their internal knowledge. Rozenblit and Keil documented overconfidence in explanatory understanding. Reviews by Bjork and colleagues, and by Soderstrom and Bjork, distinguish immediate fluency and performance from durable learning. [11, 13, 20, 26]

The risks are compatible with useful assistance. Risko and Gilbert describe cognitive offloading as a way to reduce a task’s internal demands. Storm and Stone found that saving information could improve memory for subsequently studied material under particular conditions. The important design question is where the saved effort goes. It might fund deeper comparison, or disappear into another hour of consumption. [10, 25]

Evidence about AI itself is mixed and dependent on the task. Bastani and colleagues’ high-school mathematics experiment found that unguarded assistance could damage later unaided performance, while instructional safeguards mitigated that harm. Kestin and colleagues reported positive results from a deliberately designed college physics tutor. Wang and colleagues’ Tutor CoPilot trial found benefits from supporting human tutors. None evaluated this wiki or establishes its effect on professional expertise. [34, 43, 44]

Lee and colleagues’ survey of knowledge workers adds a different warning: greater confidence in AI was associated with less reported critical-thinking effort. That is self-reported association, not proof of lasting cognitive decline. UNESCO’s guidance places human agency and pedagogical design at the centre of educational use, a useful constraint on claims that a larger archive automatically means a more capable owner. [33, 41]

For a narrow, stable subject, a small set of carefully read sources and regular practice may be enough. Maintaining thousands of pages has a cost. Secondbrain earns that cost only where revisiting changing evidence and connecting recurring topics improves work the reader needs to do.

Measure the explanation that survives

The technical alternatives remain credible. Retrieval-augmented generation can assemble evidence at question time. GraphRAG uses a graph and generated community summaries to answer broader questions. Long-context models can process substantial material, although Liu and colleagues demonstrated position-sensitive retrieval failures in the models they studied. These approaches require evaluation on the intended task, rather than dismissal through an architectural slogan. [35, 39, 40]

Recent work also studies maintenance directly. Huerta’s 2026 preprint on streaming knowledge compilation examines what to retain as information changes under a fixed budget. Its cache-oriented design differs from Secondbrain’s maintained Markdown pages. The shared problem is deciding which incoming evidence deserves attention before the next question arrives. [45]

For Secondbrain, I would measure source fidelity, the accuracy of dated explanations and the usefulness of connections alongside maintenance time. Learning would require a separate comparison: after a delay, can a reader explain a mechanism, identify a misleading analogy and apply it to an unfamiliar case? Deslauriers and colleagues’ classroom study shows why felt learning should be measured separately from tested learning. Kahneman and Klein likewise emphasise the quality of an environment and opportunities for feedback when assessing expertise. [19, 24]

The next useful experiment is small enough to run. Choose a recurring subject, record an explanation before consulting the archive, revise it against the evidence, and revisit it weeks later using a new case. Keep the original answer. If the explanation becomes more accurate, better qualified and more useful outside the examples already read, the system has supported something worth accumulating.

Sources

Foundations

1. Andrej Karpathy, "LLM Wiki" https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

2. National Academies of Sciences, Engineering, and Medicine, "How People Learn II: Learners, Contexts, and Cultures" https://www.nationalacademies.org/read/24783/chapter/7

3. Juliana Jansen Ferreira et al., "How People Manage Knowledge in their "Second Brains"- A Case Study with Industry Researchers Using Obsidian" https://arxiv.org/abs/2509.20187

4. Peter Pirolli and Stuart Card, "The Sensemaking Process and Leverage Points for Analyst Technology as Identified Through Cognitive Task Analysis" https://www.researchgate.net/profile/Peter-Pirolli/publication/215439203_The_sensemaking_process_and_leverage_points_for_analyst_technology_as_identified_through_cognitive_task_analysis/links/02bfe50f09ca94efc0000000/The-sensemaking-process-and-leverage-points-for-analyst-technology-as-identified-through-cognitive-task-analysis.pdf

5. Vannevar Bush, "As We May Think" https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881/

6. Douglas C. Engelbart, "Augmenting Human Intellect: A Conceptual Framework" https://dougengelbart.org/pubs/augment-3906-Framework.html

Learning

7. John Dunlosky et al., "Improving Students’ Learning With Effective Learning Techniques: Promising Directions From Cognitive and Educational Psychology" https://www.psychologicalscience.org/publications/journals/pspi/learning-techniques.html

8. Jeffrey D. Karpicke and Janell R. Blunt, "Retrieval practice produces more learning than elaborative studying with concept mapping" https://pubmed.ncbi.nlm.nih.gov/21252317/

9. Michelene T. H. Chi, Paul J. Feltovich and Robert Glaser, "Categorization and Representation of Physics Problems by Experts and Novices" https://onlinelibrary.wiley.com/doi/10.1207/s15516709cog0502_2

10. Evan F. Risko and Sam J. Gilbert, "Cognitive Offloading" https://pubmed.ncbi.nlm.nih.gov/27542527/

11. Matthew Fisher, Mariel K. Goddu and Frank C. Keil, "Searching for Explanations: How the Internet Inflates Estimates of Internal Knowledge" https://cpb-us-w2.wpmucdn.com/campuspress.yale.edu/dist/c/259/files/2020/05/fisher2015.pdf

12. Michelene T. H. Chi and Ruth Wylie, "The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes" https://education.asu.edu/sites/default/files/lcl/chiwylie2014icap_2.pdf

13. Nicholas C. Soderstrom and Robert A. Bjork, "Learning Versus Performance: An Integrative Review" https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/11/soderstorm_ra_learningvsperformance.pdf

14. Dedre Gentner, Jeffrey Loewenstein and Leigh Thompson, "Learning and Transfer: A General Role for Analogical Encoding" https://groups.psych.northwestern.edu/gentner/papers/GentnerLoewensteinThompson03.pdf

15. John C. Nesbit and Olusola O. Adesope, "Learning With Concept and Knowledge Maps: A Meta-Analysis" https://www.sfu.ca/~jcnesbit/research/NesbitAdesope2006.pdf

16. Nicholas J. Cepeda et al., "Distributed practice in verbal recall tasks: A review and quantitative synthesis" https://pubmed.ncbi.nlm.nih.gov/16719566/

17. Henry L. Roediger and Jeffrey D. Karpicke, "Test-enhanced learning: taking memory tests improves long-term retention" https://pubmed.ncbi.nlm.nih.gov/16507066/

18. Michelene T. H. Chi et al., "Self-Explanations: How Students Study and Use Examples in Learning to Solve Problems" https://onlinelibrary.wiley.com/doi/10.1207/s15516709cog1302_1

19. Daniel Kahneman and Gary Klein, "Conditions for Intuitive Expertise: A Failure to Disagree" https://bear.warrington.ufl.edu/brenner/mar7588/Papers/kahneman-klein-2009.pdf

20. Robert A. Bjork, John Dunlosky and Nate Kornell, "Self-regulated learning: beliefs, techniques, and illusions" https://pubmed.ncbi.nlm.nih.gov/23020639/

21. Institute of Education Sciences, "Organizing Instruction and Study to Improve Student Learning" https://ies.ed.gov/ncee/wwc/PracticeGuide/1

22. Scott Freeman et al., "Active learning increases student performance in science, engineering, and mathematics" https://www.pnas.org/doi/full/10.1073/pnas.1319030111

23. John Hattie and Helen Timperley, "The Power of Feedback" https://educacion.udd.cl/files/2018/04/The-Power-of-Feedback.pdf

24. Louis Deslauriers et al., "Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom" https://dunkin.eeb.ucsc.edu/images/documents/measuring_actual_learning_versus_feeling_of_learning_in_response_to_active_engasgement_in_a_classroom_2019.pdf

25. Benjamin C. Storm and Sean M. Stone, "Saving-Enhanced Memory: The Benefits of Saving on the Learning and Remembering of New Information" https://journals.sagepub.com/doi/abs/10.1177/0956797614559285

26. Leonid Rozenblit and Frank Keil, "The misunderstood limits of folk science: an illusion of explanatory depth" https://onlinelibrary.wiley.com/doi/abs/10.1207/s15516709cog2605_1

Evidence and maintenance

27. Nadia M. Brashier and Elizabeth J. Marsh, "Judging Truth" https://www.annualreviews.org/content/journals/10.1146/annurev-psych-010419-050807

28. Lisa K. Fazio et al., "Knowledge does not protect against illusory truth" https://scholars.duke.edu/publication/1090926

29. Ullrich K. H. Ecker et al., "The psychological drivers of misinformation belief and its resistance to correction" https://www.nature.com/articles/s44159-021-00006-y

30. Stephan Lewandowsky et al., "Misinformation and Its Correction" https://www.psychologicalscience.org/journals/pspi/1529100612451018/

31. W3C, "PROV-DM: The PROV Data Model" https://www.w3.org/TR/prov-dm/

32. W3C, "SKOS Simple Knowledge Organization System Reference" https://www.w3.org/TR/skos-reference/

AI and evidence

33. Hao-Ping Lee et al., "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers" https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf

34. Hamsa Bastani et al., "Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics" https://hamsabastani.github.io/education_llm.pdf

35. Nelson F. Liu et al., "Lost in the Middle: How Language Models Use Long Contexts" https://aclanthology.org/2024.tacl-1.9/

36. Tianyu Gao et al., "Enabling Large Language Models to Generate Text with Citations" https://aclanthology.org/2023.emnlp-main.398/

37. Sewon Min et al., "FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation" https://aclanthology.org/2023.emnlp-main.741/

38. Joshua Maynez et al., "On Faithfulness and Factuality in Abstractive Summarization" https://aclanthology.org/2020.acl-main.173/

39. Patrick Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" https://arxiv.org/abs/2005.11401

40. Darren Edge et al., "From Local to Global: A Graph RAG Approach to Query-Focused Summarization" https://arxiv.org/abs/2404.16130

41. UNESCO, "Guidance for generative AI in education and research" https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research

42. NIST, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile" https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf

43. Greg Kestin et al., "AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting" https://www.nature.com/articles/s41598-025-97652-6

44. Rose E. Wang et al., "Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise" https://arxiv.org/abs/2410.03017

45. Juan M. Huerta, "Streaming Knowledge Compilation: Proactive Materiality-Scored Pinning for Time-Evolving LLM Wikis" https://arxiv.org/abs/2606.09877