From Notetaking to Neuralink (Contrary Research)
TL;DR
A May-2023 Contrary Research deep dive (Alex Banks) that supplies the pre-agent-era problem statement beneath the vault's second-brain cluster: information volume now outruns any unaided mind (the report cites knowledge doubling every ~13 months, against every 100 years before 1900), and manual PKM — the Capture-Storage-Retrieval Pipeline pipeline done by hand in Notion/Roam/Obsidian — decays into learn → write down → forget because of three structural limits: alignment (notes lose the context that made them valuable), labor intensity (capture and organization effort deters use), and fragmentation (link webs accumulate faster than anyone analyzes them). The proposed fix is the move from static to dynamic retrieval (Dynamic Retrieval): an AI second brain that surfaces what's relevant from context and current task instead of waiting for the human to remember to ask, integrated across the user's tool ecosystem rather than standing alone. The report grounds the whole category in the Extended Mind thesis (citing Clark & Chalmers' paper directly), scopes what 2023-era LLMs can contribute via the incremental-vs-discontinuous task split from "Sparks of AGI" (Incremental vs Discontinuous Tasks) — connection-making over stored notes is incremental, hence LLM-suitable — and extrapolates the trajectory to brain-computer interfaces (Neuralink) as the eventual direct interface to the knowledge store. Read from 2026, it is upstream lineage: the "true copilot for the mind" it calls for is what the vault's later agent-era sources describe being built.
Concepts introduced
- Capture-Storage-Retrieval Pipeline — the three-component decomposition of every PKM system (capture → storage → retrieval) and the three limits that break it when done manually.
- Dynamic Retrieval — static retrieval waits for the user to recall and search; dynamic retrieval surfaces information from context and current task; the report's core claim about why AI changes the second brain.
- Incremental vs Discontinuous Tasks — the report's borrowed taxonomy (from "Sparks of AGI") for which intellectual tasks autoregressive LLMs are suited to; routes what to delegate to the model.
Held, not dropped (touched, no page yet — spin out on demand):
- knowledge-doubling curve / half-life of knowledge — Fuller's doubling estimates and Machlup's knowledge half-life; framing statistics, no operational content here.
- brain-computer interfaces — the Neuralink/BrainGate/CTRL-Labs endpoint of the trajectory; out of the vault's current operational scope.
- UX as adoption driver — same model, different interface: Playground didn't go viral, ChatGPT did; one sharp observation, not yet a treated concept.
- chunking — the chess-master memory strategy the report contrasts with tool-assisted synthesis; passing reference.
- linked thinking / networked notes — Nick Milo's "linking your thinking," Roam's productized knowledge graph; already implicitly covered by the wikilink practice inside AI Second Brain, not distinct enough here to warrant its own page.
Key claims
- The report states human knowledge doubled every ~100 years until 1900 and has recently been doubling every ~13 months (citing Buckminster Fuller's knowledge-doubling curve). (observation — the source's framing statistic; check-worthy) → Capture-Storage-Retrieval Pipeline
- The extended-mind thesis explains why note-taking expanded rather than degraded cognition (contra Socrates' prediction): developing processes for recalling information beats memorizing it, and the phone in your pocket outsources storage constraints, expanding overall cognitive ability. (principle — as asserted by the report, citing Clark & Chalmers) → Extended Mind
- "Your mind is for having ideas, not holding them" (David Allen, quoted by the report): working memory is small and long-term memory decays, so holding should be offloaded to an external system. (principle — as asserted) → AI Second Brain
- Every second brain involves three components: information capture, information storage (synthesis/extraction), and information retrieval (putting insights into action). (principle — the report's decomposition) → Capture-Storage-Retrieval Pipeline
- Manual knowledge management fails along three buckets: alignment (context-dependent value is lost by capture time), labor intensity (active tagging/organizing deters capture and adds cognitive load), and fragmentation (knowledge-graph link webs go unexplored). (observation — the report's failure analysis of 2023-era PKM tooling) → Capture-Storage-Retrieval Pipeline
- Static stores decay into learn → write it down → forget about it; retrieval that depends on the human remembering to ask reimports the very limit the store was meant to escape. (principle — as asserted) → Dynamic Retrieval
- An AI-powered second brain enables dynamic retrieval — adapting to the user's needs and presenting relevant information based on context and current task — in contrast to static tag/link/search stores. (principle — as asserted; the report's central thesis) → Dynamic Retrieval
- The report cites Coveo research that knowledge workers spend an average of 3.6 hours per day searching for information. (observation — the source's cited statistic; check-worthy) → Dynamic Retrieval
- Interoperability is critical: a second brain integrated with email, calendar, project and communication tools becomes a unified knowledge ecosystem rather than a standalone static tool. (best_practice — context: the report's 2023 product thesis for AI-native PKM; "best" where the store is meant to act on the user's behalf, not just archive) → Dynamic Retrieval
- Intellectual tasks split into incremental (solved gradually, word-by-word: summaries, factual Q&A) and discontinuous (requiring a creative leap or reframing: jokes, new hypotheses); the report states GPT-4-class autoregressive models are strong at the former and weak at the latter (citing "Sparks of AGI"). (observation — the source's 2023 characterization of model capability; check-worthy and dated) → Incremental vs Discontinuous Tasks
- Making incremental connections across a large store of existing notes is exactly the task shape LLMs could be exceptional at — so point the model at connection-making over your store, not at creative leaps. (best_practice — context: 2023-era autoregressive models; the report's routing advice for what to delegate) → Incremental vs Discontinuous Tasks
- Value increases when LLMs are treated as a thinking engine over your knowledge base — suggesting perspectives, identifying gaps, connecting unrelated pieces — rather than as a text-generation tool; the relationship compounds as more information is stored. (principle — as asserted) → AI Second Brain
- Operating with an AI-powered second brain will change the biological brain: the human's skills shift toward asking probing questions, identifying relevant patterns, and drawing connections — co-adapting to the technological appendage. (observation — the report's forward-looking claim about person+scaffold co-adaptation) → Cognitive Scaffolding
- The report cites Goldman Sachs (2023) that generative-AI productivity gains could raise global GDP by 7% over the coming decade. (observation — the source's cited projection; check-worthy) → Dynamic Retrieval
- Brain-computer interfaces (Neuralink's implanted electrodes; non-invasive approaches like CTRL-Labs) could eventually provide direct access to the knowledge repository and real-time human-AI communication — the endpoint of the second-brain trajectory. (observation — the source's speculative extrapolation; held theme) → Extended Mind
Why this builds on the second-brain cluster
The vault's existing AI Second Brain page is written from the agent era: markdown stores, routing files, extraction skills — the solution pattern, sourced from 2025-26 practitioners. This 2023 report is the layer beneath it: the clearest statement in the vault of why that pattern exists — manual Capture-Storage-Retrieval Pipeline breaks on alignment, labor, and fragmentation, and static retrieval reimports the limits of the unaided mind. Its "dynamic retrieval + interoperability" prescription is, read from 2026, a prediction the cluster's later sources describe fulfilling (Hermes' gateways, agent- maintained wikis, context routing) — the report even anticipates the agent framing, describing 2023's early autonomous agents as the mechanism. Secondary stances: it corroborates Extended Mind as an independent, non-philosophy source (investor research) grounding the second-brain category in Clark & Chalmers' paper — a third source lineage now invokes the thesis — and its person+scaffold co-adaptation claim lands on Cognitive Scaffolding. One dated tension worth noting rather than resolving: the report assumes dynamic retrieval requires AI product machinery, while AI Second Brain holds that "just files + routing" suffices for the retrieval test — the vault's later sources resolve this by putting the intelligence in the agent reading the files, not in the store.