An LLM wiki changed how I work

Casey Newton · rss · 2026-08-19 source ↗ Reader 🖥️

practical · mixed — Primarily a first-person narrative/expository account of Newton's annual productivity-tool review, organized around a practical verdict (what to keep doing, what to stop, what's new) rather than a theoretical claim; the LLM-wiki section carries a light argumentative thread about whether the system is worth its maintenance cost and whether it can be recommended to others. The trailing 'Following' news-roundup section (OpenAI safety changes, Meta addiction trial) is a separate newsletter segment not analyzed here, since it is unrelated in subject to the titled essay.

Structure

Key terms

• LLM wiki
  A folder of interlinked Markdown files, compiled and continuously updated by an LLM from source documents (Newton's own archive plus daily-clipped reading), browsable in Obsidian and queryable via a plugin (Claudian) — following the process Andrej Karpathy described.
• infohazard
  An idea so immediately compelling on exposure that simply learning of it commits the hearer to weeks of costly follow-through, with no guarantee of payoff — used self-mockingly about his own reaction to Karpathy's tweet.
• blips
  Newton's own term (credited to Andy Matuschak's influence) for hand-created Capacities pages that gather links and notes around an emerging, not-yet-fully-formed story or topic over time, to be revisited later.

Propositions

• explicitOf everything Newton tried this year to get better at the deskbound parts of his job, the LLM wiki has been the single most useful change.conf: high
  supportsLLM Knowledge Bases · p1 — Newton's month-long practitioner report — roughly 1,440 auto-generated wiki pages, one running past 12,000 words — is real-world confirmation that LLMs can sustain the autonomous-compilation workflow Karpathy described, at meaningful personal scale.
• explicitcorroboratedThe LLM wiki requires a fair degree of ongoing maintenance, comparable to a vintage sports car, and is not obviously the easiest way to build a personal knowledge base.conf: high
  refinesWow, this tweet went very viral! · p4 — Karpathy's idea file claims the human 'rarely writes or edits the wiki directly.' Newton's experience holds for wiki *content* — he doesn't hand-write entries — but not for the wiki *system*: he actively debugs a breaking pipeline and had to bring in a second LLM to rewrite the output style, so 'rarely edits' understates the ongoing human maintenance burden.
  supports ←TBM 439: Day At The Gig... · p7 — Both pieces converge on knowledge repositories requiring continuous upkeep rather than one-time construction — Cutler frames it as context eroding daily, Newton frames it as maintenance comparable to a vintage sports car.
• explicitThe wiki automates and supersedes Newton's prior manual 'blips' system: it generates new concept pages on its own from each day's reading, without him having to notice a pattern and create the page by hand.conf: high
  supportsLLM Knowledge Bases · p4 — Newton's daily auto-generated concept pages (e.g. 'AI and Congress,' 'tokenmaxxing') that surface story ideas on their own are a concrete instance of the accretive, self-surfacing knowledge base Karpathy describes as the goal.
• explicitNewton manually verifies wiki-derived claims against original source material before writing or speaking publicly, specifically to catch possible hallucinations.conf: high
• explicitTwo earlier partial solutions to the same memory problem — Notion's agentic search over saved links, and the manual Capacities 'blips' workflow — were abandoned mainly because they demanded continuous personal effort that Newton could not sustain as a habit.conf: high
• implicitPersistence through app-switching and renewed subscriptions is a better test of a tool's real productivity value than initial enthusiasm.conf: medium
• explicitOutput style and quality differ enough between LLMs to be worth actively managing: Claude produced a 'hyper-compressed, borderline-unreadable' house style for the wiki, so Newton had a different model (GPT-5.6 Sol) rewrite much of the system to approximate AP style.conf: high
• implicitThe bottleneck in Karpathy-style LLM wikis has shifted from the difficulty of initial setup to ongoing operational upkeep — pages growing too long and needing compaction, and silent pipeline failures that have to be diagnosed and fixed.conf: medium
• explicitNewton cannot yet confidently recommend the LLM wiki to a general reader, despite it being personally the most valuable change of his year, because it feels too specific to him and too clunky.conf: high

Arguments

Narratives

• Using the wiki daily during the multi-week OpenAI/Hugging Face agentic-breach story to refresh his memory of a complicated, fast-evolving timeline before writing or podcasting about it.
  → Supports the proposition that the wiki delivered concrete productivity value (p1) and grounds the argument that external memory beats reconstructing context from scratch under a fast-moving, information-dense story.
• The wiki's auto-generated 'AI and Congress' concept page surfaced a cluster of recent stories Newton had half-forgotten, which he then pitched as a podcast segment.
  → Supports the proposition that the wiki automates the story-idea-generation function his old 'blips' system used to serve (p3).
• The abandoned Capacities 'blips' system: Newton built a randomized-blip widget into his daily journal template to remind himself of neglected blips, which worked for a while but ultimately failed as the number of blips grew beyond what he could consistently track.
  → Illustrates why a purely manual, willpower-dependent knowledge-tracking system doesn't scale, motivating the shift to an LLM-automated replacement and grounding the implicit premise that discipline-dependent systems degrade as scope grows (relevant to p5, p8).
• Newton seeded the wiki with his entire Platformer archive (since 2020), producing pages like a 12,000-word Meta page with 1,300 internal links, and found himself losing 'more than one afternoon' just browsing it the way he'd browse Wikipedia.
  → Illustrates unexpected personal/reflective value beyond narrow lookup utility, and supports the claim (p1) that the wiki's value to him is deep and idiosyncratic rather than generically demonstrable.

Questions addressed

• ✅ Is an LLM-compiled personal wiki actually more useful in practice than existing knowledge-management apps for someone who needs to track recurring topics across years of work?
  Yes, for Newton specifically — he states it has been 'easily the most useful' productivity change of his year, superseding Notion's agentic search and his manual Capacities blips system.
• ✅ How does a user keep a large personal LLM-maintained knowledge base honest — free of confidently-stated errors — once it becomes a working reference used in professional output?
  Partially: Newton describes always reopening original source documents before writing or speaking to confirm nothing he's about to say was hallucinated, rather than trusting the wiki's summaries outright.
• ⭕ Should Newton, or readers generally, adopt this LLM-wiki system?
  He explicitly declines to give a confident recommendation, calling it 'too specific... and too clunky' despite it being personally his most valuable change of the year.
• ⭕ As the wiki's scope keeps growing (more pages, longer pages, more source material added daily), does its maintenance cost eventually outpace the value it returns?

Gaps

• incompleteNewton never quantifies the actual time cost of building and maintaining the wiki pipeline — the script that files new clippings, the process that 'stops running and has to be fixed,' the periodic page compaction — so the claim that the wiki nets out to a productivity win is asserted, not measured, against that upkeep cost.
  Read next: A time-tracked account of LLM-wiki maintenance overhead versus time saved would let this be evaluated; compare with other practitioner reports on sustaining a Karpathy-style knowledge base at multi-month scale.
• incompleteThe piece never addresses privacy or data-handling exposure from routing years of reporting material, sources, and possibly off-the-record notes through third-party LLM APIs (Claude, GPT) as part of the daily ingestion pipeline — a notable omission given Newton's own reporting elsewhere on AI notetakers' data practices.
  Read next: Newton's own piece on Granola's data-retention and training-use practices (extract-newton-granola-notetaker-surveillance) raises the same category of question he doesn't ask about his own wiki.
• illogicalNewton faults Notion's agent for 'often fail[ing] to cite its sources without additional prompting,' implying citation reliability is a real bar for a research tool — but he never confirms the wiki itself structurally guarantees sourcing beyond his own manual habit of reopening original documents; the standard he explicitly holds a competitor to isn't shown to be met by his own system in a built-in, non-optional way.
  Read next: Compare against tools that make citation-to-source a structural guarantee rather than a user habit (e.g. retrieval systems with enforced provenance links).
• incompleteThe essay tours five overlapping tools (Raycast, Capacities, Recall, Notion, the LLM wiki) without asking why his personal information system requires this much fragmentation in the first place, or whether the LLM wiki is best understood as a genuine replacement for the others versus one more layer added on top.
  Read next: Critiques of personal-knowledge-management tool sprawl and note-taking-app churn — including Newton's own earlier annual productivity posts (2023, 2024, 2025), which aren't in this corpus but are referenced in the piece.

Closes gaps elsewhere

• Wow, this tweet went very viral! · g2 — Same evidence as above (operation at ~1,440 pages / a 12,000-word single page) is a practitioner data point at roughly an order of magnitude beyond the ~100-article baseline the original claim ('RAG becomes unnecessary at this scale') rested on, and reports it continuing to work without a dedicated vector-search layer.
• Wow, this tweet went very viral! · g3 — Newton supplies a concrete answer to 'how does the user verify LLM-generated wiki content is correct': he routinely reopens original source documents before writing or speaking, specifically to catch hallucinations, rather than trusting wiki summaries outright.
• LLM Knowledge Bases · g3 — Newton reports real-world operation well past the '~100 articles / ~400K words' scale the original idea was tested at — about 1,440 wiki pages, one exceeding 12,000 words with 1,300 internal links, sustained for over a month — showing the approach holds up at that larger scale but starts requiring active page compaction and pipeline repair, i.e. rising maintenance cost rather than outright breakdown.

Related documents

• AI 101: Agentic Vector Databases – What Is That?
  Agent Memory Knowledge Base Retrieval & RAG
• Paweł Huryn's Post
  Agent Memory Knowledge Base Retrieval & RAG
• Adrien Taravant's Post
  Agent Memory Knowledge Base
• AI Coding Agents Get a Stack Overflow of Their Own
  Agent Memory Knowledge Base
• Claude Code Source Leak Megathread
  Agent Memory Vibe Coding
• Inside a PhD Student's Research Workflow | Studying Parasocial Relationships with Alvina Lai
  Retrieval & RAG AI Workflow Patterns
• Graphify: Unifying Codebase Context to Streamline Agentic Software Engineering
  Knowledge Base Retrieval & RAG
• Podcast: Strands Agents with Clare Liguori
  Agent Memory AI Workflow Patterns
• Wow, this tweet went very viral!
  Knowledge Base Retrieval & RAG
• Someone just built the exact tool Andrej Karpathy said someone should build
  Knowledge Base Retrieval & RAG