{ "video": { "id": "I3bpdgFJCUY", "title": "LLM Knowledge Bases: a practical guide — Ben Holmes, Warp", "duration": 1277, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "I3bpdgFJCUY", "title": "LLM Knowledge Bases: a practical guide — Ben Holmes, Warp", "one_liner": "Ben Holmes shows a four-stage pipeline — voice-dictate raw markdown notes, have an agent enrich them with tags/sources/backlinks, generate Karpathy-style wikis on top, then visualize the whole thing — all run on a nightly cloud schedule so you wake up to a fresh personal encyclopedia.", "summary": "Holmes argues that the bottleneck on a useful personal knowledge base is raw material, not organization: get as many scrappy thoughts down as possible via voice dictation (~200 wpm, faster than typing), then let agents do the structuring. He demos an 'enrich note' agent skill that stamps a timestamp, applies tags from a fixed reference list, web-searches the source URL, and inserts backlinks to related notes; then a wiki-generation step based on Andrej Karpathy's LLM-knowledge-base gist that produces browsable people/concepts/sources/organizations pages. Finally he shows the whole thing running unattended on a cloud schedule via Warp's oz.dev, syncing his markdown vault in and out with the Obsidian headless CLI, plus agent-built HTML/Tailwind graph views and habit charts over the notes.", "key_points": [ "Voice dictation is the fastest capture method — roughly 200 words per minute, faster than any non-Olympic typist — and you no longer need to pay for it: Handy is open source and runs a local model on device, Voice Ink is a ~$20 lifetime fee with a hotkey (option/function) and a mobile app.", "Don't self-edit at capture time: 'you need a lot of raw data, you need a lot of raw materials' before an LLM can generate wikis or visualizations, so rambly unformatted transcripts are fine.", "The 'enrich note' skill does four things to a raw note: write an enriched-at timestamp (so later passes can skip already-done notes), add tags, research the source URL with web tools, and find related notes by key-term search to write backlinks at the bottom.", "Tags come from a dedicated reference folder listing the allowed tags, and the skill explicitly instructs the agent to be reluctant to invent new ones — 'Claude loves to get creative' — while permitting an addition if it finds a genuine pattern.", "The wiki layer comes straight from Andrej Karpathy's public gist (searchable as his name + 'wiki'): point an agent at a raw directory plus a focus area and it emits a wiki of sources, concepts, people and organizations, each with backlinks down to the raw notes.", "Demos of the output: a 'latest in AI' wiki that surfaced unexpected entries like a Ralph loop rabbit hole and jazz musician Adam Neely on AI in music, plus a Bible-in-a-year wiki generating per-character entries to keep similar-sounding names straight.", "Scheduling: Codex app automations run locally so 'your laptop has to be cracked open'; Warp's oz.dev runs the same skills in a cloud sandbox on a schedule or trigger (Slack, iMessage). The loop is sync markdown down → run the skill → sync back up, using the Obsidian headless CLI instead of git clone/push/pull.", "Visualizations need no installed tool — he asked an agent to build an HTML + Tailwind graph view over the markdown, with clickable nodes, a restyle-to-space mode, and a GitHub-style chart tracking note-writing as a habit." ], "takeaways": [ "Install a voice dictation tool (Handy if you want free and local, Voice Ink for the hotkey + phone app) and capture notes by talking, on desktop and phone, without formatting them.", "Write enrichment as an agent skill, not a prompt you retype: timestamp each enriched note so reruns are incremental, and give the agent a closed tag list in a reference folder so the taxonomy stays stable.", "Steal Karpathy's wiki gist rather than designing a schema — hand an agent a raw notes directory and the focus area you care about, and let it produce the people/concepts/sources structure.", "Move both skills onto a nightly cloud schedule (oz.dev, or any runner) with the Obsidian headless CLI or a git clone for sync, so enrichment and wiki updates happen while you sleep instead of blocking you for minutes at a time.", "For an overview, just ask an agent to build a clickable graph view in HTML and Tailwind over your markdown — useful for spotting both your real interests and the gaps in your thinking." ], "topics": [ "knowledge-management", "agents", "agent-skills", "voice-dictation", "markdown", "automation", "personal-wikis", "visualization" ], "tools": [ "Warp", "oz.dev (warp.dev/oz)", "Hubble / hub.md", "Obsidian", "Obsidian headless CLI", "Apple Notes", "Wispr Flow", "Handy", "Voice Ink", "Exa", "GLM 5.2", "Claude", "Codex app", "Tailwind", "Docker", "GitHub", "Acquired podcast" ], "quotes": [ { "text": "voice dictation, even though it is pretty awkward to talk into your computer with a bunch of co-workers around, it is the fastest way to get your thoughts onto paper.", "at": "03:10", "url": "https://www.youtube.com/watch?v=I3bpdgFJCUY&t=190s" }, { "text": "I actually instruct the agent to be reluctant to add new tags because Claude loves to get creative.", "at": "07:13", "url": "https://www.youtube.com/watch?v=I3bpdgFJCUY&t=433s" }, { "text": "if you're someone that likes following Wikipedia rabbit holes, it kind of invents a Wikipedia rabbit hole of your own thoughts.", "at": "09:47", "url": "https://www.youtube.com/watch?v=I3bpdgFJCUY&t=587s" }, { "text": "So that way when I come back to my computer in the morning, I wake up to a perfectly fresh wiki that I can review. It's like the daily paper, but it's your own, which is so exciting.", "at": "16:33", "url": "https://www.youtube.com/watch?v=I3bpdgFJCUY&t=993s" } ], "words": 5118 }, "summary_url": "/#I3bpdgFJCUY", "transcript": { "html": "/transcripts/I3bpdgFJCUY.html", "txt": "/transcripts/I3bpdgFJCUY.txt", "vtt": "/transcripts/I3bpdgFJCUY.vtt" } }