Almost Timely News: 🗞️ Creating a Keynote Speaker Second Brain  🔗 **Source:** [Almost Timely News: 🗞️ Creating a Keynote Speaker Second Brain](https://www.christopherspenn.com/2026/09/almost-timely-news-%f0%9f%97%9e%ef%b8%8f-creating-a-keynote-speaker-second-brain-2026-09-13/?pk_campaign=feed&pk_kwd=almost-timely-news-%25f0%259f%2597%259e%25ef%25b8%258f-creating-a-keynote-speaker-second-brain-2026-09-13) > [!note]+ Document Notes > tags: > - second-brain > - local-ai > - retrieval-augmented-generation > - sqlite > - knowledge-management > - public-speaking > author: Christopher S. Penn > source: https://www.christopherspenn.com/2026/09/almost-timely-news-creating-a-keynote-speaker-second-brain-2026-09-13/ > reference: https://www.christopherspenn.com/2026/09/almost-timely-news-creating-a-keynote-speaker-second-brain-2026-09-13/ > date created: 2026-09-13 > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > <!-- deep-read:sections start --> ## 🏷️ Title Almost Timely News: Creating A Keynote Speaker Second Brain (2026-09-13) — Christopher S. Penn's build log for a local-AI slide catalog that doubles as a template for any organization's private knowledgebase. ## 🧭 Table of Contents 1. Title 2. Table of Contents 3. References (in-page) 4. Summary 5. Main Content and Notes 6. References (primary sources) 7. Paraphrase (section by section) 8. Newsletter (WordPress-focused) 9. Action Items 10. Hypotheses 11. Methodology 12. Questions 13. Future Research 14. Implications 15. Learning Styles Assessment 16. Accessibility Assessment 17. DEIB Assessment 18. Logical Validity 19. DARVO Analysis 20. Conclusion 21. Refactor Appendix ## 🔗 References (in-page) Links the author placed inside the main essay (boilerplate, sponsor, merch, and job links omitted): - Wil Reynolds working-in-public video (LinkedIn short link): https://lnkd.in/p/dSBpGjcz - Katie Robbert: https://www.katierobbert.com - 5P Framework by Trust Insights: https://www.trustinsights.ai/5p-framework - LibreOffice: https://www.libreoffice.org/ - Qwen3.6 35B A3B on Hugging Face: https://huggingface.co/Qwen/Qwen3.6-35B-A3B - oMLX: https://omlx.ai/ - SQLite: https://sqlite.org/index.html - Content authenticity / EU disclosure post: https://www.trustinsights.ai/blog/2023/08/disclosure-of-ai-and-protection-of-copyright/ - YouTube version of the issue: https://youtu.be/3S7B8r7GzWk - MP3 version: https://traffic.libsyn.com/cspenn/almost-timely-newsletter-2026-09-13.mp3 - Related ICYMI item (local AI for non-technical people): https://www.trustinsights.ai/blog/2026/09/in-ear-insights-the-non-techies-journey-to-local-ai/ Not linked in the body but named: sqlite-vec, FTS5, BM25, reciprocal rank fusion, LiteLLM, GLM-5.3, Claude Code, ChatGPT Work, Claude Cowork, PostgreSQL, Parakeet, DeepSeek v4.1 Flash, OpenCode. ## 📝 Summary **What happened.** Penn describes, step by step, how he built a searchable catalog of roughly 7,000 slides from two years of keynotes and workshops, then wired an AI coding tool to that catalog so it can assemble a draft deck from a session description, flag content gaps, and leave placeholders for new slides. The pipeline: (1) plan with the 5P Framework (purpose, people, process, platform, performance); (2) convert every PPTX to images with a LibreOffice `soffice` shell script, no AI needed; (3) have a locally run vision-language model (Qwen3.6 35B A3B on oMLX, self-quantized) write a structured record per slide — verbatim text, visual elements, one-sentence summary, keywords, slide type, domain, source deck, slide number — into SQLite with sqlite-vec for embeddings and FTS5/BM25 for lexical search, fused with reciprocal rank fusion; (4) hand the 5P brief to Claude Code, answer its questions (including hashing for de-duplication), and let it build the deck-assembly application. He is explicit that the system retrieves and arranges his existing slides; it does not author new ones. **Why it matters.** The closing section reframes the personal project as a general recipe: the same conversion-index-hybrid-search pattern applies to Dropbox, OneDrive, or SharePoint corpora, and it is how an organization can build a private, local-first knowledgebase without renting cloud AI or handing proprietary material to a vendor. For a personal-knowledge-management practitioner with an Obsidian vault, the piece is a concrete argument that the retrieval layer (hybrid BM25 + vectors over structured metadata) matters more than the model, and that a cheap local model tuned against a frontier model's output on a small sample is good enough for indexing. **Stakeholders.** Speakers and trainers with large slide libraries; marketing and analytics teams asked to "build a knowledgebase"; PKM practitioners; open-source maintainers whose project archives (docs, Slack, meeting notes) are similarly unindexed; privacy- and governance-conscious organizations. **Risks.** "Zero additional dollars" ignores electricity, hardware, and 12 hours of machine time plus tuning labor; a 10-slide calibration sample is small; quality claims are self-assessed; the closing privacy argument rests on an unresolved news anecdote; the approach assumes Apple-silicon hardware and shell comfort that many readers lack. **Bottom line.** A credible, reproducible pattern for a local-first second brain: plan with 5P, pre-process with classical tools, index with a small local VLM, store in SQLite with hybrid retrieval, and keep the human as the author. The privacy framing is persuasive but argued more by anecdote than evidence. ## 📖 Main Content The essay ("What's On My Mind") is the substantive core; the surrounding newsletter frame (course plug, poll, ICYMI, merch, jobs, events, disclosures, AI-generated glossary) is skipped here except where the glossary clarifies a term. **Part 1, Planning.** Penn credits Wil Reynolds' public build of a cosine-similarity slide finder as the spark, then explains why his situation differs: thousands of mostly image-based slides across many topics. He runs the 5P Framework. Purpose: a meta-index enabling fast assembly of new decks from a session description plus an interactive interview, with gap detection. People: himself as builder, audiences as beneficiaries. Process: replace "copy the last similar deck" with retrieval by theme, topic, and keyword, flagging stale slides. Platform: local AI (Qwen) is sufficient for indexing; a frontier model builds the scaffolding. Performance: a working catalog is "done" for now; time saved and audience satisfaction are the long-run measures. **Part 2, Slides Prep.** He distinguishes text-only LLMs, vision-language models, and omni models by cost, and chooses to rasterize slides so a smaller VLM can do the work. Conversion uses LibreOffice's `soffice` CLI in a short shell script; it finished in under 15 minutes. He argues that having an agentic tool do the conversion would burn usage limits to do the same thing, and that a reusable script is the better asset. **Part 3, Cataloging.** He defines the schema up front (eight fields) because the purpose dictates the data. Rather than narrating slides as Reynolds did with 304, he feeds each JPEG to Qwen3.6 35B A3B under oMLX with a custom 4/8-bit quantization, Lightning MTP, and Apple ANE prefill. The run took almost 12 hours and about 20 million tokens. Quality was tuned by comparing Qwen's output on 10 slides against GLM-5.3 and iterating prompts for about 15 minutes, made easy by LiteLLM model swapping. Storage is SQLite plus sqlite-vec (vectors) and FTS5 (BM25 lexical search); the two are fused with reciprocal rank fusion, which he presents as a custom RAG that avoids both the embedding bottleneck and the exact-phrase trap. **Part 4, Using the Catalog.** The 5P brief goes to Claude Code, which interviews him (hashing for duplicates among the questions) and produces, after about seven hours, an application that takes a session description and other text inputs, reads the catalog, identifies gaps, assembles a deck, and inserts placeholders. He insists the tool must not create slides, because he must know what is in his own deck. **Part 5, So What?** The techniques generalize to organizational knowledgebases; larger deployments would want PostgreSQL, governance, and probably an MCP server interface for non-technical users on desktop AI apps. Documents and spreadsheets are easier than slides; audio can be transcribed locally with models like Parakeet. He closes by tying local-first AI to data safety, citing a same-week controversy about OpenAI and a math result. ### 🗒️ Notes - The order of operations is the transferable lesson: purpose, then schema, then cheap classical preprocessing, then the smallest model that meets a quality bar set by a bigger model. - Hybrid retrieval (BM25 + vectors + RRF) is presented as strictly better than either alone for multi-topic items; this matches common practice in search engineering. - The "human remains the author" boundary is a values choice, not a technical limit, and it is the part most worth copying into any PKM system. - For an Obsidian vault: the schema (verbatim text, visual elements, summary, keywords, type, domain, source, position) maps neatly onto note frontmatter; SQLite + FTS5 is what several Obsidian search plugins already approximate. - The glossary at the end is AI-generated (DeepSeek v4.1 Flash) and is a helpful reading aid, but it is not the author's voice. ## 📚 References Primary sources cited or named in the essay (all appear in the body): - Penn, C. S. (2026-09-13). Almost Timely News: Creating A Keynote Speaker Second Brain. https://www.christopherspenn.com/2026/09/almost-timely-news-creating-a-keynote-speaker-second-brain-2026-09-13/ - Reynolds, W. Working-in-public video on slide catalog assembly. https://lnkd.in/p/dSBpGjcz - Trust Insights. 5P Framework. https://www.trustinsights.ai/5p-framework - Alibaba. Qwen3.6-35B-A3B model card. https://huggingface.co/Qwen/Qwen3.6-35B-A3B - oMLX. https://omlx.ai/ - SQLite. https://sqlite.org/index.html - LibreOffice. https://www.libreoffice.org/ - Trust Insights (2023). Disclosure of AI and protection of copyright. https://www.trustinsights.ai/blog/2023/08/disclosure-of-ai-and-protection-of-copyright/ Named but unlinked in the body (verify independently before citing): sqlite-vec, SQLite FTS5, BM25, reciprocal rank fusion, LiteLLM, GLM-5.3, Parakeet, PostgreSQL vector and BM25 extensions. The OpenAI math-problem controversy is asserted without a link. ## 🗣️ Paraphrase Because the essay runs about 3,000 words in its core, this paraphrase is section by section rather than paragraph by paragraph. **Opening.** After last week's high-level PR advice, this issue gets technical. Wil Reynolds' public build of a cosine-similarity slide assistant inspired the project, but Penn's slide library is far larger, more varied, and mostly visual, so he first needs an inventory of what he has. **Part 1.** Before touching tools, he articulates purpose, people, process, platform, and performance. The goal is a meta-index that lets an AI tool draft a deck from a brief and an interview, flag gaps, and flag outdated slides. Indexing doesn't need frontier AI; a local model suffices. "Done" means a working catalog. **Part 2.** Slides must become images so a smaller vision model can read them. That conversion is a classical task: a shell script driving LibreOffice's command-line tool did it in under fifteen minutes. Letting an agentic AI do the same thing wastes quota; a script is reusable and can later be handed to an agent. **Part 3.** Define the fields you want before running a single slide through the model. Narrating thousands of slides by hand is infeasible, so a Python script feeds each image to a locally run Qwen model with a custom quantization. The run took about half a day and cost nothing extra in cash. Quality was calibrated by comparing ten slides against a frontier model and tuning prompts. Storage is SQLite with a vector extension and full-text search; combining the two rankings gives hybrid retrieval that handles multi-topic slides better than embeddings alone and catches synonyms better than keywords alone. **Part 4.** The same 5P brief, handed to Claude Code with instructions to interview him, produced an assembly application in about seven hours. It retrieves, arranges, and flags gaps; it does not write slides, because the speaker must know his own material. **Part 5.** This is the engine of any organizational knowledgebase. Scale it with a multi-user database, governance, and an MCP interface for desktop AI users. Text documents are easier than slides; audio can be transcribed locally. Building a second brain this way keeps your knowledge in your own hands, which he links to a current dispute over OpenAI and mathematicians' work. ## 📰 Newsletter **Subject: A 7,000-slide second brain, built on a laptop, and what WordPress can steal from it** Christopher Penn spent a week turning two years of keynote decks into a searchable, AI-assisted catalog, and the punchline for the WordPress world isn't the slides. It's the architecture: rasterize with LibreOffice, index with a small local model, store in SQLite with FTS5 and sqlite-vec, fuse keyword and vector search with reciprocal rank fusion, and let a coding agent build the front end from a five-question plan. Read the original: https://www.christopherspenn.com/2026/09/almost-timely-news-creating-a-keynote-speaker-second-brain-2026-09-13/ Why you should care if you ship WordPress things: - **Your archive is a slide deck.** Make.WordPress.org posts, Slack threads, WordCamp talks, Learn WordPress lessons, and Trac tickets are an unindexed keynote library. The same eight-field schema (verbatim text, visual elements, summary, keywords, type, domain, source, position) would turn them into something a contributor could actually query. - **Hybrid search is not exotic.** BM25 has powered search since before REST endpoints existed; SQLite ships FTS5. If a solo speaker can fuse lexical and vector ranking on a Mac, a plugin can too. Elasticsearch-flavored plugins and the Core Performance team's search conversations should note that SQLite plus an extension got Penn there without a server. - **Local-first is a feature.** Penn's argument for keeping proprietary knowledge on your own machine is the same argument hosts and agencies hear from clients every week. WordPress Playground already runs WordPress in the browser; a Playground-shaped local retrieval demo would be a good story. - **The human writes the slides.** His AI is not allowed to author content, only retrieve and arrange it. That is a sane editorial policy for Core AI experiments, block-editor assistants, and anyone drafting community guidance on AI-assisted contribution. For Obsidian folks: his schema is a frontmatter template waiting to happen. Steal it. ## ✅ Action Items ### DevRel - Draft a tutorial: "Index your talk archive with LibreOffice, a local VLM, and SQLite," using Penn's schema and a WordPress-flavored corpus (WordCamp slides, Learn WordPress decks). - Add Penn's "AI retrieves, the human authors" boundary to the personal content engine's AI-use guidance. ### OSPO - Evaluate SQLite + sqlite-vec + FTS5 as a zero-infrastructure retrieval layer for internal documentation search before proposing hosted vector databases. - Note the licensing and provenance implications of running open-weight models (Qwen3.6) on internal materials versus sending them to a cloud API. ### Community - Propose a Contributor Day experiment: catalog a Make team's P2 archive with the same pipeline and measure whether new contributors find answers faster. ### WordPress subteams - **Core AI**: Compare Penn's hybrid retrieval pattern with any planned Abilities API or AI Client retrieval work; document why RRF over BM25 + vectors handles multi-topic content better than embeddings alone. - **Core Performance**: Review whether FTS5-style lexical ranking could inform native search improvements; benchmark against current MySQL fulltext behavior. - **Playground**: Scope a demo that rasterizes uploaded slides or posts and indexes them client-side, showing local-first retrieval in the browser. - **Documentation**: Pilot the eight-field record on HelpHub or DevHub articles as structured metadata for retrieval. - **Training**: Consider indexing Learn WordPress lesson decks so course authors can find reusable slides and spot gaps. - **Hosting**: Discuss local-model indexing as a privacy story for managed hosts handling client content. - **Meta**: Assess whether WordPress.org search could add a hybrid ranking layer without new infrastructure. - **Marketing**: A short piece on "local-first AI and your content" fits the current privacy conversation. ### The WPOCC - Run a small governance conversation: if community knowledge is indexed locally by members, who owns the resulting catalogs and how are they shared? ### FAIR - Explore whether a federated package or plugin index could expose BM25 + vector search over plugin readmes, using the same SQLite pattern as a reference implementation. > [!quote] Action anchor (verbatim, tagged `#action/h1`) > If you sit down and go through the 5P Framework by Trust Insights™, it's pretty much your entire project plan in a nutshell. This plus an interactive interview from the AI coding tool of your choice is enough to build requirements, a technical spec, and a workplan. ## 🧪 Hypotheses 1. **H1.** A local VLM tuned on a 10-item sample against a frontier model will reach at least 85% agreement with the frontier model's structured output across a 200-item held-out set. Test: run both on 200 slides, score field-level agreement. 2. **H2.** Hybrid RRF retrieval will outperform pure cosine similarity on multi-topic slides by at least 20% in top-10 precision. Test: build a labeled query set from real session descriptions. 3. **H3.** The "copy the last similar deck" process takes longer than catalog-assisted assembly by a measurable margin (Penn's Process claim). Test: time-box five deck builds each way. 4. **H4.** Total cost of the local run (electricity, hardware amortization, operator time) is lower than a cloud API run only above a corpus-size threshold. Test: model costs at 500, 2,000, 7,000 items. 5. **H5.** Applying the eight-field schema to Make.WordPress.org posts improves new-contributor answer-finding time. Test: A/B a search interface during a Contributor Day. ## 📊 Methodology The piece is a first-person build log, not a study. Method as described: 5P planning; classical batch conversion (shell + soffice); schema definition; local VLM inference over roughly 7,000 images with self-made quantization; small-sample prompt calibration against a frontier model (n=10, ~15 minutes); SQLite storage with vector and FTS5 extensions; RRF fusion; agentic application build from the 5P brief (~7 hours). No metrics beyond wall-clock time and token count are reported; quality is asserted ("quite good"). No held-out evaluation, no user testing, no cost accounting beyond API-dollar avoidance. ## ❓ Questions - How was "close to the same level of quality" measured between Qwen and GLM-5.3? - What embedding model produced the vectors stored in sqlite-vec? - How were duplicates hashed given that slide images with tiny differences produce different hashes? - What does the assembled deck's accuracy look like on a real session brief, and how often does it surface stale slides? - What hardware (memory, chip) did the 12-hour run require, and what would it cost a reader to replicate? - Does the workflow generalize to non-Apple hardware, where oMLX and ANE are unavailable? ## 🔮 Future Research - A public benchmark for slide-catalog retrieval (queries, relevance labels) so hybrid versus vector claims can be tested. - Cross-platform recipes (Linux, Windows) using llama.cpp or vLLM instead of oMLX. - Extending the schema to Obsidian notes and measuring recall on personal vaults. - Governance patterns for organization-scale versions: access control, provenance, and retention. - Cost-comparison studies of local versus cloud indexing that include energy and labor. ## 💡 Implications - **Practice.** Plan with purpose and schema first; use classical tools for conversion; reserve AI for the judgment step; keep humans as authors. Small local models are adequate for indexing when calibrated. - **Policy.** Local-first indexing is a defensible default for proprietary or member data; organizations should write down when cloud AI is and isn't permitted, rather than relying on anecdotes. - **Security.** Running models locally removes one exfiltration path but not others (backups, sync folders, the SQLite file itself). Threat-model the database, not just the model. - **Education.** The essay is a strong teaching case for retrieval basics (BM25, embeddings, RRF) and for the difference between indexing and generation; the AI glossary helps but should be checked. - **Funding.** The pattern lowers the barrier for volunteer-run projects (WordPress Make teams, The WPOCC, FAIR) to build search over their own archives without recurring API bills. ## 🧠 Learning Styles Assessment - **Visual.** Moderate. No diagrams or screenshots in the captured text; the architecture is described in prose. A pipeline diagram would help. - **Aural.** Strong. YouTube and MP3 versions of the issue are provided. - **Verbal.** Strong. Clear sequential narrative with headed parts and defined terms. - **Kinesthetic.** Strong for practitioners: concrete steps that can be reproduced immediately. - **Logical.** Strong. Cause-and-effect ordering, explicit trade-offs (model size, cost, search method). - **Social.** Moderate. Credits collaborators (Reynolds, Robbert) and invites replication, but no group exercise. - **Solitary.** Strong. Designed for a single builder working alone. - **Naturalistic.** Weak. No pattern-in-nature framing beyond the "second brain" metaphor. - **Reading level.** Roughly grade 11-13 (college-entry). Sentences are conversational but dense with acronyms; the glossary lowers the barrier to about grade 10 for motivated readers. ## 📊 Accessibility Assessment Mapped to WCAG 2.2 AA, based on the captured text and links (the rendered email or web page was not inspected). 1. **Text alternatives (1.1.1).** Only one image (newsletter subscribe graphic) with alt text; the essay is text-first. Pass. 2. **Headings and structure (1.3.1, 2.4.6).** Clear H2/H3 hierarchy for parts. Pass. 3. **Link purpose (2.4.4).** Most links are descriptive; the LinkedIn short link and "Click here" phrasing are weaker. Partial. 4. **Readability and language (3.1.3, 3.1.5).** Heavy jargon, mitigated by the appended glossary. Partial. 5. **Emoji use (1.3.3, 4.1.2).** Emoji in headings and the poll may be read aloud verbosely; decorative use should be limited. Partial. 6. **Multiple modalities (1.2.x).** Audio and video versions exist; captions on the video are not verifiable here. Likely pass. 7. **Contrast and reflow (1.4.3, 1.4.10).** Not assessable from text; depends on the newsletter template. ## 🌍 DEIB Assessment - **Cultural breadth.** Uses models from Alibaba (Qwen) and Zhipu (GLM) alongside US tools; acknowledges non-US ecosystems. Assumes Apple hardware and USD budgets. - **Inclusive language.** Neutral, first-person, no exclusionary terms. "Folks" is used throughout. - **Neurodiversity support.** Headed parts, numbered lists, and a glossary help; long paragraphs and dense acronyms may burden some readers. - **Representation.** Credits a woman colleague (Katie Robbert) as the framework author and Wil Reynolds as inspiration; audience examples are generic. - **Psychological safety.** Tone is encouraging ("this is how you do it") and self-deprecating about clip art; no shaming of non-technical readers, though shell familiarity is assumed. ## ✅ Logical Validity **Strong reasoning.** - Choosing the smallest model class that can do the job (VLM rather than omni) follows from a correct cost ordering. - Preferring a reusable script to agent-driven conversion is sound: same output, no quota burn, reusable artifact. - Calibrating a small model against a larger one on a sample is a legitimate cheap evaluation, even if the sample is small. - The hybrid retrieval argument (BM25 catches exact terms, vectors catch related concepts, RRF blends them) is standard and correct. - Keeping the human as author is well justified: the speaker must know the deck. **Fallacies.** 1. **Anecdotal evidence / appeal to fear** (Part 5, final paragraph). An unresolved, unlinked news story about OpenAI and mathematicians is offered as the reason local AI "is how you keep your data safe." One contested anecdote does not establish that cloud AI leaks data generally. Fix: cite vendor data-retention terms or documented incidents, and state the specific threat the local approach mitigates. 2. **Incomplete accounting** (Part 3, "cost zero additional dollars"). Not a formal fallacy, but the comparison omits electricity, hardware, and operator time, so the cloud-versus-local cost claim is overstated. Fix: present a total-cost comparison including time and energy. A further caution: the claim that the brain handles "one language stream at a time" is a simplification of multimedia-learning research and is presented without citation. **Tone of voice.** Practical, confident, conversational, lightly self-deprecating, with mild advocacy for local-first tooling. Not hostile or defensive. ## 🛡️ DARVO Analysis - **Deny.** Not present. The author does not deny any criticism; he pre-empts "so what?" and answers it. - **Attack.** Not present. The reference to OpenAI is critical but reports others' allegations rather than attacking a person. - **Reverse victim-offender.** Not present. No victim framing. Overall: no DARVO pattern detected; the piece is expository, not adversarial. ## 🎯 Conclusion **Bottom line.** Penn's build is a reproducible template for a local-first second brain: purpose and schema first, classical preprocessing, a calibrated small model for indexing, SQLite with hybrid retrieval, and an agent-built interface that retrieves but never authors. The technical reasoning is sound; the privacy argument is directionally right but argued by anecdote. **Next steps.** 1. Adopt the eight-field schema as an Obsidian frontmatter template for talk and lesson notes. 2. Prototype SQLite + FTS5 + sqlite-vec over one Make team's archive and measure answer-finding time. 3. Write the WordPress-flavored tutorial and pitch a Playground demo. 4. Before repeating the privacy claim publicly, replace the anecdote with vendor terms or documented incidents. ## 🧱 Refactor Appendix ### Literature Split Pad ### h1 — One language stream at a time > [!quote] > Our language processors in our brains can only handle one language stream at a time, so your brain has to decide whether you want to read or listen, and if a slide is a wall of text, then you'll tune me out. #atom/literature/h1 ### h2 — Twelve hours, twenty million tokens, no API bill > [!quote] > It took almost 12 hours for Qwen to index the slides, at a run rate of nearly 20 million tokens. This is why I went the Qwen route, by the way. If I had attempted to do with with a cloud API, this would have been a bill in the hundreds of dollars. Instead, it took longer but cost zero additional dollars and the results are quite good. #atom/literature/h2 ### h3 — BM25 plus vectors equals reciprocal rank fusion > [!quote] > If you pair BM25 with the sqlite-vec embeddings, you get something called reciprocal rank fusion (RRF), which is the best of both worlds. #atom/literature/h3 ### Atomic Split Pad #### z1 — AI retrieves; the speaker authors A retrieval system for creative material should arrange and flag, never generate, because the presenter's ownership of the content is the test of whether the content is usable on stage. Source passage: > [!quote] > What's important here is that it's not creating new slides for me. That's my job, for good or ill (there are days when folks like Katie would rather me use AI than fill slides with varying forms of clip art, a favorite pastime of mine), and having AI do my slides for me would fail the ultimate test of any content – I wouldn't know what's in my own speaking deck. #atom/zettel/z1 ### Fleeting Queue #### f1 — Second brain without the vendor Check how far this maps onto the Obsidian vault: local model, local database, classical search. Source passage: > [!quote] > For those folks who want to build an "AI second brain", this is how you do it without spending big bucks on proprietary software or handing over your most valuable knowledge to a third party – with the help of local AI, tools you run on your machine, and classical databases enabled for AI. #atom/fleeting/f1 ### Fallacy anchor > [!quote] > This has been center stage this past week because OpenAI claims to have solved a hundred-year-old math problem, and some mathematicians are saying OpenAI stole their work that they were apparently using with OpenAI tools. If you've got sensitive proprietary information, building a second brain like this using local AI is how you keep your data safe. #fallacy/anecdotal <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Almost Timely News 🗞️ Creating a Keynote Speaker Second Brain.md|Almost Timely News: 🗞️ Creating a Keynote Speaker Second Brain by Christopher S Penn]] ## Highlights - Our language processors in our brains can only handle one language stream at a time, so your brain has to decide whether you want to read or listen, and if a slide is a wall of text, then you’ll tune me out. ([View Highlight](https://read.readwise.io/read/01m2dr4bt1hvhegqppny10n4jc)) - Note: Why his slides are image-first: one language stream at a time (uncited simplification of multimedia-learning research). - Tags: atom/literature/h1 - If you sit down and go through the 5P Framework by Trust Insights™, it’s pretty much your entire project plan in a nutshell. This plus an interactive interview from the AI coding tool of your choice is enough to build requirements, a technical spec, and a workplan. ([View Highlight](https://read.readwise.io/read/01m2dr4rg8xyn4v876v8k54n6y)) - Note: Action: run 5P plus an agent interview before building any knowledgebase or catalog project. - Tags: action/h1 - It took almost 12 hours for Qwen to index the slides, at a run rate of nearly 20 million tokens. This is why I went the Qwen route, by the way. If I had attempted to do with with a cloud API, this would have been a bill in the hundreds of dollars. Instead, it took longer but cost zero additional dollars and the results are quite good. ([View Highlight](https://read.readwise.io/read/01m2dr4e057gp49afnbt9h5t59)) - Note: The local-vs-cloud cost claim; note it omits energy, hardware, and operator time. - Tags: atom/literature/h2 - If you pair BM25 with the sqlite-vec embeddings, you get something called reciprocal rank fusion (RRF), which is the best of both worlds. ([View Highlight](https://read.readwise.io/read/01m2dr4g2r99npvczhq5f93a78)) - Note: The retrieval architecture in one sentence: hybrid lexical plus vector search fused with RRF. - Tags: atom/literature/h3 - What’s important here is that it’s not creating new slides for me. That’s my job, for good or ill (there are days when folks like Katie would rather me use AI than fill slides with varying forms of clip art, a favorite pastime of mine), and having AI do my slides for me would fail the ultimate test of any content – I wouldn’t know what’s in my own speaking deck. ([View Highlight](https://read.readwise.io/read/01m2dr4jewy9f3t33hjg4a3sp8)) - Note: Zettel: AI retrieves and arranges; the human authors. The ownership test for any content system. - Tags: atom/zettel/z1 - For those folks who want to build an "AI second brain", this is how you do it without spending big bucks on proprietary software or handing over your most valuable knowledge to a third party – with the help of local AI, tools you run on your machine, and classical databases enabled for AI. ([View Highlight](https://read.readwise.io/read/01m2dr4p36ekx1d1102350cps8)) - Note: Fleeting: map this local-first recipe onto the Obsidian 2nd Brain vault. - Tags: atom/fleeting/f1 - This has been center stage this past week because OpenAI claims to have solved a hundred-year-old math problem, and some mathematicians are saying OpenAI stole their work that they were apparently using with OpenAI tools. If you’ve got sensitive proprietary information, building a second brain like this using local AI is how you keep your data safe. ([View Highlight](https://read.readwise.io/read/01m2dr4tqry2steetgx5xj7m55)) - Note: Anecdotal evidence / appeal to fear: one unresolved, unlinked news story used to establish that cloud AI is unsafe generally. - Tags: fallacy/anecdotal