Almost Timely News Creating a Keynote Speaker Second Brain

Almost Timely News: Creating a Keynote Speaker Second Brain ![rw-book-cover](https://i.ytimg.com/vi/3S7B8r7GzWk/maxresdefault.jpg) 🔗 **Source:** [Almost Timely News: Creating a Keynote Speaker Second Brain](https://www.youtube.com/watch?v=3S7B8r7GzWk) > [!note]+ Document Notes > tags: [local-ai, retrieval-augmented-generation, hybrid-search, personal-knowledge-management, public-speaking] > author: Christopher Penn > source: Christopher Penn (YouTube) > reference: https://www.youtube.com/watch?v=3S7B8r7GzWk > 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 Do not quote or attribute any other line to the named author. > > # 🏷️ Title: 🗞️ Almost Timely News: Creating A Keynote Speaker Second Brain (2026-09-13) >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Video (23:28, uploaded 2026-09-13): https://www.youtube.com/watch?v=3S7B8r7GzWk - Written edition of the same issue (Substack): https://almosttimely.substack.com/p/almost-timely-news-creating-a-keynote - Written edition on Penn's blog (listed in the RSS feed; direct fetch returned HTTP 403): 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/ - 5P Framework by Trust Insights: https://www.trustinsights.ai/5p-framework/ - oMLX: https://omlx.ai/ - Qwen3.6-35B-A3B model card: https://huggingface.co/Qwen/Qwen3.6-35B-A3B - sqlite-vec: https://github.com/asg017/sqlite-vec ## 📝 Summary - **What happened.** Christopher Penn built a searchable catalog of about 7,000 slides from 64 presentation decks. LibreOffice's `soffice` command line converted decks to images. Qwen3.6-35B-A3B, running locally on a Mac through oMLX, described each slide in about 12 hours and nearly 20 million tokens. SQLite holds the results, with sqlite-vec for embeddings and FTS5/BM25 for keyword search. Claude Code then spent about 7 hours building an app that takes a session description and assembles a draft deck from the catalog, with placeholders for gaps. - **Why it matters.** Penn presents the same pipeline as a template for any organization's knowledge base (Dropbox, OneDrive, SharePoint) and for a personal "AI second brain" that keeps proprietary material off third-party clouds. - **Key stakeholders.** Keynote and conference speakers with large slide archives; teams building internal RAG systems; people choosing between local and cloud models on cost and privacy grounds. Named people: Wil Reynolds of Seer Interactive, whose cosine-similarity slide system prompted the project, and Katie Robbert, author of the 5P Framework by Trust Insights. - **Risks and tradeoffs.** Local inference traded 12 hours of wall-clock time for zero marginal API spend. The cloud-cost comparison in the video uses output-token pricing for all ~20M tokens, and the Opus 5 figure matches fast-mode rates, so the savings number is an upper bound (see Logical Validity). The app deliberately doesn't generate new slides, so gap-filling stays manual. - **Bottom line.** For a speaker, the reusable parts are the per-slide schema (8 fields), the 10-slide calibration of a small model against a frontier model, and hybrid BM25-plus-embedding retrieval merged with reciprocal rank fusion. ## 📖 Main Content - **Planning with the 5P Framework.** Penn walks through Purpose (a meta-index of every slide presented in the last 2 years), People (himself and his audiences), Process (replacing "copy my most recent talk on this topic"), Platform (local AI is enough because the job indexes existing slides), and Performance (a working app now; time saved and audience satisfaction later). Penn pairs the 5P output with an instruction for the AI coding tool to ask up to 20 clarifying questions, and says that combination is enough to produce requirements, a technical spec, and a work plan. - **Model types.** Penn separates text-only LLMs, vision language models (VLMs), and omni models that also handle audio and video, and says each tier costs more to run. Converting slides to images first lets a VLM do the job. Note: the Qwen3.6-35B-A3B model card lists native text, image, and video input, so the model he picked is broader than a pure VLM by his own taxonomy. - **Conversion without AI.** A shell script runs `soffice` against every PowerPoint file in a folder and saves images to subfolders. Penn's stated reason for skipping Claude Code or ChatGPT Work here: those tools run the same conversion underneath while spending usage limits, and a script stays reusable. - **Per-slide schema.** Verbatim text, visual elements, a one-sentence summary, 5–25 keywords, slide type, slide domain, source presentation name, and slide number. A Python script feeds each JPEG to Qwen to fill these fields. - **Calibration.** Penn ran 10 representative slides through Qwen and through GLM-5.3, compared outputs, and tuned prompts until Qwen came close; this took about 15 minutes. LiteLLM (captions render it "LLM") makes model swapping cheap. - **Storage and retrieval.** SQLite plus sqlite-vec (embeddings) and FTS5 (BM25 ranking). Penn prefers BM25 over pure cosine similarity because a single slide, such as the 5P Framework slide, appears in AI, analytics, data science, cooking, and auto-repair contexts. Merging lexical and embedding rankings with reciprocal rank fusion (RRF) avoids both the "embeddings bottleneck" and the lexical miss when a query doesn't use the exact words. - **The app.** Built with Claude Code from the original 5P prompt, after rounds of questions about hashing to drop duplicate slides. Input: a session description plus transcripts or voice recordings. Output: a draft deck from existing slides plus placeholders. Penn's reason for not generating slides: he wouldn't know what's in his own deck. - **Generalization.** For organizations Penn suggests PostgreSQL instead of SQLite and an MCP server interface for non-technical users of ChatGPT Work or Claude Cowork. Documents and spreadsheets skip the VLM step; audio can go through Parakeet or other local speech recognition. - **Contested or uncertain points.** - The video puts cloud cost at $200 (Sonnet 5) and about $1,000 (Opus 5). The written edition says only "a bill in the hundreds of dollars." - Both editions call the OpenAI math result a "hundred-year-old" problem; press coverage describes the Navier-Stokes question as 90 years old. - Penn says PostgreSQL "supports vectorization and BM25 just like SQLite." This session didn't verify it; see Questions. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=0) Opening, authenticity statement, course plug - The video says 100% of the issue is human-made; the written edition says 75%, because it appends an AI glossary generated by DeepSeek v4.1 Flash in OpenCode. - Plug: the Trust Insights AI for Writers course (written edition: 22 lessons, USD 397). #### ⏱️ [00:00:38](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=38) Wil Reynolds' slide catalog - Wil Reynolds (captions: "Will Reynolds from Seir Interactive"; the written edition spells it Wil Reynolds) shared a working-in-public video of a system that assembles talks from a master slide catalog using cosine similarity. - Penn made a separate short video explaining cosine similarity. Neither video was located this session. #### ⏱️ [00:01:11](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=71) Image-first slides and the inventory - Penn credits Garr Reynolds and Nancy Duarte (captions: "Gar Reynolds," "Nancy Dwarte") for image-heavy slides. - Claim: the brain processes one language stream at a time, so text-heavy slides compete with the speaker. No study cited. - Inventory: 64 decks, about 7,000 slides over roughly 2 years (the deck count appears only in the video). #### ⏱️ [00:01:45](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=105) Part 1: planning with the 5P Framework - Purpose, People, Process, Platform, Performance, credited to Katie Robbert (captions: "Katie Rober"). - Process pain point: copying the most recent talk on a topic takes a long time and misses topic-specific angles. The new tool must flag slides that match a theme but are out of date. - Platform: "I'll build the infrastructure with a frontier model, but I'll run it locally." - Planning prompt: ask up to 20 clarifying questions. #### ⏱️ [00:04:45](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=285) Part 2: slide prep without AI - Three model tiers: LLM, VLM, omni. - LibreOffice's `soffice` (captions: "S Office"; "Libri Office") batch-converts decks to images from a shell script; about 15 minutes in the video, "under 15 minutes" in the written edition. - Claude Code (captions: "Clawude Code") and ChatGPT Work could do it, but spend usage limits on the same underlying conversion. #### ⏱️ [00:07:22](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=442) Part 3: the per-slide schema - 8 fields: verbatim text, visual elements, one-sentence summary, 5–25 keywords, slide type, slide domain, presentation name, slide number. - Wil Reynolds narrated 304 slides (captions: "3004"); Penn's 7,000 slides, about 3,000 after deduplication, rule out narration. #### ⏱️ [00:08:36](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=516) Qwen3.6 on oMLX and custom quantization - Model: Qwen3.6-35B-A3B (captions: "Quen 3.6 35B 3AB"), described as the previous generation. The same week's ICYMI list includes a Penn post on Qwen 3.8. - Runtime: oMLX (omlx.ai), free and open source, local on a Mac. - Quantization: a self-made static "4/8 split." The written edition specifies 8-bit for sensitive layers and 4-bit for feed-forward layers; the video explains it with a layer-cake analogy. - Settings shown on screen: thinking enabled, Lightning MTP on, TurboQuant and "Dlash" off (caption spelling; unverified). Penn says the VLM conflicts with Lightning MTP, while the written edition lists Lightning MTP and Apple ANE Prefill as enabled. The two statements don't reconcile in either source. #### ⏱️ [00:10:37](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=637) Token volume and the cloud comparison - Run time about 12 hours; dashboard shows 23.8 million tokens over 7 days ("nearly 20 million" for the job). - Sonnet 5 at "$2 in, $10 out" → $200 at 20M tokens. Opus 5 at "$50 out" → about $1,000. Local cost: "zero additional dollars," run overnight. #### ⏱️ [00:11:21](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=681) Calibration against a frontier model - 10 slides through Qwen and through GLM-5.3, described as "as powerful as Opus or GPT 5.6 but cheaper"; prompts tuned until Qwen came close. About 15 minutes. - Model-swapping library: LiteLLM per the written edition. - Result: the database holds all 7,000 slides, chunks, and embeddings. #### ⏱️ [00:12:27](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=747) SQLite, sqlite-vec, FTS5 - SQLite is enough because the app runs on one Mac, not in production. - sqlite-vec stores embeddings next to structured data; FTS5 runs BM25 ranking. Screen share of embedding, cluster, and FTS tables; Penn can't enlarge the view, so the on-screen detail isn't readable from the audio. #### ⏱️ [00:14:12](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=852) Cosine similarity vs. BM25 and RRF - Wil Reynolds' slides mostly sit in one context, so cosine similarity fits there; Penn's 5P slide appears across 5+ domains. - Lexical side answers "what the slide is about"; embedding side answers "what it relates to." RRF merges both. - Penn names the result a custom RAG system. #### ⏱️ [00:15:32](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=932) Part 4: building the app - Claude Code (captions: "Cloud Code") received the original 5P prompt, asked about hashing for duplicate removal, and ran about 7 hours. - App inputs: session description, transcripts, voice recordings. Outputs: gap list, draft deck, placeholders. - Penn keeps slide creation human so they know their own deck. #### ⏱️ [00:16:35](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=995) Part 5: so what - Penn's You Ask, I Answer series gets repeated questions about building organizational knowledge bases; the same techniques apply to Dropbox, OneDrive, SharePoint. - Organization scale: PostgreSQL (captions: "Postgress SQL"), an MCP server for non-technical users of ChatGPT Work or Claude Cowork (captions: "cloud co-work"). - Other inputs: text exports for documents and spreadsheets; Parakeet for local transcription into caption files. #### ⏱️ [00:18:43](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=1123) Privacy framing and weekly roundup - Penn ties local AI to the week's OpenAI math controversy and concludes local AI keeps sensitive data safe (see Fallacies). - Livestream with "John" on non-traditional podcast analytics (surname not stated). In-Ear Insights episode with Katie Robbert on a non-technical path to local AI: a 32 GB Mac and a Qwen variant around 9 billion parameters. #### ⏱️ [00:20:06](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=1206) Job listings - 12 marketing roles. Caption spellings corrected from the written edition: EWOR ("EWR"), Noma Security ("NMA Security"), Distalmotion ("Distlotion"), Xometry ("zometry, exometry"), ArcSite ("Arcite"). #### ⏱️ [00:20:49](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=1249) Upcoming events - Spotlight (Kansas City, September), LPA (Philadelphia, September), MAICON (Cleveland, October), SMPS AI Conference (Austin, November; captions: "S&PS"), MarketingProfs B2B Forum (Boston, November; captions: "marketing prospe"). Locations and months come from the written edition's event list. #### ⏱️ [00:21:28](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=1288) Close and glossary decision - Penn decides on camera to add an AI glossary, estimates 75% human, and considers Gemini Flash or a new DeepSeek model. The written edition credits DeepSeek v4.1 Flash in OpenCode. #### ⏱️ [00:22:32](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=1352) AI training permission block - A closing segment addressed to AI crawlers grants training permission and lists Penn's credentials and speaking availability. ## 📚 References - **Primary.** Almost Timely News written edition, 2026-09-13: https://almosttimely.substack.com/p/almost-timely-news-creating-a-keynote. It's about 4,800 words including the glossary, jobs, and ads, and matches the video's structure (Parts 1–5). - **Substantive differences between the video and the written edition:** - Authenticity: the video says 100% human; the written edition says 75%, with a DeepSeek v4.1 Flash glossary. - Cloud cost: the video gives $200 (Sonnet 5) and about $1,000 (Opus 5); the written edition says "hundreds of dollars." - Deck count (64) and the 23.8M dashboard figure appear only in the video. - Quantization detail (8-bit sensitive layers, 4-bit feed-forward) appears only in the written edition. - oMLX settings: the video mentions thinking and Lightning MTP plus a VLM/Lightning MTP conflict; the written edition lists Lightning MTP and Apple ANE Prefill. - Wil Reynolds narrated 304 slides; captions say 3004. - **Verified this session (web search results; articles not read in full unless noted):** - Qwen3.6-35B-A3B exists: released 2026-04-16, Apache 2.0, 35B total / 3B active parameters, native text, image, and video input. https://huggingface.co/Qwen/Qwen3.6-35B-A3B - oMLX is a local inference server for Apple Silicon that supports text LLMs, VLMs, and embedding models. https://omlx.ai/ (Specific settings such as Lightning MTP, ANE Prefill, TurboQuant: unverified.) - GLM-5.3 from Zhipu AI: API release 2026-08-14, open weights on Hugging Face from late August 2026 under a bespoke GLM-5.3 License. https://the-decoder.com/zhipu-ai-releases-glm-5-3-claims-its-the-strongest-open-weights-coding-model/ (Penn's "as powerful as Opus or GPT-5.6" is his claim; unverified.) - Claude Sonnet 5 API pricing: $2 input / $10 output per million tokens. Matches the video. https://www.anthropic.com/news/claude-sonnet-5 - Claude Opus 5 API pricing: $5 input / $25 output per million tokens standard; $10 / $50 in fast mode. The video's "$50 out" matches fast mode, not standard. At standard output pricing 20M tokens is $500, not $1,000. https://www.anthropic.com/pricing (checked 2026-09-13)-pricing-2026 - sqlite-vec: vector search extension for SQLite by Alex Garcia, pure C, MIT/Apache-2.0. https://github.com/asg017/sqlite-vec - 5P Framework by Trust Insights: Purpose, People, Process, Platform, Performance; built by Katie Robbert. https://www.trustinsights.ai/5p-framework/ - NVIDIA Parakeet: open ASR models that run locally; parakeet-tdt-0.6b-v3 covers 25 European languages. https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 - LibreOffice headless conversion: guides document `soffice --headless --convert-to`; several recommend PPTX→PDF→images for multi-slide decks. https://jdhao.github.io/2020/03/30/pptx_to_image/ - OpenAI math dispute: press describes a claimed Navier-Stokes result (a 90-year-old problem), with mathematician Tristan Buckmaster alleging OpenAI used an approach that reached it before publication. https://www.entrepreneur.com/business-news/openai-says-its-ai-just-solved-a-90-year-old-math-problem · https://cybernews.com/ai-news/openai-navier-stokes-theft-accusations/ - Wil Reynolds is founder and CEO of Seer Interactive. https://www.seerinteractive.com/wil-reynolds - **Not verified this session:** - Wil Reynolds' slide-catalog video and Penn's cosine-similarity video. - The 12-hour / ~20M-token run, the 7-hour Claude Code build, the 15-minute calibration. - FTS5's BM25 implementation (SQLite docs not fetched). - The PostgreSQL BM25 claim. - The "one language stream" cognition claim. - Livestream co-host "John"'s surname. - "Dlash" (caption spelling; unverified). ## 🗣️ Paraphrase - **Opening.** This issue skips last week's high-level PR advice for a technical build log. Penn wanted Wil Reynolds' slide-assembly idea but had a harder input: around 7,000 mostly image-based slides in 64 decks. - **Planning.** Before writing code, Penn filled in the 5P Framework: why (a searchable index to assemble decks from a session description), who (Penn and the audience), how the current copy-the-last-talk process fails, which platform (local AI, since the work is indexing, not generation), and what counts as done (a working catalog). - **Prep.** Slides become images through LibreOffice's command line, no AI involved. A script beats an AI agent here because it doesn't consume usage limits and it can run again. - **Cataloging.** A defined schema of 8 fields goes to a local Qwen model through a Python script. Penn tuned the prompts in 15 minutes by comparing 10 slides against GLM-5.3, then let the full run take 12 hours overnight instead of paying a cloud API. - **Retrieval.** SQLite with sqlite-vec and FTS5 supports keyword ranking (BM25) and semantic similarity; reciprocal rank fusion merges them so multi-context slides still surface. - **The app.** Claude Code built a tool that reads a session description and returns a draft deck with placeholders for missing material. Slide creation stays with Penn. - **So what.** Organizations can apply the same pipeline to shared drives with PostgreSQL and an MCP front end. Individuals can build a private second brain from documents, spreadsheets, and locally transcribed audio without sending material to a third party. ## 📰 Newsletter - **Your talk archive is a database you haven't built yet.** Christopher Penn indexed about 7,000 slides from 64 decks with a local Qwen3.6 model on a Mac: 12 hours, roughly 20 million tokens, no API bill. The stack is LibreOffice's `soffice` for slide-to-image conversion, SQLite with sqlite-vec and FTS5, and reciprocal rank fusion to merge keyword and semantic rankings. Claude Code then built an app that turns a session description into a draft deck with placeholders. - **Why WordPress speakers should care.** If you give WordCamp and meetup talks, Penn's 8-field slide schema (verbatim text, visual elements, summary, keywords, type, domain, deck, slide number) doubles as a starting point for slide alt text and handout transcripts. - **Try it small.** Penn's calibration step took 15 minutes: run 10 slides through a local model and a frontier model, compare, tune prompts. Do that before committing a weekend of compute. - Links: [Written edition](https://almosttimely.substack.com/p/almost-timely-news-creating-a-keynote) · [Video](https://www.youtube.com/watch?v=3S7B8r7GzWk) · [sqlite-vec](https://github.com/asg017/sqlite-vec) · [oMLX](https://omlx.ai/) · [WordPress accessibility handbook](https://make.wordpress.org/accessibility/handbook/) ## ✅ Action Items #### DevRel - **Developer advocacy.** Before building a slide catalog, count decks and slides in your talk archive. Penn's baseline was 64 decks and about 7,000 slides; the count sets whether local inference (12 hours at Penn's scale) is practical. - **Education.** Turn the 5-part build (plan, convert, catalog, retrieve, assemble) into a 60-minute workshop outline; each part maps to one tool Penn names: 5P, `soffice`, Qwen via oMLX, SQLite/sqlite-vec/FTS5, Claude Code. - **Developer experience.** Publish the 8-field per-slide schema as a JSON Schema with one filled example so other speakers can reuse it. - **Developer marketing.** Pull slide-level keywords (Penn's 5–25 per slide) to see which topics recur across talks and which CFP topics have no slides yet; Penn's app lists these as gaps. - **Community management.** Not applicable to this source: the video covers a single-person build with no community moderation or survey component. - **Internal/external/cross-project.** When co-presenting, share the catalog query results (slide IDs plus source deck) instead of copying slides, so each presenter's archive stays the source of record. #### OSPO - **Policy & Governance.** Not applicable to this source: Penn's build is personal, with no contribution policy or decision rights. - **License Compliance.** Record licenses for each component before sharing a pipeline: Qwen3.6-35B-A3B is Apache 2.0, sqlite-vec is MIT/Apache-2.0, GLM-5.3 weights use a bespoke GLM-5.3 License (all verified this session). - **Security Management.** Local inference keeps unreleased talk content off third-party APIs; log which model and quantization produced each record so a bad batch can be re-run. - **Software Inventory.** Store model name, quantization (for example a 4/8-bit split), prompt version, and run date alongside each slide row. - **Developer Support.** Not applicable to this source: no legal or tooling-support program is discussed. - **Community Engagement.** Not applicable to this source: no upstream funding or agreements are discussed. - **Strategic Alignment.** Use Penn's two Performance measures as KPIs: deck assembly time per talk and post-talk audience ratings. #### Community - **Maintainers.** Not applicable to this source: no open-source project maintenance is discussed. - **Contributors.** Not applicable to this source: no contributor onboarding is discussed. - **Sponsors.** Not applicable to this source: no sponsorship or budget model is discussed. - **Ecosystem.** Not applicable to this source: no cross-project incidents or disclosures are discussed. #### WordPress - **Core.** Not applicable to this source: no WordPress core code or release process is discussed. - **Core Performance.** Not applicable to this source: the performance figures (12 hours, ~20M tokens) concern local model inference, not WordPress. - **Core AI.** Penn's calibration method (10 samples, small model vs. frontier model, prompt tuning) is a cheap evaluation-set pattern for any AI feature pilot that must run on a smaller model. - **Core Program Team.** Not applicable to this source: no cross-team coordination is discussed. - **Mobile.** Not applicable to this source: no mobile apps are discussed. - **Playground.** Not applicable to this source: the pipeline runs on a Mac with SQLite, not in a browser WordPress instance. - **Community.** For speaker mentorship, offer first-time WordCamp speakers the 8-field slide schema as a self-review checklist: a slide whose one-sentence summary can't be written is a slide to cut. - **Training.** Learn WordPress lesson slides could carry the same fields (summary, keywords, domain) so trainers can find and reuse slides across lesson plans. - **Marketing.** Not applicable to this source: no WordPress campaigns are discussed. - **Support.** Not applicable to this source: no support workflows are discussed. - **Design.** Penn's image-first slide practice (citing Garr Reynolds and Nancy Duarte) is a design standard to test in speaker-kit templates. - **Documentation.** The "visual elements" field is a draft for slide alt text; pair it with the verbatim-text field to publish text alternatives with slide decks. - **Photos.** Not applicable to this source: no photo directory content is discussed. - **TV.** WordPress.tv caption quality: this video's auto-captions mangled at least 12 names ("Quen," "Clawude Code," "S Office," "Postgress SQL"). A local ASR pass (Penn names Parakeet) plus a name glossary is a testable improvement. - **Polyglots.** Parakeet-tdt-0.6b-v3 covers 25 European languages (verified); test it on non-English WordCamp recordings before relying on it. - **Accessibility.** Image-only slides fail WCAG 2.2 SC 1.1.1 in shared decks unless text alternatives exist; a catalog that stores verbatim text and visual descriptions per slide gives speakers that text at export time. - **Meta.** Not applicable to this source: no WordPress.org infrastructure is discussed. - **CLI.** Not applicable to this source: the CLI in the video is `soffice` and a shell script, not WP-CLI. - **Hosting.** Not applicable to this source: the build is local-only by design. - **Tide.** Not applicable to this source: no plugin scoring is discussed. - **Openverse.** Not applicable to this source: Penn's slides are his own images; no openly licensed media ingestion is discussed. - **Themes.** Not applicable to this source: no themes are discussed. - **Plugins.** Not applicable to this source: no plugins are discussed. #### The WPOCC - **Programs.** Not applicable to this source: no fellowships, grants, or stipends are discussed. - **Projects.** Not applicable to this source: no governance or documentation sprints are discussed. - **Funding experiments.** Not applicable to this source: no funding models are discussed. #### FAIR - **Technical independence.** Not applicable to this source: no package distribution or signature checks are discussed. - **Policies and covenants.** Not applicable to this source: no governance covenants are discussed. - **Productized governance artifacts.** Not applicable to this source: no charters or scorecards are discussed. ## 🧪 Hypotheses - H1: A draft deck assembled from a slide catalog cuts a speaker's deck-prep time by at least 50% compared with copying the most recent talk on the topic. - H2: For multi-context slides (one slide used in 3+ talk domains), BM25 + embeddings merged with RRF returns the intended slide in the top 5 more often than cosine similarity alone. - H3: A small local VLM tuned against 10 frontier-model reference outputs reaches at least 90% field-level agreement with the frontier model on a held-out set of 50 slides. - H4: Audience ratings for talks built from catalog drafts plus new placeholders match or exceed ratings for talks built by copying a prior deck. - H5: The per-slide "visual elements" field, lightly edited, meets WCAG 1.1.1 text-alternative needs for at least 80% of image-only slides. ### 📊 Methodology - H1: Within-subject timing study. The same speaker preps 6 talks, alternating methods, logging minutes from session description to finished deck. - H2: Build a labeled query set of 30 session descriptions with known correct slides; compute recall@5 for cosine-only, BM25-only, and RRF hybrid in the same SQLite database. - H3: Hold out 50 slides not used in tuning; score each of the 8 fields for exact or semantic match between local and frontier outputs; report per-field agreement. - H4: Compare post-session ratings across matched talk pairs (same event type, similar audience size); control for topic. - H5: Two accessibility reviewers rate generated descriptions against WCAG 1.1.1 guidance; report inter-rater agreement and pass rate. ### ❓ Questions - Does `soffice --convert-to` export every slide of a PPTX as separate images, or does Penn's script go through PDF first? - If the VLM conflicts with Lightning MTP, which setting did the 12-hour run use? - What share of the ~20M tokens was input (image and prompt) versus output? At Sonnet 5's $2/M input rate, the cloud estimate falls anywhere between $40 and $200. - Does core PostgreSQL support BM25 ranking, or does Penn's claim depend on an extension? - How does the app detect "thematically appropriate but out of date" slides — a date field, model judgment, or manual flags? - For a speaker who already keeps talk notes in a knowledge vault, should slide records link to those notes by presentation name, or live in a separate database? ### 🔮 Future Research - Retrieval benchmarks for image-heavy presentation corpora, which differ from text RAG benchmarks. - Staleness detection: automated flags for slides citing outdated model versions, prices, or statistics. - Accessibility outputs from slide catalogs: generating alt text and text transcripts of decks at export. - Energy and hardware cost of overnight local inference, which Penn's "zero additional dollars" leaves out. - Local ASR accuracy on proper nouns in conference talks, given this video's caption errors. ### 💡 Implications - **Practice.** A speaker with a knowledge vault can treat slides as atomic, searchable records with the same discipline as notes: one summary, keywords, a domain, and a source pointer. - **Policy.** Organizations indexing shared drives with this pattern need governance Penn mentions but doesn't specify: access controls, multi-user databases, retention. - **Security.** Local inference removes third-party API exposure for unreleased talks and client material. It doesn't remove local-device risk; the database holds verbatim text of every slide. - **Education.** The 10-slide calibration step is teachable in one session and generalizes to any "small model for bulk, frontier model for reference" workflow. - **Funding models.** Penn's cost math frames local AI as free at the margin; at Sonnet 5 pricing the same job ranges from $40 to $200 depending on the input/output split, which shifts the case from savings to privacy for smaller archives. ## 🧠 Learning Styles Assessment - **Reading level estimate.** - Flesch-Kincaid Grade Level 6.7 (Flesch Reading Ease 73.2) for the caption transcript: 4,220 words, 286 sentence marks. - Grade 9.9 (Reading Ease 64.5) for the written edition's main section: 2,542 words. - Computed this session with a heuristic syllable counter. Caption punctuation is machine-inserted, so the transcript score understates difficulty. #### 👀 Visual learners - **Present.** Screen shares of the slide image folders, the oMLX settings panel, the token dashboard (23.8M), and SQLite tables. - **Improve.** Penn couldn't enlarge the database view at 12:27. A pipeline diagram (PPTX → images → VLM → SQLite → RRF → deck) would carry the architecture better than table browsing. #### 🔊 Aural learners - **Present.** Full narration; the written edition links an MP3-only version. - **Improve.** Spoken field names in Part 3 run together in one sentence; pausing between the 8 fields would help. #### 🗣️ Verbal learners - **Present.** Definitions of LLM, VLM, omni model, BM25, RRF; the layer-cake analogy for quantization; a 30-term AI-generated glossary in the written edition. - **Improve.** Define "embeddings bottleneck" directly; Penn names it without explaining it. #### 🧩 Kinesthetic learners - **Present.** Named, reproducible steps and free tools (LibreOffice, SQLite, sqlite-vec, oMLX). - **Improve.** Share the shell script, the Python cataloging script, and the schema prompt so viewers can run the steps. #### 🧠 Logical learners - **Present.** 5P structure, numbered Parts 1–5, cost arithmetic. - **Improve.** A table comparing cosine-only, BM25-only, and RRF on sample queries would test the retrieval claim. #### 🧑‍🤝‍🧑 Social learners - **Present.** Framed as "working in public," responding to Wil Reynolds' shared video; Analytics for Marketers Slack mentioned in the written edition. - **Improve.** Invite speakers to share their own slide schemas for comparison. #### 🧘‍♀️ Solitary learners - **Present.** A single-builder project that can be followed alone. - **Improve.** A self-check list: deck count, slide count, schema defined, 10-slide calibration done. #### 🌿 Naturalistic learners - **Present.** Not applicable to this source: no environmental or ecosystem framing. - **Improve.** Report energy use for the 12-hour local run next to the API cost comparison. ## 📊 Accessibility Assessment #### Screen reader semantics - **Present.** The written edition uses headings for Parts 1–5 and a glossary with term headings. - **Improve.** Not verified this session whether images in the written edition carry alt text; the Substack HTML wasn't audited for it. #### Keyboard navigation - **Present.** Not applicable to this source: the video plays in YouTube's player, and Penn controls neither player navigation nor Substack's page chrome. - **Improve.** Not applicable to this source: no custom interactive UI is offered. #### Contrast and readability - **Present.** Not verified: this analysis used captions and the written text, not the video frames. - **Improve.** Screen shares of small database tables at 12:27 are likely unreadable at typical playback size; Penn notes he can't enlarge the window. #### Text resizing and reflow - **Present.** The written edition is a text newsletter page, which reflows as HTML. - **Improve.** Not applicable to this source for the video beyond providing the written edition. #### Media and captions - **Present.** YouTube auto-captions, a written edition covering the same content, and an MP3-only version. - **Improve.** - Auto-captions misrender at least 12 proper nouns (Qwen, Claude Code, soffice, PostgreSQL, Garr Reynolds, Nancy Duarte, Katie Robbert, and others) and one number (304 → 3004). Corrected captions would meet WCAG 2.2 SC 1.2.2 more fully. - Screen-share segments (8:57–10:37, 12:27–13:30) show UI state that narration only partially describes; this relates to SC 1.2.5 audio description. #### Forms and interactions - **Present.** The written edition's one-click rating links (Good / Neutral / Bad) use emoji plus text labels. - **Improve.** Not applicable to this source for the video. #### Standards reference - **1.1.1 Non-text Content.** Penn's image-only slides need text alternatives when decks are shared; the catalog's text and visual-element fields could supply them. - **1.2.2 Captions (Prerecorded).** Auto-captions exist, but name errors reduce accuracy. - **1.2.5 Audio Description (Prerecorded).** Screen-share segments lack full description. ## 🌍 DEIB Assessment #### Cultural breadth - **Present.** Tools span US (LibreOffice community, SQLite), Chinese (Alibaba Qwen, Zhipu GLM-5.3), and NVIDIA (Parakeet) origins; events listed are all US cities (Kansas City, Philadelphia, Cleveland, Austin, Boston). - **Improve.** Note Parakeet v3's 25-language coverage for non-English archives. #### Inclusive language - **Present.** Penn defines LLM, VLM, BM25, RAG, and RRF on first use; the written edition adds a 30-term glossary. - **Improve.** Idioms such as "cost me an arm and a leg" and "down and nerdy" may not translate for non-native English speakers. #### Neurodiversity support - **Present.** Numbered Parts 1–5 give predictable structure. - **Improve.** The oMLX settings tangent at 9:56 ("ooh, there's a new feature") breaks the sequence; chapter markers would let viewers skip it. #### Representation - **Present.** Named people: Wil Reynolds, Garr Reynolds, Nancy Duarte, Katie Robbert, "John," Tristan Buckmaster (in press coverage, not named by Penn). - **Improve.** Not assessable further: no demographic information is given in the source, and none should be inferred. #### Psychological safety - **Present.** Self-deprecating framing ("clip art, which is a favorite pastime of mine") and "for good or ill" about Penn's own slide design. - **Improve.** The OpenAI theft allegations are presented as settled context ("OpenAI stole our work"); labeling them as allegations would treat all parties fairly. ## ✅ Logical Validity #### Strong reasoning - **Platform choice (3:29).** Deductive fit between task and tool: indexing existing slides doesn't require generation, so a smaller local model suffices. Penn backs it with a calibration test (10 slides against GLM-5.3). Clarify by publishing the agreement rate. - **Script over agent (7:02).** Causal and practical: an AI agent converting files runs the same underlying tool while spending usage limits; a script is reusable at zero marginal cost. - **Hybrid retrieval (14:12).** Analogical contrast with Wil Reynolds' single-context slides: when items appear in many contexts, one embedding per item blurs them, so adding lexical ranking helps. Clarify with retrieval metrics on real queries. - **Keeping slide creation human (16:04).** Sound criterion: a speaker who didn't make the slides won't know the deck's content on stage. #### Fallacies - **Non sequitur (18:43).** > [!quote] ID:rw_01m2eg9jqhd3dqv17288mhqkfp > If you've got sensitive proprietary information, building a second brain like this using local AI is how you keep your data safe. - Weakness: the OpenAI dispute as reported involves an unpublished mathematical approach reaching OpenAI before publication, per Cybernews and Entrepreneur; it isn't a case of data leaking from a user's files through a cloud AI tool. The press also gives the problem's age as 90 years, not 100. - Fix: cite a documented case of cloud AI exposing user data, or argue from the providers' data-retention terms directly. - **Inflated comparison (10:37–11:21).** Cost figures treat all ~20M tokens as output and use Opus 5's fast-mode output rate ($50/M) instead of standard ($25/M). - Weakness: overstates cloud cost by up to 5× for Sonnet 5 ($40 at all-input vs. $200 at all-output) and 2× for Opus 5. - Fix: report the input/output token split, as the written edition partly does by saying "hundreds of dollars." - **Hasty generalization (1:11).** "Our language processors in our brains can only handle one language stream at a time." - Weakness: stated as settled cognitive science with no source. - Fix: cite research on reading-while-listening redundancy, or soften to a claim about audience attention that Penn has observed. - **Equivocation (13:30–14:12).** Penn attributes the "embedding bottleneck" to cosine similarity. - Weakness: cosine similarity is a distance metric; the bottleneck comes from representing a multi-topic slide as one vector. - Fix: name the representation as the limit, and note that multi-vector or chunked embeddings are another remedy besides BM25. #### Tone of voice - Casual and self-deprecating: "let's get down and nerdy," "I ain't paying that," "My slide decks ain't worth a thousand bucks." - Promotional in places: - Repeated full trademark name ("we trademark that term. It's why I always have to say it all the way out loud"). - A course plug. - A closing block addressed to AI crawlers with speaker-booking language. - Live-demo register: unscripted asides while browsing oMLX settings at 9:56. ## 🛡️ DARVO Analysis - **Deny.** Not applicable to this source: Penn faces no accusation and denies none. - **Attack.** Not applicable to this source: no accuser's credibility is attacked. Penn relays mathematicians' allegations against OpenAI second-hand but doesn't attack either party. - **Reverse victim/offender.** Not applicable to this source: no role reversal appears. - **Emotional cues and rhetoric.** The OpenAI example carries a fear appeal ("how you keep your data safe") tied to the non sequitur above; this is persuasion, not DARVO. ## 🎯 Conclusion - **Bottom line.** Penn's build shows a local, low-cost path from about 7,000 image slides to a searchable catalog and a draft-deck generator. The durable parts: - An explicit schema. - A 10-sample calibration against a frontier model. - Hybrid BM25-plus-embedding retrieval with RRF. - Keeping slide authorship human. - **Next steps for a speaker with a second brain:** 1. Count decks and slides in your archive. 2. Adopt or adapt the 8-field schema. 3. Run a 10-slide local-vs-frontier comparison before any bulk run. 4. Treat the cost savings in the video as an upper bound and the privacy benefit as the stronger reason. ## 🔄 Refactor Appendix — 🗞️ Almost Timely News: Creating A Keynote Speaker Second Brain (2026-09-13) ### 📚 Literature Atoms #### [H1] Indexing existing slides needs only a local model > [!quote] ID:rw_01m2eg8yvjwqq5h8tny6988rjc > Local AI like Quen 3.6 running on a Mac is more than enough to do indexing and cataloging. **Context & use**: - [3:29](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=209), Part 1, Platform step of the 5P Framework. "Quen" is the caption spelling of Qwen. - Sets the model tier by task: cataloging existing material doesn't need frontier generation. **Tags**: #xref **Highlight**: [view highlight](https://read.readwise.io/read/01m2eg8yvjwqq5h8tny6988rjc)  #### [H2] AI-written scripts beat agents for repeat file conversion > [!quote] ID:rw_01m2eg93abpncccrjetjjyeqe7 > you have a piece of static, reusable code for the next time you want to do the task. **Context & use**: - [7:02](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=422), Part 2, after converting 64 decks with `soffice`. - Applies to any recurring conversion in a vault pipeline: have AI write the script once, then run it without spending usage limits. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://read.readwise.io/read/01m2eg93abpncccrjetjjyeqe7)  #### [H3] Rank fusion avoids the single-embedding retrieval bottleneck > [!quote] ID:rw_01m2eg9d53y0v2e8qm1jpg3qvx > by using reciprocal rank fusion, we're avoiding the problems with a pure embedding rag solution, which is the embeddings bottleneck. **Context & use**: - [14:50](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=890), Part 3, contrasting Penn's multi-context slides with Wil Reynolds' single-context set. - Relevant when one note or slide serves several topics. **Tags**: #xref **Highlight**: [view highlight](https://read.readwise.io/read/01m2eg9d53y0v2e8qm1jpg3qvx)  #### [H4] Prompt tuning lifts small-model output near frontier quality > [!quote] ID:rw_01m2eg97xj6gy5fqp9g897zvy2 > You can make a small model deliver high quality if you tune you invest time to tune it. **Context & use**: - [11:54](https://www.youtube.com/watch?v=3S7B8r7GzWk&t=714), after the 10-slide comparison with GLM-5.3 (about 15 minutes of tuning). - Supports running bulk jobs on local models once a reference sample exists. **Tags**: #xref #articleseed **Highlight**: [view highlight](https://read.readwise.io/read/01m2eg97xj6gy5fqp9g897zvy2)  ### 🧠 Permanent Candidates #### [P1] A frontier-model reference sample makes a small model trustworthy for bulk work **Claim** (one sentence, Courtney's voice): Before handing a bulk job to a small local model, I run 10 representative items through a frontier model and tune prompts until the small model's output matches. **Evidence pointer**: [[#[H4] Prompt tuning lifts small-model output near frontier quality]] **Scope & conditions**: Holds for extraction and description tasks with a fixed schema (Penn's 8 slide fields). It's weaker for open-ended generation, where "close to the same quality" has no field-level check. Penn reports no agreement metric, so the claim stays a working method. **Links** (≥1 with annotation): - [[Engineering Managers Ditch Cloud AI for Local LLMs]] — because extends - [[You Ask, I Answer — Replacing Fine Tuned Models With RAG]] — because contrasts on tuning prompts versus tuning weights **Optional**: `proof_state: working` - Tests / falsifiers: field-level agreement below 80% on a 50-item held-out set falsifies it for that task.  #### [P2] Items reused across contexts need lexical and semantic retrieval together **Claim** (one sentence, Courtney's voice): When one slide or note serves several topics, I retrieve it with keyword ranking and embeddings merged, because one embedding per item blurs its contexts together. **Evidence pointer**: [[#[H3] Rank fusion avoids the single-embedding retrieval bottleneck]] **Scope & conditions**: Matters when items appear in 3+ domains, like Penn's 5P slide. A single-context collection, like Wil Reynolds' 304 narrated slides, may do fine with cosine similarity alone. **Links** (≥1 with annotation): - [[You're doing RAG wrong (Akshay)]] — because extends - [[Building a Second Brain and the Zettelkasten Method]] — because applies-in personal knowledge vault retrieval - [[You Ask, I Answer AI Tools for Messy Notes to Slides]] — because applies-in turning notes into talk slides  ### ✏️ Fleeting Queue - [F1] Check whether `soffice --convert-to` exports every slide of a PPTX or only the first; guides suggest PPTX → PDF → images — https://jdhao.github.io/2020/03/30/pptx_to_image/ #fleeting #inbox - [F2] Verify whether core PostgreSQL ranks with BM25 or needs an extension, before repeating Penn's claim — https://almosttimely.substack.com/p/almost-timely-news-creating-a-keynote #fleeting #inbox - [F3] Find Wil Reynolds' slide-catalog video to compare the cosine-similarity design — https://www.seerinteractive.com/wil-reynolds #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Count decks and slides in my talk archive — baseline before building a slide catalog (Penn: 64 decks, ~7,000 slides) #things - [ ] Run a 10-slide test of Penn's 8-field schema on a local model vs. a frontier model — https://almosttimely.substack.com/p/almost-timely-news-creating-a-keynote #things  📄 **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 Penn]] ## Highlights - [3:29] on themes, topics, and keywords with the understanding that some slides may be thematically appropriate but wildly out ofd. When we hit a situation like that, it has to flag that slide for me to outright replace platform. For an application like this, we're not assembling net new slides, per se. We're merely indexing the data that already exists. That means we don't need expensive state-of-the-art AI. Local AI like Quen 3.6 running on a Mac is more than enough to do indexing and cataloging. I'll build the infrastructure with a frontier model, but I'll run it locally. And finally, performance. In the short term, a working application is the definition of [4:06] done, which means a knowledge catalog of all my slides. In the long term, time savings and increased audience satisfaction of the benchmark. But for today, a working application itself is our measure of success. If you sit down and you go through the 5P framework by trust insights, it's pretty much your entire project plan in a nutshell. This plus an interactive interview. You say to the AI, ask me up to 20 clarifying questions until you have enough information to complete the task [snorts] is enough uh to to for any AI coding tool of your choice to build requirements, a technical spec, and a work plan. Part two, slides prep. In ([View Highlight](https://read.readwise.io/read/01m2eg8yvjwqq5h8tny6988rjc)) - Note: Indexing existing slides needs a local vision model (Qwen 3.6 on a Mac), not a frontier API; frontier model builds, local model runs. - Tags: atom/literature/h1 - [7:02] I'd rather skip the middleman or middle machine, I guess, uh, and do it myself, or barring that, have AI write the script for you, but then you run it yourself. The bonus of it doing it this way is you have a piece of static, reusable code for the next time you want to do the task. And if you have a good process down, you can even give the script to your favorite AI tool to operate on your behalf. [7:22] Part three, cataloging the slides. To catalog slides, it's not enough to tell AI to transcribe the slide. If our purpose is to use the index for finding themes and assembling new decks, then we need to be very, very specific about what data we're collecting. Before passing a single slide through our VLM, we need to be very clear about what we need. And I came up with this list. the verbatim text of the slide. What visual elements are on the slide, a one-s sentence summary of the slide, five to 25 keywords, a slide type like image, uh video, animation, chart, graph, etc. A slide domain, data science, AI, ([View Highlight](https://read.readwise.io/read/01m2eg93abpncccrjetjjyeqe7)) - Note: Action: have AI write the deck-to-image script once, run it yourself, and define the per-slide schema before any model sees a slide. - Tags: atom/literature/h2, action/h1 - [11:54] tune the prompts until Quen reached close to the same level of quality. This took probably 15 minutes up front. If you design your software with tools like like LLM that enable to swap models quickly, you don't have to reinvent the wheel and you you get really good quality. You can make a small model deliver high quality if you tune you invest time to tune it. Quen 3.6 fully populated the database with all 7,000 slides, their chunks, and their embeddings. The database itself is also really important. Let's bring up the database here. There's the database. [12:27] The database is really important because look at all those slides. Can I make that bigger? I don't think I can make the interface bigger, can I? Window. No, doesn't it doesn't get bigger. >> [clears throat] >> um because this is running locally on my Mac and isn't a production application. SQLite which is what I'm using here uh is the free open- source database. If you go to uh SQLite.org u is very very very powerful local database uh is perfectly fine with two add-ons. There's a vector storage component called SQLite vec that allows you to store embeddings. So if I go to my slide embeddings here ([View Highlight](https://read.readwise.io/read/01m2eg97xj6gy5fqp9g897zvy2)) - Note: Calibrate a small local model against a frontier model on 10 representative items before a bulk run. - Tags: atom/zettel/z1 - [14:12] dealing with a slide that has multiple topics and multiple uses for example the 5p framework by trust insights is one of my slides. It's seen a lot that is not in one context. If you look at and watch will's video a lot of his slides are in the same context. So cosign similarity makes total sense there. in mind, I can apply the 5P framework to AI, to analytics, to data science, to cooking, to auto repair. And so BM25 allows us to to pair the embeddings with uh lexical search. And if you pair lexical search with embeddings, you get something what uh the technical term is [14:50] called reciprocal rank fusion, which is the best of both worlds. It allows you to narrow down very quickly. Here's what this slide is about, the lexical side, and here's what it's related to, the embedding side. And you want to have both. For the technically minded, what we've created is essentially a custom retrieval augmented generation or rag solution for a massive number of of slide decks. And by using reciprocal rank fusion, we're avoiding the problems with a pure embedding rag solution, which is the embeddings bottleneck. And [snorts] by using embeddings, we're avoiding the lexical trap of if you don't use the exact phrase or words in a ([View Highlight](https://read.readwise.io/read/01m2eg9d53y0v2e8qm1jpg3qvx)) - Note: Hybrid retrieval: BM25 lexical plus embeddings, merged with reciprocal rank fusion, for slides reused across many contexts. - Tags: atom/literature/h3 - [18:43] with the help of local AI tools you run on your machine and classical databases enabled for AI. This has been center sta stage this past week because OpenAI claims to have solved a 100-year-old math problem and some mathematicians are saying OpenAI stole our 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. All right, what else happened this week? We had a a fun live stream this week. John and I talked through how to do non-traditional podcast analytics and we're looking at the Trust Insights podcast and all the [19:28] different ways you could unify your data and that was a lot of fun. We actually uh I learned some stuff uh on that episode that was worth uh spending some time with on the uh in-ear insights podcast I do with Katie every week. Uh we talked about uh her journey towards local AI, the non- techies journey as she's starting to build out what it would take for her to run models like Quinn. Now she has a 32 GB Mac and so she wants her computer to be at all usable. She's probably going to be using one of the much smaller ver variants of Quinn, like the 9 billion parameter ([View Highlight](https://read.readwise.io/read/01m2eg9jqhd3dqv17288mhqkfp)) - Note: Press reports on the Navier-Stokes dispute describe an approach reaching OpenAI by word of mouth, not data leaking through OpenAI tools, so the local-AI safety conclusion doesn't follow from this example. Also: the problem is reported as 90 years old, not 100. - Tags: fallacy/non-sequitur