Clown Cast

Clown Cast

By Joey MusselmanTechnologyTech News
Download on the App Store

Clown Cast episodes

  • Where Your Brain Ends: Obsidian and the Art of Keeping Knowledge
    Episode 200: We're drowning in scattered notes across six apps and zero organization. Where do you actually keep what you know? Today we're exploring Obsidian, the note-taking tool that's quietly become essential for developers, writers, and researchers building their entire intellectual lives inside one place. We trace the lineage from Renaissance commonplace books through modern AI knowledge layers, and ask: is Obsidian industry tool or personal obsession?
    0:00 - The Knowledge Crisis: Where does it all live?
    2:30 - The Problem: 700 notes, zero memory of them
    5:15 - Obsidian Introduction: The conspiracy board graph view
    8:00 - Commonplace Books: 500 years of knowledge management
    12:30 - Obsidian in Practice: How developers actually use it
    16:45 - The AI Connection: Why knowledge layers matter now
    ---
    Sources & further reading:
    • [eesel AI — Obsidian AI explained: How it enhances knowledge management](: https://www.eesel.ai/blog/obsidian-ai
    • [Volodymyr Pavlyshyn — Obsidian, Supercharged: The AI Revolution in Personal Knowledge Management](: https://volodymyrpavlyshyn.substack.com/p/obsidian-supercharged-the-ai-revolution
    • [Eric Ma — Mastering Personal Knowledge Management with Obsidian and AI](: https://ericmjl.github.io/blog/2026/3/6/mastering-personal-knowledge-management-with-obsidian-and-ai/
    • [Obsibrain — What Is Obsidian Used For? A Practical 2026 Guide](: https://www.obsibrain.com/blog/what-to-use-obsidian-for-in-2026
    • [DEV Community — The Missing Piece Every Obsidian User Needs: Local RAG That Actually Works in 2026](: https://dev.to/numbpill3d/the-missing-piece-every-obsidian-user-needs-local-rag-that-actually-works-in-2026-2gfp
    • [FrankX — Obsidian Second Brain: AI Agents, MCP and Agent Skills](: https://www.frankx.ai/guides/agentic-obsidian-second-brain
    • [MorphLLM — Obsidian MCP Server: Connect Your Vault to AI Agents](: https://www.morphllm.com/obsidian-mcp-server
    • [NexaSphere — Best Note-Taking Apps for Developers in 2026](: https://nexasphere.io/blog/best-note-taking-apps-developers-2026
    • [DasRoot — Obsidian vs Logseq vs Notion: PKM Systems Compared 2026](: https://dasroot.net/posts/2026/03/obsidian-logseq-notion-pkm-systems-compared-2026/
    • [eesel AI — Obsidian pricing 2026: Free vs paid plans compared](: https://www.eesel.ai/blog/obsidian-pricing
    • [obra/knowledge-graph — GitHub: Query and traverse an Obsidian vault as a knowledge graph](: https://github.com/obra/knowledge-graph
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    21 min
  • Everything in Its Place: How AI Finds Knowledge
    When you ask an AI about something new, where does it find the answer? Host A and Host B explore the fundamental difference between RAG (Retrieval-Augmented Generation) and context skills—two competing philosophies for teaching AI agents knowledge beyond their training data. Using a professional kitchen's mise en place discipline as their guide, they break down how knowledge layers work, why trust matters, and why the future belongs to layered approaches, not single solutions.
    00:00 - The core problem: Knowledge cutoffs and frozen training data
    03:20 - Mise en place: The kitchen metaphor explained
    05:50 - RAG: Real-time retrieval and dynamic lookup
    10:15 - Context skills: Pre-loading and structured knowledge
    13:45 - Layering strategies: When to use each approach
    15:30 - 2026 frontier: Hybrid architectures and knowledge trust
    ---
    Sources & further reading:
    • Patrick Lewis et al. — "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020): https://arxiv.org/abs/2005.11401
    • VentureBeat — "The RAG era is ending for agentic AI — a new compilation-stage knowledge layer is what comes next": https://venturebeat.com/data/the-rag-era-is-ending-for-agentic-ai-a-new-compilation-stage-knowledge-layer-is-what-comes-next
    • arXiv — "Skill Retrieval Augmentation for Agentic AI": https://arxiv.org/html/2604.24594v1
    • Andrej Karpathy — context engineering framing (June 2025, expanded 2026) — (originally Twitter/X posts)
    • The Context Graph — "Context Engineering in 2026: From Karpathy's Tweet to Production Infrastructure": https://thecontextgraph.co/memos/context-engineering-2026-from-tweet-to-infrastructure
    • AI Builder Club — "Andrej Karpathy's AI Engineering Playbook (2026)": https://www.aibuilderclub.com/blog/karpathy-ai-engineering-playbook
    • Towards Data Science — "Is RAG Dead? The Rise of Context Engineering and Semantic Layers for Agentic AI": https://towardsdatascience.com/beyond-rag/
    • RAGFlow — "From RAG to Context — A 2025 year-end review of RAG": https://ragflow.io/blog/rag-review-2025-from-rag-to-context
    • Redis — "Context Retrieval for AI Agents: What It Is & Why It Matters": https://redis.io/blog/context-retrieval-for-ai-agents/
    • Neo4j — "Knowledge graph vs. vector RAG: Benchmarking, optimization levers": https://neo4j.com/blog/developer/knowledge-graph-vs-vector-rag/
    • CMARIX — "RAG & AI Trust Statistics 2026: Beating Hallucinations": https://www.cmarix.com/blog/rag-ai-statistics/
    • yAI — "Why Citation-Based RAG Still Hallucinates": https://yaihq.com/research/citation-based-rag-still-hallucinates
    • arXiv — "FACTUM: Mechanistic Detection of Citation Hallucination in Long-Form RAG": https://arxiv.org/pdf/2601.05866
    • arXiv — "Knowledge Activation: AI Skills as the Institutional Knowledge Primitive": https://arxiv.org/pdf/2603.14805
    • Startup GTM — "Build Karpathy's Self-Updating AI Knowledge Base in Just 90 Minutes": https://startupgtm.substack.com/p/self-updating-ai-wiki-knowledge-base
    • Atlan — "Knowledge Graph vs RAG: When Each One Wins (2026)": https://atlan.com/know/knowledge-graphs-vs-rag-for-ai/
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    18 min
  • The Devil You Know: Why Dead Solutions Haunt Our Keyboards
    Why do we still use a keyboard layout designed to prevent typewriter jams that haven't existed for a century? This week, we explore lock-in effects—the phenomenon where solutions persist long after their original problems vanish, trapping entire industries and technologies through the sheer cost of switching. From QWERTY to railroad gauges to programming languages, discover how historical accidents become permanent features of our world, and why the economics of change often defeat the logic of progress.
    00:00 - Why Your Keyboard Layout Exists
    03:45 - The Typewriter Jam That Changed Everything
    07:20 - When the Problem Disappeared (But the Solution Stayed)
    10:15 - Defining Lock-In: Paul David & W. Brian Arthur
    13:30 - The Economics of Being Stuck
    ---
    Sources & further reading:
    • Paul David — "Clio and the Economics of QWERTY," American Economic Review 75(2), 1985: https://econ.ucsb.edu/~tedb/Courses/Ec100C/DavidQwerty.pdf
    • W. Brian Arthur — "Competing Technologies, Increasing Returns, and Lock-In by Historical Events," The Economic Journal 99(394), 1989: http://www.critical-juncture.net/uploads/2/1/9/9/21997192/arthur_competing_technologies_increasing_returns_and_lock-in_by_1989.pdf
    • Stan Liebowitz & Stephen Margolis — "The Fable of the Keys," Journal of Law and Economics 33(1), 1990: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1069950
    • W. Brian Arthur — "Increasing Returns and Path Dependence in the Economy" (collected essays), University of Michigan Press, 1994: https://sites.santafe.edu/~wbarthur/increasingreturns.htm
    • Encyclopedia.com — "Railroad Gauges, Standardization of": https://www.encyclopedia.com/history/encyclopedias-almanacs-transcripts-and-maps/railroad-gauges-standardization
    • Snopes — "Are U.S. Railroad Gauges Based on Roman Chariots?" (debunking the chariot myth): https://www.snopes.com/fact-check/railroad-gauge-chariots/
    • Gareth Dennis — "The not-so-glamourous origins of standard track gauge": https://garethdennis.medium.com/the-not-so-glamourous-origins-of-standard-track-gauge-2b5f1ae7e3bc
    • ScienceBlog — "A 65-year-old programming language called COBOL still quietly processes over $3 trillion in banking transactions every single day": https://scienceblog.com/a-65-year-old-programming-language-called-cobol-still-quietly-processes-over-3-trillion-in-banking-transactions-every-single-day-and-because-the-original-engineers-are-rapidly-retiring-ban/
    • The New Stack — Brendan Eich on creating JavaScript in 10 days: https://thenewstack.io/brendan-eich-on-creating-javascript-in-10-days-and-what-hed-do-differently-today/
    • The History of the Web — "The 10-Day Programming Language (Is Kind of a Myth)": https://thehistoryoftheweb.com/the-10-day-programming-language-is-a-myth/
    • ANSI Blog — "Why Doesn't the U.S. Use the Metric System?": https://blog.ansi.org/ansi/why-does-the-u-s-not-use-the-metric-system/
    • LegalClarity — "Why Doesn't the US Use the Metric System? History and Costs": https://legalclarity.org/why-doesnt-the-us-use-the-metric-system-history-and-costs/
    • VaasBlock — "Enterprise AI Vendor Lock-In: The Switching Cost Problem No One Is Measuring" (2026): https://www.vaasblock.com/research/enterprise-ai-vendor-lock-in-switching-costs-copilot-agentforce-2026/
    • CEPR — "Abolishing Daylight Saving Time is easy, setting a permanent time is not": https://cepr.org/voxeu/columns/abolishing-daylight-saving-time-easy-setting-permanent-time-not
    • Costin Smilovici (Bactra) — "QWERTY, Lock-In and Path Dependence" (notebook/synthesis): https://bactra.org/notebooks/qwerty.html
    • Galbraith & Kay — "QWERTY, the search for optimality and IP complementarity" (2025): https://strathprints.strath.ac.uk/92487/1/Galbraith-Kay-BH-2025-QWERTY-the-search-for-optimality-and-IP-complementarity.pdf
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    17 min
  • Skeleton Crew: Inside Machine Vision
    What the machine actually sees is almost nothing like what you see. When a volleyball AI watches a broadcast, it doesn't see athletes and a ball—it sees 16 skeletal dots per person and misses the ball 73% of the time. In this trilogy closer, hosts explore the gap between human perception and machine input, revealing how the representation you hand a model becomes its permanent ceiling.
    00:00 - Introduction: The fundamentals trilogy conclusion
    01:00 - The perception gap: What humans see vs. what machines see
    03:00 - Skeleton detection: 16 skeletal points per athlete
    06:00 - Ball detection: The 27% problem
    10:00 - Court mapping: Converting pixels to meters
    14:00 - Why representation shapes everything
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md §§1.3, 2, 5 (the
    • input inventory; the flat-projection autopsy; the ranker domain-shift and leakage
    • stories)
    • game-senser-fable repo — reference/FINDINGS-2026-06-04.md (ball vs feet projection; the
    • 19–26m measurement; per-skill breakdown)
    • game-senser-fable repo — baseline/BASELINE_SCORES.md (25.4 vs 17.9; the degenerate
    • sparse homography)
    • game-senser-fable repo — JOURNAL.md iters 3–5 (pose candidates 99.6%; the ranker's two
    • poisons and the scale-free fix)
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    15 min
  • Generous Grading: The Invisible Rule Behind Every Number
    When a tennis AI system scored 90.8% on identical footage with zero code changes, the rescoring dropped to 62.9%. Nothing broke—the matching rule just tightened. Host A and Host B uncover why every metric hides a procedure, tracing seven episodes of dead numbers back to this fundamental truth: a metric is a procedure, not a number.
    00:00:00 - Opening: The 90.8% to 62.9% collapse
    00:01:00 - What actually changed: The matching rule
    00:04:00 - A chain of dead metrics from previous episodes
    00:07:00 - Core lesson: Metrics are procedures
    00:10:00 - Hidden tolerances and invisible criteria
    00:12:00 - Closing thoughts
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md §§1.1, 1.3, 2, 6
    • (the 90.8→62.9 re-measure; the zone floor and its denominator; the evaluation onion
    • the taste-check record)
    • game-senser-fable repo — baseline/BASELINE_SCORES.md (25.4% over 63/182, per-skill
    • splits)
    • game-senser-fable repo — JOURNAL.md iters 13–15, 36–38 (the yardstick calibration
    • chain; the three-stage honesty ladder)
    • game-senser-fable repo — HUMANINPUTLOG.md (the taste-check institution, costs and
    • impacts)
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    13 min
  • Poison Gold: The Truth About Ground Truth
    What actually IS ground truth? And who makes it? In this episode, we finally answer the questions we've been hinting at all season: the human work that powers machine learning models, why it matters infinitely more than fancy algorithms, and what happens when your foundation is contaminated. This is the poison gold episode—where four hours of human attention powered 14 months of measurement.
    00:00:00 - Where does ground truth come from?
    00:02:30 - The single most leveraged artifact
    00:04:15 - The 719 touches: hand-tagging at scale
    00:10:00 - Why labeling is the real return on investment
    00:14:45 - When poison contaminates the gold
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md §§1.2, 3.3 (the
    • golden session; the prefilled labeler autopsy)
    • game-senser-fable repo — GAPS.md (the circular-GT premise autopsy)
    • game-senser-fable repo — HUMANINPUTLOG.md (the 4-hour session; the unused outcome
    • labels; the rally-67 catch)
    • game-senser repo — app/labeler (the prefill mechanism, cited via GAPS.md)
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    17 min
  • The Frozen Bar: When Good Isn't Enough
    B3 achieved 72% after human audit—a significant jump from V2's 46%—but fell short of the frozen 80% bar set before grading began. Using the high jump metaphor, the hosts explore why an impressive personal best still means no advancement when the line is drawn in advance, and how the system itself worked even when the pipeline didn't. A meditation on standards, measurement, and why the bar doesn't move.
    00:00 - B3 Results & The Frozen 80% Line
    03:45 - The High Jump: Why Personal Bests Don't Lower Bars
    07:20 - V2 vs V3: Architecture Changes
    10:30 - Window Placement & Rally Detection Problems
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md §9 (the precondition
    • session and the §9.6 audit)
    • game-senser-fable repo — src/gstaadwindowsv3.py + src/out/gstaadwindowsv3.json (the
    • frozen build)
    • game-senser-fable repo — src/out/gstaadv3eval.json + gstaadv3humancheckkey.json +
    • HUMANINPUTLOG.md 2026-07-28 (every verdict, the pre-committed key, the audit)
    • game-senser-fable repo — src/out/saqpresence517318.json (the banked Saquarema
    • anatomy)
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    14 min
  • Feeding the Beast: What Boundary Models Actually Eat
    What does a boundary model actually eat? This episode serves up the answer: a carefully curated menu of portable features that actually train models with minimal data. Host A and Host B unpack the human-machine collaboration at the heart of their project—Joey sets the questions and validates the labels, while an overnight AI agent session catalogs every signal the pipeline can compute, from whistle detection to scoreboard OCR. A practical deep dive into feature engineering on a sticky note budget.
    00:00:00 - Opening: Breaking the curse
    00:02:45 - The human-machine division of labor
    00:06:30 - The feature menu: Inside the overnight session
    00:12:00 - Portability: The secret ingredient
    00:16:15 - Closing thoughts
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md §8 (the ep-04 work
    • session: feature menu, label census, JP pool/semifinal scans, containment/slop
    • measurement, the generalization bet; commit c5d53d6)
    • game-senser-fable repo — src/out/jppoolsfbaranatomy.json +
    • src/out/gstaadcontainmentslop.json (the night's primary measurements)
    • game-senser-fable repo — src/out/gstaadsample10v2verdicts.json + HUMANINPUT_LOG.md
    • (the ~20 boundary verdicts; the label-cost ladder)
    • game-senser-fable repo — data/514945/rallybounds.json + goldparsed_514945.json +
    • data/468157 manual rows (the label estate census)
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    18 min
  • From the Stands to the Algorithm: When Your Pipeline Catches Up
    The experiment series reached a turning point: Joey watched the G-Stod 2020 volleyball finals live in Switzerland, then went home and fed that same match footage to the machine learning pipeline. This episode documents the technical breakthrough when AI analysis caught up to real-time—and what the cleanest scoreboard reads yet revealed about the project's trajectory.
    00:00 - Correction: Attribution for AI contributions
    01:15 - The bottleneck that just broke: from 50-minute manual capture to same-day pipeline
    03:45 - Joey's Swiss Alps moment: watching volleyball live
    05:30 - Pipeline staging chain: marking, scoreboard detection, whistle recognition
    07:00 - The results: only 8 misreads across the entire match
    09:20 - Evaluation: 10/10 accuracy on random spot-check clips
    12:45 - What this win means for the next phase
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md §7 ADDENDUM
    • (2026-07-27 evening: the dress-rehearsal split, the bar-presence discovery, the
    • reclassification)
    • game-senser-fable repo — JOURNAL.md 2026-07-27 entries + data/gstaad2026/STAGING.md
    • (staging chain, OCR quality, v2 windows)
    • game-senser-fable repo — data/gstaad2026/barpresence532121.json +
    • src/out/gstaadsample10v2verdicts.json + HUMANINPUT_LOG.md (the two 10-clip verdict
    • passes)
    • guist-bot repo — data/series/ml-volleyball/ERRATA.md (the ep-2 attribution correction
    • Joey's series directive on clip cutting)
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    14 min
  • Systems and Gaps: Why Bad Data Breaks ML
    A correction to episode 1 becomes the perfect teaching moment: the one untagged number sailed through quality checks while all the tagged numbers held up. The hosts explore why data provenance is everything in ML systems, then dive into ground truth problems lurking in labeling pipelines. The system worked. The system also had a gap.
    00:00 - Episode 1 correction and the gap that caught them
    03:30 - Why data provenance tags actually matter
    08:00 - Ground truth and the circular labeling problem
    12:00 - What makes something truly an ML problem
    ---
    Sources & further reading:
    • game-senser-fable repo — docs/podcast/LEARNING-CONTEXT.md, including the
    • guist-bot repo — data/series/ml-volleyball/ERRATA.md (the three-years correction
    • 2026-07-27 corrections block (source document for the series)
    • and the ground-truth clarification, from Joey, 2026-07-27)
    • game-senser-fable repo — JOURNAL.md (89-variant sweep; skill classifiers parked twice
    • rally-grammar decoder 44%→64.8%; serve identity 100%)
    • game-senser-fable repo — GAPS.md (the prefilled-labeler circularity; digs 1/22)
    • game-senser-services repo — docs/CV_PIPELINE.md (the 18-experiment history and the
    • hand-written fusion constants)
    • The history of email spam filtering, rules-to-statistical-learning — [GENERAL-PRACTICE]
    • not checked against a primary source for this episode
    This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.
    15 min

About Clown Cast

From the publisher's feed

Podcasts about whatever I find interesting — history, tech, weird rabbit holes. I was making these for myself anyway, so I figured I'd share. Fully AI-generated: Claude researches and writes each…