---
name: whats-viktor-watching
description: "Knowledge brief for the 38 videos Viktor Gamov watched with the /watch skill and archived at watching.ai-assisted.engineering, plus the wiki built from them (51 people and products, 77 concepts). Topics include developer-tools, ai-agents, ai, terminal, agentic-engineering, herdr, tmux, heart-rate, wearables, claude-code, hrv, multiplexer. Use to answer what one of these videos claimed, what the wiki says about a person, product or concept, or which videos cover a topic."
---

# What's Viktor Watching: Archive Knowledge

Process steps in order. Do not skip ahead.

## Step 1 — Match the question

This brief covers 38 videos Viktor Gamov watched between 2026-08-28 and 2026-11-16 with Claude Code and the /watch skill, plus the wiki built from them. Each video has a report at https://watching.ai-assisted.engineering/watched/<dir>/ with a summary, key moments, quotes, entities and frames. A wiki note covers one person, product or concept and cites the reports it came from.

Supported requests:
- what a video or its creator claimed
- what the wiki says about a person, product or concept
- which videos cover a topic, and how they relate
- where to find the original video

For anything else, say this brief does not cover it and finish.

Continue to Step 2.

## Step 2 — Find the sources

Look the question up in the Videos and Wiki lists below. A wiki line holds the note's first paragraph, trimmed if long; its link points to the full note in Markdown. When the line is not enough, fetch the full note; go to the report page only for detail the note lacks. If nothing in the lists matches, say so and finish.

Continue to Step 3.

## Step 3 — Answer

Answer from the notes and reports, and match depth to the question. Keep three things apart: what a video's creator claims, what the report or note summarises, and your own interpretation. Reports and notes repeat what the videos claim without checking it. Quotes in reports come from auto-captions and can contain transcription errors. Never invent quotes or timestamps; for exact wording, link the report, which links the original video. Cite the report or note URL for each claim.

Finish.

## Videos (38, newest first)

- 2026-11-16 · [This Tool Replaced 7 CLIs (and killed my opensource)](https://watching.ai-assisted.engineering/watched/2026-11-16-television-fuzzy-finder-devops-toolbox/) · DevOps Toolbox · Television's TOML channels (source, preview, actions) replace Ranger, a Kubernetes log reader, kubectx and Omer's tmux-sessionx; the key bindings need a leader key.
- 2026-11-15 · [Build a Better Software Factory with Claude Code](https://watching.ai-assisted.engineering/watched/2026-11-15-better-software-factory-claude-code-zazencodes/) · ZazenCodes · Labelled GitHub issues go through Fable intake, Opus spec and build, a checks.sh gate, risk-routed reviewers and an approver that merges or asks a human; only a script writes job state.
- 2026-11-15 · [How AI Agent Memory Works: MemGPT & Memory Types | Stanford CS329Z Lecture 7 with Michael Ryan](https://watching.ai-assisted.engineering/watched/2026-11-15-agent-memory-memgpt-stanford-cs329z/) · Stanford Online · Michael Ryan sorts agent memory into working, episodic, semantic and procedural types, kept in files, databases or graphs, then covers skill libraries, MemGPT tiers, reflection and maintenance; on MemoryArena the memory agents still scored worst.
- 2026-11-14 · [Why I'm betting on Rex](https://watching.ai-assisted.engineering/watched/2026-11-14-why-betting-on-rex-reemus/) · reemus_dev · Rex loses a feature checklist against herdr and Orca. It mirrors raw output to a libghostty client where tmux composites, and reemus backs it for treating the multiplexer and its API as the product.
- 2026-10-29 · [macOS: Essential Productivity Hacks for Developers](https://watching.ai-assisted.engineering/watched/2026-10-29-macos-productivity-hacks-developers-mehdio/) · mehdio debugs · skhd app shortcuts, yabai tiling, and a home space for each app so one shortcut always lands in the right place; SketchyBar shows what is where and Raycast covers the rest.
- 2026-10-22 · [How To Make An Amazing Custom Menu Bar For Your Mac With Sketchybar](https://watching.ai-assisted.engineering/watched/2026-10-22-sketchybar-custom-menu-bar-josean-martinez/) · Josean Martinez · Builds a SketchyBar menu bar from an empty config: items in `items/`, scripts in `plugins/` run on timers or subscribed events, app-icon font for spaces and the front app, and a scrolling media item.
- 2026-10-22 · [Aerospace Is Probably The Best MacOS Tiling Manager I've Ever Used](https://watching.ai-assisted.engineering/watched/2026-10-22-aerospace-tiling-manager-devops-toolbox/) · DevOps Toolbox · Omer swaps yabai and skhd for AeroSpace (tiling plus app launching), adds binding modes, an fzf window switcher, SketchyBar and JankyBorders, and keeps tmux for the terminal.
- 2026-10-21 · [S1E3: An Architect's Verdict on OpenSpec](https://watching.ai-assisted.engineering/watched/2026-10-21-openspec-architects-verdict-tim-berglund/) · Tim Berglund · Architect Michael Carducci went from joking about spec-driven development to running OpenSpec on a production SaaS: a repeatable process, specs that let the agent recover from detours, and a record of why decisions were made.
- 2026-10-18 · [Run Multiple Claude Code Sessions From a Stream Deck (Full Demo)](https://watching.ai-assisted.engineering/watched/2026-10-18-stream-deck-claude-code-not-replaced-yet/) · Not Replaced Yet · Jesse gives each client work stream its own Stream Deck key, configured by Claude: one press reopens the Claude Code thread, two restart it from a saved prompt, three close it; less context switching, many more tokens.
- 2026-10-16 · [The big lie about AI coding](https://watching.ai-assisted.engineering/watched/2026-10-16-markdown-industrial-complex-less-bitter/) · Less Bitter · Mo argues there is no method to AI coding: shared skill files (gstack, Matt Pocock's skills) are superstition, post-training is what makes agents act like engineers, and a 2,500-PR month means 10 to 12 hour days.
- 2026-10-16 · [Don't Buy a Mac Mini For Local AI (Do This)](https://watching.ai-assisted.engineering/watched/2026-10-16-mac-mini-local-ai-kai/) · Kai · Apple charges $25 per GB and even 64 GB can't hold gpt-oss-120b with context; buy the M6 at 32 GB, switch inference servers, run sparse models, and rent before buying big.
- 2026-10-16 · [How to set up Herdr for multi-agent coding (full guide)](https://watching.ai-assisted.engineering/watched/2026-10-16-herdr-setup-guide-dave-ebbelaar/) · Dave Ebbelaar · Coming from the Claude Code and Codex apps, Dave moves to Herdr in Ghostty to mix harnesses, with persistent sessions, a config.toml that Codex wrote, Cmd-key shortcuts, Codex handing reviews to Claude through the Herdr skill, and Neovim plus lazygit in place of an IDE.
- 2026-10-16 · [Every Size Local AI In 24 Minutes](https://watching.ai-assisted.engineering/watched/2026-10-16-every-size-local-ai-tina-huang/) · Tina Huang · Tina runs local models on six hardware classes, from an ESP32 to a rented 8x H100: a model has to fit in RAM, its speed comes from bandwidth and GPUs, and 75% of RAM divided by 0.6 gives the largest model in billions.
- 2026-10-04 · [I Don't Want a New Terminal. I Want THIS for Tmux.](https://watching.ai-assisted.engineering/watched/2026-10-04-workmux-devops-toolbox/) · DevOps Toolbox · Omer stays on tmux and adds workmux: a worktree per agent, a dashboard where `i` lets you type into an agent's preview, a Herdr-style sidebar, and session mode that shows agent trees in his session manager.
- 2026-10-04 · [TUIOS Makes Herdr and Tmux Feel Ancient](https://watching.ai-assisted.engineering/watched/2026-10-04-tuios-devops-toolbox/) · DevOps Toolbox · TUIOS turns the terminal into a small desktop (draggable floating panes, minimize, tiling on demand, which-key style prefix menu, web UI); sessions are unfinished and it has no agent features, so Omer stays on tmux and recommends workmux.
- 2026-09-26 · [Rails World 2026 Opening Keynote - DHH](https://watching.ai-assisted.engineering/watched/2026-09-26-dhh-rails-world-2026-keynote/) · Ruby on Rails · 37signals goes pencils down on hand-written code; Opus 4.5 as the Kodak Brownie of software; HEY moves to native apps plus a Rust backend; DHH's own 150k lines in a month; every app needs a CLI for your own agent; P(bloom) over P(doom).
- 2026-09-24 · [Stanford CS229 I Machine Learning I Building Large Language Models (LLMs)](https://watching.ai-assisted.engineering/watched/2026-09-24-stanford-cs229-building-llms-yann-dubois/) · Stanford Online · Yann Dubois on what industry actually spends effort on: filtered web data, evaluation and GPU systems over architecture. Scaling laws set the size (20 tokens/param optimal, ~150 in production); Llama 3 405B ≈ 3.8e25 FLOPs and ~$75M; SFT only picks a persona, DPO replaced PPO.
- 2026-09-23 · [Tmux is Still Better Than Herdr (Here’s Why)](https://watching.ai-assisted.engineering/watched/2026-09-23-tmux-still-better-than-herdr-workmux/) · Sam Natale · First dissent to the Herdr wave: workmux gives plain tmux worktree-per-branch (auto .env copy + node_modules symlink via .workmux.yaml), an agent sidebar and a dashboard; it composes with tmux display-menu + fzf scripts. Augment the multiplexer you already know.
- 2026-09-23 · [Full Walkthrough: Workflow for AI Coding — Matt Pocock](https://watching.ai-assisted.engineering/watched/2026-09-23-matt-pocock-ai-coding-workflow-ai-engineer/) · AI Engineer · 96-min workshop: smart zone ends ~100k tokens regardless of window; grill-me alignment interviews replace plan-reading; PRD → Kanban DAG of vertical-slice issues → Docker-sandboxed Ralph loops with TDD; feedback loops are the ceiling, deep modules over shallow sprawl, QA is where taste stops slop; "the code is your battleground."
- 2026-09-23 · [How I use LLMs](https://watching.ai-assisted.engineering/watched/2026-09-23-karpathy-how-i-use-llms/) · Andrej Karpathy · Practical sequel to the Deep Dive: "I am a one-terabyte zip file" as the lens; pick features by what must enter the context window (search for fresh facts, uploads for exact text, Python for math, thinking models only when the fast answer fails, deep research as a first draft); every demo shows a failure, so verify like a junior analyst; new chat per topic, voice for half of queries, few-shot custom GPTs.
- 2026-09-23 · [Deep Dive into LLMs like ChatGPT](https://watching.ai-assisted.engineering/watched/2026-09-23-karpathy-deep-dive-llms-like-chatgpt/) · Andrej Karpathy · A 3.5h general-audience explainer: base model = lossy compression of the internet, assistant = simulation of a data labeler, params = vague recollection vs context = working memory, models need tokens to think, reasoning emerges from verifiable-domain RL (DeepSeek-R1's "aha moments," AlphaGo's move 37) while RLHF is a gameable bounded fine-tune; Swiss-cheese capabilities — check work, own output.
- 2026-09-23 · [AlphaGo - The Movie | Full award-winning documentary](https://watching.ai-assisted.engineering/watched/2026-09-23-alphago-the-movie-deepmind/) · Google DeepMind · AlphaGo beats Fan Hui 5-0 then Lee Sedol 4-1 (Seoul 2016): policy net + value net + tree search, imitation then self-play RL; move 37 (1-in-10,000, beyond its human prior) vs Lee's move 78 exposing superhuman blind spots; maximises win probability not margin; both humans come out changed.
- 2026-09-17 · [Apple Watch Ultra 4 In-Depth Review: As Good As the Hype?](https://watching.ai-assisted.engineering/watched/2026-09-17-apple-watch-ultra-4-in-depth-review-dc-rainmaker/) · DC Rainmaker · The accuracy verdict: sensor very accurate vs Polar H10 across six sports — but Apple's study only grades the quality-gated values the watch chose to record, leaving 11-40s gaps unscored ("Okay, Apple"); readiness realistic yet super buggy, Fenix 9 still wins GPS, 50h battery holds; upgrade from Ultra 1, skip from Ultra 3.
- 2026-09-11 · [How I Film Cinematic Cycling Videos With the DJI Osmo Action 6 (Settings + Workflow)](https://watching.ai-assisted.engineering/watched/2026-09-11-cinematic-cycling-dji-osmo-action-6-philip-carneiro/) · Philip Carneiro · Whole video shot on one action cam to prove the point: the 180° shutter rule breaks on action cams (stabilization needs sharp edges, so 1/200 fixed on the bike + ND), full D-Log M settings dump, ND→mouth mount→mic buy order, and a 4-node free / 9-node paid Resolve grade.
- 2026-09-11 · [Apple Watch Ultra 4 & Series 12: BETTER HR, Steps & ECG! (All You Need To Know)](https://watching.ai-assisted.engineering/watched/2026-09-11-apple-watch-ultra-4-series-12-quantified-scientist/) · The Quantified Scientist · Keynote claims analysis: S11 sensor redesign takes background HR from 5min→5s and HRV to every 5min, feeding a dynamic readiness score and Health Age; Rob ordered both to test the "most accurate step counting" claim against reference gear.
- 2026-09-11 · [Apple Watch Ultra 4 Hands-On: 24 New Things Tech Deep-Dive](https://watching.ai-assisted.engineering/watched/2026-09-11-apple-watch-ultra-4-hands-on-dc-rainmaker/) · DC Rainmaker · Ultra 4 companion to the Series 12 hands-on: the LED array regrouped by color so background sensing moved from infrared to green LEDs (the actual mechanism behind 5-second HR), daytime HRV called an industry first, 50h battery with Extended 25h / Max Extended 48h keeping 1s GPS; ships Sept 18 at $799.
- 2026-09-11 · [Apple Watch Series 12 Hands-on: What's Actually New?](https://watching.ai-assisted.engineering/watched/2026-09-11-apple-watch-series-12-hands-on-dc-rainmaker/) · DC Rainmaker · Apple Park hands-on: sensor teardown (+24% ECG area, 8 double-photodiode windows), SDNN and RMSSD both now exposed (answering Rob's question same-day), readiness/Health Age walkthroughs, and a skeptical read on the manufacturer HR study and 24/7 Siri recap privacy; $399, ships Sept 18.
- 2026-08-31 · [25 Cycling Kit Brands ROBBING You Blind (And 7 That Are Actually Worth The Money)](https://watching.ai-assisted.engineering/watched/2026-08-31-cycling-kit-brands-robbing-you-cycleworth/) · CycleWorth · Researched content-farm countdown: one Italian company's $6 pad sits in $174-295 shorts; fake RRPs everywhere; Rapha's £35M losses recovered in jersey prices; worth-it picks: Van Rysel, Assos, Castelli, Galibier, Santini, Endura, Sportful.
- 2026-08-30 · [You need to switch to Linux RIGHT NOW!!](https://watching.ai-assisted.engineering/watched/2026-08-30-networkchuck-switch-to-linux-omarchy/) · NetworkChuck · Omarchy's mainstream breakout: 1:47 Arch install, agent-built themes live during the sponsor read, AI crash diagnoser, Windows-in-Docker; DHH interview clips + typecraft cameo; declares the year of the Linux desktop while honestly keeping Windows for one production app.
- 2026-08-28 · [Herdr Crash Course — Step-by-Step Setup](https://watching.ai-assisted.engineering/watched/2026-08-28-herdr-crash-course-alejandro-ao/) · Alejandro AO · Structured crash course: workspace/tab/pane mental model, prefix-key rationale, --remote VPS setup (UI local, processes remote), and an agent driving Herdr via the official skill; fourth independent Herdr adoption in the archive.
- 2026-08-28 · [This Tmux "Rewrite" Is Actually Brilliant](https://watching.ai-assisted.engineering/watched/2026-08-28-herder-tmux-rewrite-devops-toolbox/) · DevOps Toolbox · Skeptical tmux power user reviews Herdr (Rust agent-native multiplexer by Can Çelik): beautiful defaults, workspace layer, agents panel, remote attach, live handoff; adopts it for his home lab, keeps tmux locally.
- 2026-08-28 · [Garmin CIRQA Ultimate Scientific Review!](https://watching.ai-assisted.engineering/watched/2026-08-28-garmin-cirqa-scientific-review/) · The Quantified Scientist · CIRQA is great on the biceps (HR 98, sleep 73), worse on the wrist (HR 93, sleep 64); recommended with a biceps band.
- 2026-08-28 · [Garmin Cirqa In-Depth Review: Reality vs Expectations](https://watching.ai-assisted.engineering/watched/2026-08-28-garmin-cirqa-dc-rainmaker-review/) · DC Rainmaker · Good $199 no-subscription band with unique auto-GPS detection, but buggy launch, hit-or-miss battery vs the 10-day claim, and missed detections permanently lose training load; he's keeping his Fenix.
- 2026-08-28 · [The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron](https://watching.ai-assisted.engineering/watched/2026-08-28-ed-zitron-diary-of-a-ceo-ai-bubble/) · The Diary Of A CEO · 2.5h adversarial interview: Zitron argues generative AI is a subsidized, circularly financed "con"; predicts OpenAI cash crunch and tech depression ~2027; Bartlett pushes back with adoption and improvement data.
- 2026-08-28 · [how i film my bike rides like a movie — with the DJI Pocket 4P](https://watching.ai-assisted.engineering/watched/2026-08-28-dji-pocket-4p-cycling-filmmaking/) · Gerrit Knein · Hands-free cinematic ride footage recipe (DJI-sponsored): Pocket 4P bar-mounted at eye level, 4K24 @ 1/50 + ND (180° rule), gimbal horizon-lock mode, D-Log 2 → DJI LUT in Resolve; for creators who want quality without a production.
- 2026-08-28 · [I'm ditching tmux for herdr!](https://watching.ai-assisted.engineering/watched/2026-08-28-ditching-tmux-for-herdr-typecraft/) · typecraft · 10-year tmux user makes a full switch: keybinding parity, state-filtered agent fuzzy-finding, agent skill files, and the --remote thin-client workflow; third independent Herdr endorsement in the archive.
- 2026-08-28 · [DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501](https://watching.ai-assisted.engineering/watched/2026-08-28-dhh-lex-fridman-501-agentic-engineering/) · Lex Fridman · 5¼-hour conversation: DHH went from AI skeptic to shipping Omarchy Quattro with 100% agent-written code; ranks Fable #1/Opus 5 #2, Claude Code best harness; argues implementation is no longer the bottleneck and Linux will win the desktop.
- 2026-08-28 · [60 Cool Bike Gadgets on Amazon You’ll Instantly Want to Try](https://watching.ai-assisted.engineering/watched/2026-08-28-60-bike-gadgets-amazon-fix-home/) · Fix Home · Faceless compilation: ~50 products (not 60) in identical ad-copy blurbs, zero testing or prices; real brands (RockBros, Silca, Knog) lend credibility to drop-ship items; a working specimen of slop content.

## Wiki: people and products (51)

- [AeroSpace](https://watching.ai-assisted.engineering/wiki/aerospace.md): An i3-style tiling-window-manager for macOS (github.com/nikitabobko/AeroSpace), configured in `~/.config/aerospace/aerospace.toml`.
- [Alejandro AO](https://watching.ai-assisted.engineering/wiki/alejandro-ao.md): AI educator YouTuber; builds tau-agent at Hugging Face. Structured tutorial format (agenda as a `cat in-this-video.md`). A second-chance-adoption case for herdr: rejected it six weeks earlier ("wasn't adding anything to my workflow"), converted once running multiple agents became Alejandro's daily reality.
- [AlphaGo](https://watching.ai-assisted.engineering/wiki/alphago.md): DeepMind's Go program, the first to beat top professionals: 5-0 over European champion fan-hui (Oct 2015, done in secret) and 4-1 over lee-sedol in Seoul (March 2016), about a decade earlier than experts predicted.
- [Andrej Karpathy](https://watching.ai-assisted.engineering/wiki/andrej-karpathy.md): Former OpenAI founding member and Tesla AI director, now an educator; coined "vibe coding" (vibe-coding).
- [Apple Watch (Series 12 / Ultra 4)](https://watching.ai-assisted.engineering/wiki/apple-watch.md): Apple's 2026 generation — Series 12 ($399) and Ultra 4 ($799) — built on the S11 chip with a redesigned optical + electrical HR sensor (repositioned photodiodes, larger green LEDs, new ECG electrodes).
- [Dave Ebbelaar](https://watching.ai-assisted.engineering/wiki/dave-ebbelaar.md): AI engineer and founder of Datalumina, who teaches on YouTube. They describe themselves as not a terminal wizard and not from a software engineering background: before Herdr they worked in VS Code, Cursor and the Claude Code and Codex desktop apps.
- [DC Rainmaker (Ray)](https://watching.ai-assisted.engineering/wiki/dc-rainmaker.md): Veteran sports-tech reviewer; long-form real-world testing (Ray + wife, hot/cold climates, multiple devices worn simultaneously). Signature conventions: "reality vs expectations" framing, a dedicated bug section signals a rough launch (launch-polish-signal), refuses to grade sleep stages against references that are themselves only ~80% accurate.
- [DHH (David Heinemeier Hansson)](https://watching.ai-assisted.engineering/wiki/dhh.md): Creator of Ruby on Rails, CTO of 37signals (Basecamp, HEY), creator of the omarchy Linux distro, author, race car driver.
- [DJI Osmo Action 6](https://watching.ai-assisted.engineering/wiki/dji-osmo-action-6.md): Action camera; small sensor, front-facing display (which is what makes it usable for self-shot talking heads), RockSteady digital stabilization, D-Log M 10-bit color, threaded ND filters. Sold in an "Adventure Combo" with spare batteries and a selfie stick.
- [DJI Osmo Pocket 4P](https://watching.ai-assisted.engineering/wiki/dji-osmo-pocket-4p.md): Pocket-sized 3-axis gimbal camera (European launch Aug 2026). Two cameras: 20mm wide on a 1-inch sensor (up to 17 stops dynamic range in D-Log 2) + 60mm portrait lens. Internal storage (no SD cards), screw+magnetic ND filters, subject tracking (double-tap), FPV and horizon-lock gimbal modes, battery adapter with hidden 1/4" mount.
- [Ed Zitron](https://watching.ai-assisted.engineering/wiki/ed-zitron.md): Tech-industry PR veteran (~16 years) turned AI skeptic; writes the "Where's Your Ed At" newsletter (on Ghost, not Substack) and hosts the "Better Offline" podcast. Coined the "rot economy" / rot-com-bubble framing.
- [Elastic Interface](https://watching.ai-assisted.engineering/wiki/elastic-interface.md): Italian chamois-pad maker near Treviso with a quiet near-monopoly on premium cycling-short inserts: ~3.2M pads/year on ~€20M revenue (≈$6/pad); the whole company changed hands for $30.5M. Its catalog pads sit inside shorts retailing $174 (Velocio) to $295 (MAAP), and Chinese direct sellers advertise the same pads by name under $100.
- [Fan Hui](https://watching.ai-assisted.engineering/wiki/fan-hui.md): European Go champion (2-dan pro), the first professional beaten by alphago, 5-0 in secret in London in Oct 2015.
- [Garmin CIRQA](https://watching.ai-assisted.engineering/wiki/garmin-cirqa.md): Garmin's $199 screenless smart band (European launch mid-2026): Gen 4 Elevate optical HR sensor with an added skin-temperature layer, physical button (manual workouts, HR broadcast, alarm dismiss), vibration motor, no subscription required. Optional biceps/arm band (~€60, slow to ship). Competes with whoop-strap, Fitbit Air, Amazfit Helio, Polar Loop.
- [Gerrit Knein](https://watching.ai-assisted.engineering/wiki/gerrit-knein.md): German cyclist and YouTube creator; films cinematic cycling content, partnered with DJI (sponsor of the Pocket 4P video). Built a hands-free bar-mounted gimbal setup for ride footage (started on the Pocket 3 in South Africa, refined ~6 months); has worked with cinema crews (Granada/Sierra Nevada shoot).
- [gstack](https://watching.ai-assisted.engineering/wiki/gstack.md): Garry Tan's (Y Combinator president) public Claude Code setup on GitHub: 23 opinionated tools that make an agent act as CEO, designer, engineering manager, release manager, doc engineer and QA. About 135k stars in October 2026.
- [Herdr](https://watching.ai-assisted.engineering/wiki/herdr.md): Agent-native terminal multiplexer (herdr.dev) — "a tmux rewrite in Rust that's a little more user-friendly" — by can-celik, a Turkish developer on it full-time who also contributes Pi extensions. Open source, 11MB binary, curl install, sub-second startup; CLI wraps a JSON API for full scripting.
- [Josean Martinez](https://watching.ai-assisted.engineering/wiki/josean-martinez.md): YouTuber making macOS and terminal setup tutorials. The SketchyBar one is a screen recording of iTerm2, tmux and Neovim with a webcam overlay, plus a companion blog post holding every command. Uses yabai as a window manager.
- [Lee Sedol](https://watching.ai-assisted.engineering/wiki/lee-sedol.md): Korean 9-dan Go professional, 18 world championships, "the Roger Federer of Go". Lee Sedol faced alphago in March 2016 expecting to win 5-0 or 4-1, lost the first three games, then won game 4 with move 78, a "God move" AlphaGo rated at about 1 in 10,000 for a human that exposed its superhuman-blind-spots. That remains humanity's only win against it.
- [Lex Fridman](https://watching.ai-assisted.engineering/wiki/lex-fridman.md): Podcast host (episode #501 with dhh, Aug 2026). Uses voice-first prompting workflows (Plaud recorder + transcription pipeline with custom dictionaries); received a Dell XPS 14 with omarchy preinstalled as an on-air birthday gift.
- [Matt Pocock](https://watching.ai-assisted.engineering/wiki/matt-pocock.md): Teacher (formerly the TypeScript educator — "you see why I stopped teaching TypeScript, cuz AI knows everything now"), now teaching AI-assisted development via AI Hero and conference workshops.
- [mehdio](https://watching.ai-assisted.engineering/wiki/mehdio.md): YouTube channel mehdio debugs, about software engineering. Keeps a public dotfiles repository (github.com/mehd-io/dotfiles).
- [MemGPT](https://watching.ai-assisted.engineering/wiki/memgpt.md): Agent memory architecture (Packer et al., 2024) that treats the model like an operating system and gives it tools to manage its own memory. The context window holds read-only system instructions, a writable working context and a first-in, first-out message queue.
- [Michael Carducci](https://watching.ai-assisted.engineering/wiki/michael-carducci.md): Software architect, architecture consultant, speaker, teacher and magician, who runs a production SaaS with paying customers and SLAs. A regular at No Fluff Just Stuff conferences such as UberConf, where Michael spoke in nearly every slot in 2026.
- [Michael Ryan](https://watching.ai-assisted.engineering/wiki/michael-ryan.md): Stanford CS PhD student, a contributor to the DSPy project, and one of three instructors of CS329Z: Engineering AI Agents with Diyi Yang and John Yang. Ryan taught the course's lecture on agent memory architectures, recorded on 14 October 2026.
- [Mitchell Hashimoto](https://watching.ai-assisted.engineering/wiki/mitchell-hashimoto.md): Founder of HashiCorp and author of the Ghostty terminal. In 2026 Mitchell announced a company for terminal multiplexing, said to be built with hand-written code and no AI (Omer doubts that will hold), and shared demos of smooth panes, fuzzy pickers and curved edges with auto-tiling. omer-hamerman points out that tuios already ships much of that.
- [Mo (Less Bitter)](https://watching.ai-assisted.engineering/wiki/mo-less-bitter.md): YouTuber (channel Less Bitter, @atmoio on X) building enjoy.dev, a tool for building apps with AI. In "The big lie about AI coding" (Oct 2026) Mo argues that coding with AI has no special method: skill files are superstition, agents learn engineering from the labs' post-training, and output tracks hours worked.
- [NetworkChuck](https://watching.ai-assisted.engineering/wiki/networkchuck.md): Mainstream tech/networking YouTuber (coffee-branded, closing-prayer signature); longtime Ubuntu/Kali user, never riced ("I've got a job, two companies, six kids").
- [Not Replaced Yet](https://watching.ai-assisted.engineering/wiki/not-replaced-yet.md): A show by the brothers Nick and Jesse, who run NonFiction, a marketing agency where AI agents do a real share of the work. The episodes cover what they use day to day and what didn't work. Jesse built NFOS, the agency's internal system, and runs its agent threads from Claude Code in PowerShell.
- [Omarchy](https://watching.ai-assisted.engineering/wiki/omarchy.md): Opinionated Arch Linux + Hyprland desktop distribution by dhh ("omakase" = chef's choice). Successor to Omakub (Ubuntu-based).
- [Omer Hamerman (DevOps Toolbox)](https://watching.ai-assisted.engineering/wiki/omer-hamerman.md): Terminal-tooling YouTuber ("omerxx"); tmux power user who maintains several tmux plugins (floating panes/floax, session management) and popularized a catppuccin-style tmux setup. Review style: skeptic-first ("too many projects claiming to be the tmux killers"), allergic to AI-slop projects and 50-agent-swarm LinkedIn posts, converts only on user experience.
- [OpenSpec](https://watching.ai-assisted.engineering/wiki/openspec.md): A command-line tool plus a set of agent skills for spec-driven-development. It installs with Homebrew, puts its skills into the coding harness's directory and supports a couple dozen harnesses.
- [Philip Carneiro](https://watching.ai-assisted.engineering/wiki/philip-carneiro.md): Cyclist and filmmaker (Black Forest / southern Germany), ~10 years shooting photo and video. Films cinematic road-cycling content and teaches the workflow; Philip's signature move is the one-camera proof — shooting an entire tutorial, talking heads included, on the camera being discussed, with the exception shots disclosed on screen (one-camera-proof).
- [Raine Virta](https://watching.ai-assisted.engineering/wiki/raine-virta.md): Software engineer from Finland and author of workmux, the Rust tool that pairs git worktrees with tmux for running coding agents in parallel. omer-hamerman tried the project on a colleague's recommendation and joked about now looking up open source authors online to check they're legit.
- [reemus](https://watching.ai-assisted.engineering/wiki/reemus.md): YouTube channel reemus_dev. In "Why I'm betting on Rex" (Nov 2026), reemus moves from a herdr setup to Rex, arguing that a multiplexer built as a primitive will be more stable than agent workspaces still finding their identity. The video is a narrated motion-graphics essay, and reemus does not appear on camera in the sampled frames.
- [Rex](https://watching.ai-assisted.engineering/wiki/rex.md): Terminal multiplexer from Superlogical, a company co-founded by Ghostty's creator, Mitchell Hashimoto. Rex has been in invite-only beta since October 2026. A server runs your programs in the background, and the Rex app mirrors each one with libghostty, where tmux paints all panes into one screen (compositing-vs-mirroring).
- [Rob ter Horst (The Quantified Scientist)](https://watching.ai-assisted.engineering/wiki/rob-ter-horst.md): Postdoctoral scientist in biological data analysis; YouTube channel applying reference-device-validation to wearables — correlation scores vs a Polar H10 ECG chest strap (exercise HR) and ZMax EEG headband (sleep stages), with percentile comparisons against every device Rob has tested. Scores HR as average correlation over three cardio exercises.
- [Sam Natale](https://watching.ai-assisted.engineering/wiki/sam-natale.md): Small developer-tools YouTuber (~4.5k subscribers), a Neovim and tmux power user whose dotfiles appear on screen.
- [SketchyBar](https://watching.ai-assisted.engineering/wiki/sketchybar.md): Scriptable replacement for the macOS menu bar by Felix Kratz (github.com/FelixKratz/SketchyBar), configured with a shell script, `~/.config/sketchybar/sketchybarrc`. Items have an icon and a label and are updated by plugin scripts (sketchybar-event-scripts).
- [skhd](https://watching.ai-assisted.engineering/wiki/skhd.md): Hotkey daemon for macOS (github.com/koekeishiya/skhd), from the same developer as yabai. Every shortcut lives in one dotfile, `skhdrc`; install with `brew install skhd`.
- [Steven Bartlett](https://watching.ai-assisted.engineering/wiki/steven-bartlett.md): Host of The Diary Of A CEO. In the Zitron episode Bartlett plays informed counterweight: adoption statistics, hallucination-leaderboard improvements, innovator's-dilemma and iPhone/Ballmer analogies, AV safety data, and 95% AI usage at the host's own company — one of the show's more adversarial episodes.
- [Tau agent](https://watching.ai-assisted.engineering/wiki/tau-agent.md): Coding agent by alejandro-ao at Hugging Face: a Python port of pi-agent ("exactly as minimalist... scores pretty much exactly the same in most benchmarks"), with native local-model support via llama.cpp and an extension system whose extensions can add widgets to the TUI sidebar. A Tau-Herder extension (by Ryan Dolphin) integrates it with herdr;
- [Television](https://watching.ai-assisted.engineering/wiki/television.md): Fuzzy finder for the terminal (`tv`), written in Rust. Like fzf it accepts piped input, but it is built around channels: TOML files that each define a source command, a preview, key bindings and actions (picker-channels).
- [Tim Berglund](https://watching.ai-assisted.engineering/wiki/tim-berglund.md): Developer relations leader, conference speaker and YouTuber. Tim gave an OpenSpec talk at UberConf in July 2026 and says the day job is managing people, so Tim's coding happens on hobby projects.
- [Tina Huang](https://watching.ai-assisted.engineering/wiki/tina-huang.md): YouTuber who teaches AI to builders, with quizzes, summary slides and free guides, in a course-like format. Their Hermes agent, which runs much of their life and work, uses local Qwen3 models on their own hardware.
- [TUIOS](https://watching.ai-assisted.engineering/wiki/tuios.md): Open-source "terminal UI operating system" by a developer from the Emirates: a terminal multiplexer that behaves like a small desktop. Panes float and can be dragged with the mouse, minimized to a label, snapped to halves or quarters, or switched into tiling mode.
- [typecraft (Chris)](https://watching.ai-assisted.engineering/wiki/typecraft.md): Linux/Neovim/terminal YouTuber ("hey gang... thanks, nerds"); ~10-year tmux user whose channel identity is built on terminal tooling. Self-described "noob with coding agents" learning in public.
- [workmux](https://watching.ai-assisted.engineering/wiki/workmux.md): A git-worktree manager for parallel (often AI-agent) development that runs on top of your existing terminal multiplexer (tmux, Zellij, Kitty, WezTerm). Its docs call it a "giga opinionated zero-friction workflow tool"; the philosophy is to build on tools you already use and let workmux orchestrate them.
- [yabai](https://watching.ai-assisted.engineering/wiki/yabai.md): tiling-window-manager for macOS (github.com/koekeishiya/yabai), configured in `yabairc` and usually paired with skhd for its key bindings.
- [Yann Dubois](https://watching.ai-assisted.engineering/wiki/yann-dubois.md): Stanford CS PhD student who worked on Alpaca (Llama 7B fine-tuned on 52k LLM-generated instructions), AlpacaFarm and AlpacaEval (GPT-4 as a judge, with a length correction). Dubois's Summer 2024 CS229 guest lecture, "Building Large Language Models", is the archive's most technical walk through the llm-training-pipeline.
- [ZazenCodes](https://watching.ai-assisted.engineering/wiki/zazencodes.md): ZazenCodes is a YouTube channel. Its video "Build a Better Software Factory with Claude Code" (Nov 2026, sponsored by Anthropic) builds the second version of a software factory (software-factory) that installs into a project from one prompt and runs as Claude Code sub-agents, a factory skill and a folder of shell scripts.

## Wiki: concepts (77)

- [180-degree shutter rule](https://watching.ai-assisted.engineering/wiki/180-degree-shutter-rule.md): Shutter speed should be double the frame rate (1/50 at 24fps) to produce motion blur that matches human vision. Needs stabilized footage — on shaky action cams the blur smears, on a gimbal it reads as cinematic. Outdoors this forces ND filters to hold ISO 100.
- [Action-cam shutter exception](https://watching.ai-assisted.engineering/wiki/action-cam-shutter-exception.md): The 180-degree-shutter-rule inverts on action cams, and the archive has both halves of the argument.
- [Agent memory types](https://watching.ai-assisted.engineering/wiki/agent-memory-types.md): The Cognitive Architectures for Language Agents (CoALA) paper sorts agent memory into four kinds. Working memory is the context window the agent reasons over. Episodic memory is the raw record of past behaviour, such as trajectories, tool calls and prior conversations, while semantic memory holds facts about the user and the world.
- [Agent-native multiplexer](https://watching.ai-assisted.engineering/wiki/agent-native-multiplexer.md): Emerging tool category: a terminal multiplexer that treats running AI agents as first-class objects alongside panes: agent status panels, done/waiting notifications with sounds, direct agent attach, and orchestration APIs.
- [Agent-orchestrating multiplexer](https://watching.ai-assisted.engineering/wiki/agent-orchestrating-multiplexer.md): The inversion of the agent-native-multiplexer: an agent drives the multiplexer. Via herdr's official agent-skill-file, it can create workspaces/sessions, start other agents inside them, prompt them, and read their replies.
- [Agent skill file](https://watching.ai-assisted.engineering/wiki/agent-skill-file.md): Tools shipping machine-readable skills that teach AI agents to operate them. herdr's NPX installer asks which agents you use and symlinks per-agent skill files (Claude, Cursor, OpenClaw...) so an agent can open panes, run tests, or manage workspaces itself — demoed by telling Claude "make a new pane in Herdr" and watching it comply.
- [Agentic engineering](https://watching.ai-assisted.engineering/wiki/agentic-engineering.md): AI agents do the implementation; the human steers with high-level vision, problem statements, and taste. dhh hates the term ("marketing slop") and argues it should eventually just be called programming — in English, "the one programming language more beautiful than Ruby."
- [AI removes the maintenance tax](https://watching.ai-assisted.engineering/wiki/ai-removes-the-maintenance-tax.md): The claim that AI agents eliminate the hidden cost that kept normal people off Linux/Arch: configuration drift, breaking updates, arcane errors. Evidence across the archive: omarchy's crash diagnoser correctly triaging a Steam/Proton crash ("not an Omarchy thing, don't file it" — networkchuck);
- [Ambient audio privacy](https://watching.ai-assisted.engineering/wiki/ambient-audio-privacy.md): The trade-off in always-listening wearable features: Apple's Series 12 / Ultra 4 audio intelligence (apple-watch) processes conversations 24/7 in the S11's hardware-isolated secure Exclave — raw audio inaccessible even to Apple, no speaker annotation, summaries deliberately non-quotable (the one exception: a 15-second Live Rewind snippet sent to Siri…
- [Anchor-price theater](https://watching.ai-assisted.engineering/wiki/anchor-price-theater.md): Never-charged list prices as sales scenery: "the crossed-out figure exists to make the number next to it feel like a rescue rather than a purchase." Specimens: Ecohi's simultaneous permanent 60/70% sales; DHB relaunching at "half price" on day one before a single full-price sale ("RRP is a number somebody types in");
- [App-pinned spaces](https://watching.ai-assisted.engineering/wiki/app-pinned-spaces.md): Give each app a home virtual desktop, so a shortcut for the app always lands in the right space. mehdio's case is that developers need more spaces, not more screens.
- [Augment over replace](https://watching.ai-assisted.engineering/wiki/augment-over-replace.md): When a new tool category appears (here, agent-native-multiplexers like herdr), you can often get its features by layering a small, composable tool onto the one you already know, keeping years of keybindings, sessions and scripts, instead of migrating.
- [Binding modes](https://watching.ai-assisted.engineering/wiki/binding-modes.md): Named sets of key bindings, each entered with its own key combination, like layers on a keyboard. The same key does different things in different modes, and nothing outside its mode.
- [Bring your own agent](https://watching.ai-assisted.engineering/wiki/bring-your-own-agent.md): An app ships a CLI, and users drive it with whatever agent they already use. dhh at Rails World 2026: "Built-in agents are a local maxima... Let me bring my own butler to the party." DHH asked every app builder in the room for a CLI "by next Friday".
- [Capacity vs bandwidth](https://watching.ai-assisted.engineering/wiki/capacity-vs-bandwidth.md): Two separate limits on local AI hardware: capacity (RAM) decides whether a model fits at all, while memory bandwidth and the number of processors (GPU, NPU) decide how fast it runs.
- [Circular financing](https://watching.ai-assisted.engineering/wiki/circular-financing.md): Money loops that manufacture demand signals inside the AI economy: Nvidia invests in CoreWeave, which borrows to buy Nvidia GPUs, which Nvidia rents back ($1.3B contract) so CoreWeave can show a customer to lenders; hyperscalers fund OpenAI/Anthropic, whose spending becomes the hyperscalers' reported "AI revenue" (~70% of it).
- [Compositing vs mirroring](https://watching.ai-assisted.engineering/wiki/compositing-vs-mirroring.md): Two ways a terminal multiplexer gets its programs onto the screen. Both keep programs alive in a background server. Compositing, the tmux model, works out what every pane should look like and paints it into one screen of text, borders and status bar included, so any terminal can show it, locally or over SSH.
- [Control surface](https://watching.ai-assisted.engineering/wiki/control-surface.md): A layer of buttons over the agent harness, so moving between agent threads takes one press. It works like a broadcast switcher, where the director calls a camera and someone presses its button.
- [Deterministic spine](https://watching.ai-assisted.engineering/wiki/deterministic-spine.md): Keeping an agent pipeline's state and transitions in plain scripts, so agents change state only by calling those scripts. In the ZazenCodes software factory (software-factory, Nov 2026), each job is a `job.json` file that only `factory/bin/job` may write, through subcommands such as `queue`, `create`, `advance`, `merge` and `feedback`.
- [E-bike conversion kits](https://watching.ai-assisted.engineering/wiki/ebike-conversion-kits.md): Retrofit systems that electrify a regular bike — the one genuine product trend visible in the Fix Home catalog: friction/tire-roller drives (bottle-battery + handlebar controller, ~18 mi range), rear-frame motor adapters, and the notable skarper-style disc-brake drive (replaces the rear rotor, ~37 mi, 15 mph, cadence-responsive).
- [Feedback-loop ceiling](https://watching.ai-assisted.engineering/wiki/feedback-loop-ceiling.md): "The quality of your feedback loops influences how good your AI can code. That is the ceiling." — matt-pocock. Without tests/typechecks the AI codes blind; improving loop quality beats blaming the model.
- [Full-fidelity recording, reduced display](https://watching.ai-assisted.engineering/wiki/full-fidelity-recording-reduced-display.md): A battery-saving pattern worth separating from the usual "low power mode" compromise: keep recording everything at full rate, and save power on the display and radios instead of on the data.
- [Gatekeeping vs accessibility](https://watching.ai-assisted.engineering/wiki/gatekeeping-vs-accessibility.md): What happens to earned-identity communities when the earning gets automated: "I use Arch btw" derived its meaning from setup pain; omarchy delivers the result in seconds, and parts of the Arch community reject its users' claim to the badge. networkchuck's position: earning it the hard way "is kind of a waste of time now";
- [Green-LED background sensing](https://watching.ai-assisted.engineering/wiki/green-led-background-sensing.md): The hardware mechanism behind the Series 12 / Ultra 4's headline jump in sampling rate — the detail that the announcement coverage left as a black box.
- [Grilling session](https://watching.ai-assisted.engineering/wiki/grilling-session.md): matt-pocock's inversion of AI planning: instead of asking the model for a plan (which it eagerly "poofs" out), a tiny skill makes it interview you relentlessly — one question at a time, each with a recommended answer — until human and AI share what Fred Brooks (The Design of Design) calls a design concept. Sessions run 22-100 questions;
- [Health age (biological age)](https://watching.ai-assisted.engineering/wiki/health-age.md): Wearable vendors' longevity metric: compare measured health (VO2max, resting HR, sleep, HRV, optionally lab values like A1C and LDL) against chronological age — "is your body younger or older than you are?" Whoop age came first; Apple's version headlines the new Health app longevity tab (apple-watch).
- [Launch-polish signal](https://watching.ai-assisted.engineering/wiki/launch-polish-signal.md): dc-rainmaker's reviewer meta-heuristic: "if I have a dedicated bug section, that means it was pretty rough" — bugs scattered inline mean a normal launch, a standalone section means the product shipped unfinished.
- [Listicle blurb template](https://watching.ai-assisted.engineering/wiki/listicle-blurb-template.md): The content-farm unit of production: product name → spec recitation → superlative → next, ~25-30 seconds per item, no opinion, no testing, no prices, no host. Infinitely scalable precisely because it requires nothing a human would have to think about; interleaves recognizable brands with drop-ship items so credibility rubs off (slop-content).
- [LLM training pipeline](https://watching.ai-assisted.engineering/wiki/llm-training-pipeline.md): How a chat model like ChatGPT is built, per andrej-karpathy's Deep Dive. There are three stages, mapped onto a textbook: exposition (pre-training), worked examples (supervised fine-tuning), practice problems (reinforcement learning).
- [Log + LUT workflow](https://watching.ai-assisted.engineering/wiki/log-plus-lut-workflow.md): Shoot in the camera's log profile (D-Log 2) to capture maximum dynamic range, convert with the manufacturer's official LUT to Rec.709, then add style (e.g. Resolve's Film Look Creator) and correction on top. Payoff: exposure mistakes become recoverable — a 2+ stop overexposed sunset clip graded back to clean color.
- [Losses recovered in price](https://watching.ai-assisted.engineering/wiki/losses-recovered-in-price.md): "Quality does not set the price of cycling clothing. What sets the price is how much the brand has to recover." Rapha: £35M of losses over two years, eight years in the red, valuation written down £169M→£67M by its Walmart-heir owners — recovered through $350 bib shorts.
- [Lossy compression of the internet](https://watching.ai-assisted.engineering/wiki/lossy-compression-of-the-internet.md): andrej-karpathy's core metaphor for a pre-trained LLM: pre-training chops the internet into tokens and compresses it into roughly a terabyte of parameters, "a lossy, probabilistic zip file." You get the gist back, with the exact wording lost.
- [Malleable computer](https://watching.ai-assisted.engineering/wiki/malleable-computer.md): The thesis that the agentic age needs an OS you can reshape by talking to it — vibe-code your operating system, not just apps. Requires Linux (everything is a config file or CLI tool); locked-down platforms (macOS) become hostile territory. omarchy is the flagship experiment;
- [Manufacturer-study skepticism](https://watching.ai-assisted.engineering/wiki/manufacturer-study-skepticism.md): A vendor-published accuracy study is a starting point for independent testing. dc-rainmaker's formulation: "anytime you have a study sponsored by the company that's releasing the study, yeah, who knows about the accuracy of the study" — while still crediting Apple for publishing one at all and naming its comparison set (Garmin Forerunner 970, Samsung Watch,…
- [Markdown Industrial Complex](https://watching.ai-assisted.engineering/wiki/markdown-industrial-complex.md): mo-less-bitter's name for the trade in shared Markdown skill packs (agent-skill-file) that promise to make a coding agent behave like a senior engineer, a CEO or a QA lead. Mo dates it to Garry Tan's gstack and puts matt-pocock's skills repo in the same group.
- [Memory-bandwidth-bound generation](https://watching.ai-assisted.engineering/wiki/memory-bandwidth-bound-generation.md): Writing each new token means reading the model's active weights out of memory, so on local hardware the generation speed follows memory bandwidth (GB/s) far more than CPU or GPU core counts. Kai cites Apple's machine learning research for this: twice the bandwidth gives about twice the tokens per second, and the Neural Engine or extra cores don't help.
- [Mixture of experts](https://watching.ai-assisted.engineering/wiki/mixture-of-experts.md): A sparse model architecture: the network holds many expert sub-networks and routes each token through a few of them, so only a small share of the parameters is active per token. A dense model reads all of its parameters for every token.
- [Model parameters (weights)](https://watching.ai-assisted.engineering/wiki/model-parameters.md): The parameters are the numbers inside a neural network. Input tokens get mixed with them through a large fixed expression of multiplications, additions and similar operations, and the result is a probability for every possible next token. andrej-karpathy's image: "knobs on a DJ set".
- [Models need tokens to think](https://watching.ai-assisted.engineering/wiki/models-need-tokens-to-think.md): A roughly fixed, small amount of compute happens per forward pass, so a model cannot do arbitrary work inside one token — reasoning must be distributed across many tokens. andrej-karpathy's labeling demo: an answer-first solution ("The answer is $3. Here's why...") is training poison because it teaches single-token guessing;
- [Move 37](https://watching.ai-assisted.engineering/wiki/move-37.md): Game 2, alphago vs lee-sedol (March 2016): a fifth-line shoulder hit that commentators first called a mistake. AlphaGo's own policy network rated it at a 1 in 10,000 chance of a human playing it, so the system knew it was un-human and played it anyway because search and value said it was good.
- [One-camera proof](https://watching.ai-assisted.engineering/wiki/one-camera-proof.md): Make the gear claim by filming the entire argument on the gear. philip-carneiro shoots the whole Action 6 tutorial — riding footage, B-roll, and the talking-head sections — on the camera under discussion, so every frame the viewer is judging is itself the evidence.
- [Open weights](https://watching.ai-assisted.engineering/wiki/open-weights.md): A model is open-weight when its creator publishes the trained parameters, so anyone can download it, host it, and build on it. With a closed model like ChatGPT or Claude you only get the vendor's app or API.
- [Parameters vs context memory](https://watching.ai-assisted.engineering/wiki/params-vs-context-memory.md): andrej-karpathy's two-memory-systems model: knowledge in a model's parameters is "vague recollection" — like something you read a month ago, quality proportional to training-data frequency (lossy-compression-of-the-internet) — while the context window is "working memory," directly accessible to attention.
- [Pencils down](https://watching.ai-assisted.engineering/wiki/pencils-down.md): dhh's name for the 37signals policy announced at Rails World 2026: nobody writes code by hand as a normal part of the job. Hand-written code becomes "an exceptional state... like seeing a bug in Sentry".
- [Picker channels](https://watching.ai-assisted.engineering/wiki/picker-channels.md): A fuzzy picker where each list lives in a small config file holding a source command that produces the entries, a preview command, key bindings, and actions that run on the pick. The file takes the place of a shell pipeline. Television's channels are TOML files in `~/.config/television/cable/`, and an action can replace the picker or run and return to it.
- [Programming as theory building](https://watching.ai-assisted.engineering/wiki/programming-as-theory-building.md): Peter Naur's argument that a program is the team's theory of the problem, and the code captures only the part that runs in the machine. The value is delivered outside the computer, so reading only the code shows only part of the program; in Naur's terms, code is lossy.
- [Quality-gated data omission](https://watching.ai-assisted.engineering/wiki/quality-gated-data-omission.md): The selection-bias mechanism dc-rainmaker uncovered in Apple's heart-rate accuracy study (apple-watch Ultra 4 review): the sensor writes a reading only when an internal confidence bar is met, and writes nothing otherwise. Apple's advertised "every 5 seconds" is an average across a workout — sometimes 1-second cadence, sometimes 20-second silences.
- [Quality without production](https://watching.ai-assisted.engineering/wiki/quality-without-production.md): The sweet spot between action-cam convenience and cinema-camera quality: gear good enough that "I'm not constantly wishing that I brought something else," light enough that filming doesn't turn life into a shoot. The test of an everyday camera is how it handles the mundane moments: "I found myself filming way more since I have it."
- [RAM-to-parameters rule](https://watching.ai-assisted.engineering/wiki/ram-to-parameters-rule.md): tina-huang's rough way to size a local model: subtract 25% of the advertised RAM for the operating system and other programs, then divide what's left by 0.6. The result is the largest model, in billions of parameters, that fits.
- [Readiness score](https://watching.ai-assisted.engineering/wiki/readiness-score.md): The recovery-score convergence: nearly every wearable vendor now compresses activity, training load, vitals, sleep, and HRV into a single daily "how hard should you go" number — Oura and Whoop pioneered it, Garmin has Body Battery/training readiness, and with Series 12 / Ultra 4 Apple joins in (apple-watch): a 0-10 score mapped to four recommendations…
- [Reference-device validation](https://watching.ai-assisted.engineering/wiki/reference-device-validation.md): Judge a wearable by correlating it against a trusted reference, not marketing claims: ECG chest strap (Polar H10) for exercise HR, EEG for sleep.
- [RLHF is not RL](https://watching.ai-assisted.engineering/wiki/rlhf-is-not-rl.md): andrej-karpathy's distinction: in verifiable domains (math, code — a boxed answer to check) RL can run indefinitely and discover superhuman strategies, because the scoring function can't be gamed.
- [Rot-com bubble](https://watching.ai-assisted.engineering/wiki/rot-com-bubble.md): Zitron's thesis: big tech ran out of hypergrowth ideas, so GPU capex became the growth story itself — the market rewards the spending as proof of demand ("they bought GPUs and people went: they're doing AI, we'd better buy the stock"). Unlike dot-com dark fiber, GPUs have no post-bubble utility story and data centers stay expensive to run forever.
- [Scaling laws](https://watching.ai-assisted.engineering/wiki/scaling-laws.md): Plot test loss against training compute, dataset size or parameter count on log-log axes and you get close to a straight line (OpenAI's 2020 scaling-laws paper). Bigger models trained on more data keep getting better, and by a predictable amount.
- [SDNN vs RMSSD](https://watching.ai-assisted.engineering/wiki/sdnn-vs-rmssd.md): The two competing HRV statistics: Apple historically reported SDNN while most of the wearable industry (and most users) use RMSSD, making cross-device HRV numbers incomparable — "all sorts of comparison controversy." A same-day cross-video thread from the Series 12 launch: rob-ter-horst flagged the ambiguity in the announcement analysis and asked Apple…
- [Second-chance adoption](https://watching.ai-assisted.engineering/wiki/second-chance-adoption.md): Tools that land on retry: initial rejection because the workflow need didn't exist yet, later conversion when it does. alejandro-ao on herdr: "didn't think it was for me... then I tried it again and I haven't been able to stop using it" — the missing piece was running multiple agents daily.
- [Sensor placement effect](https://watching.ai-assisted.engineering/wiki/sensor-placement-effect.md): Where a wearable sits can matter more than which device you buy. Both CIRQA reviews confirm it independently: rob-ter-horst measured biceps > wrist across every exercise (weightlifting 0.91 vs 0.73 correlation) and for sleep staging too (possible mechanism: fewer degrees of freedom → less motion artifact → cleaner optical signal);
- [Session per worktree](https://watching.ai-assisted.engineering/wiki/session-per-worktree.md): Give each git worktree (and the agent working in it) its own tmux session instead of a window, so the session manager you already use doubles as the agent switcher.
- [SketchyBar event scripts](https://watching.ai-assisted.engineering/wiki/sketchybar-event-scripts.md): How a sketchybar item stays current: its script runs on a timer, on a subscribed event, or both, and updates only that item.
- [Sleep-time compute](https://watching.ai-assisted.engineering/wiki/sleep-time-compute.md): Using the time an agent sits idle to reason over its past context before anyone asks, so later answers need less test-time reasoning (Lin et al., 2025). A word problem discussed yesterday gets solved overnight and answered instantly the next day; four scattered facts about a user's colour preferences get rewritten into one ranked list.
- [Slop content](https://watching.ai-assisted.engineering/wiki/slop-content.md): Mass-produced media optimized for algorithms and affiliate clicks. ed-zitron traces the lineage: SEO slop (blogs "built to rank rather than be read") preceded AI slop — "we've had slop for years, we've just found a slop machine" — and notes bosses who never cared about output quality are the ones replacing art directors and transcribers with it.
- [Smart zone / dumb zone](https://watching.ai-assisted.engineering/wiki/smart-zone-dumb-zone.md): Dex Horthy's (Human Layer) framing, spread by matt-pocock: LLM quality degrades as context fills because attention relationships scale quadratically ("every token added is a team added to a football league").
- [Software factory](https://watching.ai-assisted.engineering/wiki/software-factory.md): A pipeline of fixed, agent-run stages that carries work items through to merged code, with a human pulled in only when a stage asks. The ZazenCodes version (Nov 2026) turns labelled GitHub issues into pull requests.
- [Spec-driven development](https://watching.ai-assisted.engineering/wiki/spec-driven-development.md): Writing each change as a proposal, design, task list and formal specs before an agent writes the code, and keeping those documents in the repository so later changes are checked against them. The term suffers from semantic diffusion: it covers everything from one giant prompt to opinionated, tool-driven workflows like openspec.
- [Subsidized-demand test](https://watching.ai-assisted.engineering/wiki/subsidized-demand-test.md): Adoption statistics are meaningless while usage is subsidized: a $200/month ChatGPT plan can burn $8-14k of true token cost. The only honest demand signal is willingness to pay per-token prices — and when enterprises were moved to them (March 2026), they revolted (Uber burned its annual token budget in three months).
- [Superhuman blind spots](https://watching.ai-assisted.engineering/wiki/superhuman-blind-spots.md): A system can be superhuman almost everywhere and still hold "tricky lumps of knowledge" where it is confidently wrong. alphago would sometimes judge a dead group alive (or the reverse), and the team couldn't predict when it would wander into one of these lumps or fix them before the match. fan-hui found them in practice games;
- [Thin-client workflow](https://watching.ai-assisted.engineering/wiki/thin-client-workflow.md): Run the multiplexer, agents, and work on a remote box; the laptop is just a window onto it. herdr's `--remote user@ip` makes it one command: SSH in, auto-install if missing, apply your local config and theme. Close the lid and agents keep working — "finally be a responsible father again" (typecraft).
- [Thread rehydration](https://watching.ai-assisted.engineering/wiki/thread-rehydration.md): Clearing an agent thread and restarting it from a saved prompt that fetches the context again, so a fresh session starts from the same place every time. It handles context-window limits without losing the work stream.
- [Tiling window manager](https://watching.ai-assisted.engineering/wiki/tiling-window-manager.md): Windows fill the screen as non-overlapping tiles, and the keyboard moves focus, swaps tiles, changes layout and sends windows to workspaces. Long associated with Linux (i3); aerospace and yabai bring it to macOS.
- [Time to first token](https://watching.ai-assisted.engineering/wiki/time-to-first-token.md): The wait before a model starts answering, mostly prefill: the model reading the whole prompt and context. Tokens-per-second figures in benchmarks measure generation after that wait, so at long context they can hide most of the delay.
- [Tool selection by freshness](https://watching.ai-assisted.engineering/wiki/tool-selection-by-freshness.md): andrej-karpathy's working heuristic for which LLM feature to reach for, based on what needs to get into the context window (params-vs-context-memory):
- [Training-load capture gap](https://watching.ai-assisted.engineering/wiki/training-load-capture-gap.md): Event-gated load tracking (Garmin: load accrues only during detected workouts, sensor power raised only then) loses data permanently on a missed or self-ended detection — dc-rainmaker's mountain ride ended itself mid-descent and the remaining intervals vanished from training load with no way to recover them.
- [Verify like a junior analyst](https://watching.ai-assisted.engineering/wiki/verify-like-a-junior-analyst.md): andrej-karpathy's stance on trusting LLM output: treat it as the work of "a very, very junior data analyst... a little bit absent-minded and not quite right all the time." It's useful and fast, but it's a first draft.
- [Vibe coding](https://watching.ai-assisted.engineering/wiki/vibe-coding.md): Telling an agent to build software without looking at the implementation. Coined by andrej-karpathy. Distinct from programming (which implies understanding loops, conditions, primitives). dhh's experiment: wrote a personal writing app in C++/Qt without reading a line, treating it as a black box.
- [White-label economics](https://watching.ai-assisted.engineering/wiki/white-label-economics.md): When one supplier's catalog part sits inside an entire industry's premium products, the retail price spread measures marketing and investor recovery: elastic-interface's $6 pad inside $174-295 bib shorts;
- [Win probability, not margin](https://watching.ai-assisted.engineering/wiki/win-probability-not-margin.md): alphago maximises its probability of winning and doesn't care whether it wins by one point or fifty. When ahead, it plays "slack moves" that look lazy to humans, giving up territory for certainty. Commentators learned to read a slow move as a tell that AlphaGo thinks it's winning.
- [Worktree per task](https://watching.ai-assisted.engineering/wiki/worktree-per-task.md): Give each branch (feature, code review, parallel AI agent) its own `git worktree` checkout so switching never means stashing, WIP commits, or losing unstaged context. That's why worktrees became the default isolation unit for running several agents at once.
