Made
Crystal ClearFor teams

The diagrams

Every idea in the books is drawn before it is described. This is the drawings, one a day, each with the paragraphs that explain it. RSS feed.

Figure 3.2: Legacy names mapped to current ones. During the transition window, both old names quietly pointed at deepseek-v4-flash's two modes.

DeepSeek

Legacy names mapped to current ones

2026-10-12

GLM

2026-10-12

Figure 4.1: The meter map, tracing a locally generated token and a hosted token to their real cost sources.

Llama

The meter map, tracing a locally generated token and a hosted token to their real cost...

2026-10-12

Figure 3.1: The watching surface, annotated. The running feed shows each step as it happens; an approval gate interrupts the feed and waits for you before a sensitive action goes out.

Grok Bot

The watching surface, annotated

2026-10-11

Figure 3.1: The lethal trifecta. Left, OpenClaw's default install sits in all three circles at once. Right, pulling the outbound channel out to a small, allowlisted destination breaks the overlap that makes prompt injection dangerous.

OpenClaw

The lethal trifecta

2026-10-11

Figure 5.1: The price ladder from Flash to Max, output rate per million tokens, log scale. Note qwen3.8-max sits lower than qwen3.7-max, the model it replaced.

Qwen

The price ladder from Flash to Max, output rate per million tokens, log scale

2026-10-11

Figure 2.3: Anycast in one picture. One address reaches every datacenter; the request always lands on the nearest healthy Machine, and a nearest-but-unhealthy datacenter is skipped without the client ever noticing.

Fly.io

Anycast in one picture

2026-10-10

Figure 2.3: The decision framework. Your task, context need, budget, and data policy feed a twenty-minute evaluation on your own inputs, which points to a recommended surface and model tier, including the honest case where that tier is not Gemini at all.

Gemini

The decision framework

2026-10-10

Figure 2.3: Deployment lifecycle. A build that fails never reaches ready. A ready deployment may stay a preview or be promoted. Rolling back repoints production at an older ready deployment; nothing here is ever edited in place.

Vercel

Deployment lifecycle

2026-10-10

Figure 2.3: The escalation ladder. Tab, Inline Edit, and Agent rise together in cost and review burden. Ask sits beside the ladder, not on it, because it never touches a file.

Cursor

The escalation ladder

2026-10-09

Figure 2.2: The multiplying effect, using the worked example from Make's own documentation. One trigger check plus 10 bundles through 3 modules comes to 31 credits for a single run.

Make.com

The multiplying effect, using the worked example from Make's own documentation

2026-10-09

Figure 2.2: The post lifecycle. Scheduled posts wait, publishNow skips the wait, drafts wait for you. Publishing ends in one of three terminal states, and partial has its own recovery path: a retry endpoint that re-attempts only the platforms that failed.

Zernio

The post lifecycle

2026-10-09

Figure 2.3: Harness anatomy. System prompts (including your AGENTS.md) and managed context feed the model; the model drives its tools inside a sandbox that limits what is physically possible; approvals sit at the edge deciding when you get asked; checkpoints snapshot the repo around every task. Chapters 4, 10, and 13 each take one of these layers deep.

OpenAI Codex

Harness anatomy

2026-10-08

Figure 2.3: Transformers as the ecosystem pivot. Training frameworks, inference engines, and adjacent runtimes all build on the same model definitions, which is why a model published once runs in so many places.

Hugging Face

Transformers as the ecosystem pivot

2026-10-08

Figure 2.2: Life of a request. The container exists, and bills, only from boot through the last idle second. Stages 2 to 4 together are the cold start; a warm container skips straight to stage 5.

Modal

Life of a request

2026-10-08

Figure 3.1: The GitHub flow as a cycle. Steps 1 and 2 are Git on your machine. Steps 3 to 6 are GitHub. The only two commands that cross the line are push and pull.

GitHub

The GitHub flow as a cycle

2026-10-07

Figure 3.1: Preview vs published. Two copies of the same app; the Publish button is the only bridge between them.

Lovable

Preview vs published

2026-10-07

Figure 3.1: the same three objects, three ways in. Start at the left, move right as the work starts repeating.

Vapi

the same three objects, three ways in

2026-10-07

Figure 3.1: The decision tree. Question 1 comes from Anthropic's rule that agents are for problems whose number of steps cannot be predicted. Questions 2 and 3 are what makes autonomy survivable once you get there.

AI Agents

The decision tree

2026-10-06

Figure 3.2: Tag anatomy. Every image reference is the same four-part address; anything you leave out is filled by a default (docker.io, library, latest). The expansions are the docs' own examples, including a reference on a non-Hub registry where nothing defaults.

Docker

Tag anatomy

2026-10-06

Figure 3.2: Test run versus production run: what fires, what counts toward the quota, and what you get to see.

n8n

Test run versus production run: what fires, what counts toward the quota, and what you...

2026-10-06

Figure 3.1: The quickstart path. Key, .env, and SDK install happen once per project; from then on, every generation is one convert() call carrying the four-part order form, and MP3 audio comes back.

ElevenLabs

The quickstart path

2026-10-05

Figure 5.1: The ramp made visual. Day N is N reps of each of the five moves, so the day's total is five times the day number. Times are mine, at my pace, rests included; yours will be different. The day axis is not to scale, so the bars jump: the real growth is a straight line of one rep per move per day. The dotted extension past day 180 is arithmetic, not a plan.

The +1 Method

The ramp made visual

2026-10-05

Figure 3.2: The true-cost iceberg. A $120K salary is a roughly $170K-a-year commitment before overhead, plus months of ramp and a management load that never expires.

Tech Hiring

The true-cost iceberg

2026-10-05

Figure 3.2: Training and inference. The factory runs once: mountains of text tune billions of dials, then the settings freeze. The storefront runs on every request: your prompt goes through the frozen model's prediction loop and the answer is generated live, metered in tokens.

AI

Training and inference

2026-10-04

Figure 3.1: The zero to first change timeline. Six steps from a bare terminal to a reviewed commit in git log. Steps 1 to 3 are setup you do once; steps 4 to 6 are the loop you will repeat forever, and step 5 is where you, the human, gate everything.

Claude Code

The zero to first change timeline

2026-10-04

Figure 4.1: The autonomy dial. Left to right, what gets auto-approved grows and what still prompts shrinks. YOLO keeps three prompt classes alive: sensitive actions, questions, and plan exits. Only /auto silences all of them.

Kimi Code

The autonomy dial

2026-10-04

Figure 4.1: init starts an empty repo in the current folder; clone copies a remote repo with its full history.

Git

init starts an empty repo in the current folder; clone copies a remote repo with its...

2026-10-03

Figure 6.1. The skill lifecycle: observe repeated success, draft a SKILL.md, reuse it, and let the self-improvement review refine it.

Hermes Agent

The skill lifecycle: observe repeated success, draft a SKILL.md, reuse it, and let the...

2026-10-03

Figure 2.1: The big schema. (1) Your app builds one state and a map of questions keyed by IDs your code chooses. (2) One request carries both, under the 64k and 32k context budgets, with a usage meter that bills input tokens only. (3) Jev reads the state once and judges each question in its own walled-off lane, in parallel. (4) One response returns a typed answer per ID. (5) Your code maps each answer onto a switch, a threshold, or an if, and acts, confirms, or escalates. Answer values are the docs' illustrative ones from the System One page; price as of 2026-09-21. Every later architecture figure zooms into one region of this one.

Jev

The big schema

2026-10-03

Figure 2.3: Two ways to build the denoiser. Left, the U-Net lineage described in Weng (2024); right, the diffusion transformer over spacetime patches introduced by Peebles and Xie (2022) and used by Sora (2024), Veo 3 and Wan (2025). Both take a noisy latent plus the encoded prompt and return a slightly cleaner latent.

AI Video

Two ways to build the denoiser

2026-10-02

Figure 3.1: Quickstart sequence, from account to a verified first response in the dashboard.

OpenRouter

Quickstart sequence, from account to a verified first response in the dashboard

2026-10-02

Figure 2.2: One request, end to end. Lane A is the browser with a publishable key and a user JWT: key check, role switch, grants, RLS, one SQL statement. Lane B is your server with the secret key: same path, but the role becomes service_role and the RLS gate is skipped.

Supabase

One request, end to end

2026-10-02

Figure 3.1: The request and response anatomy of a DeepSeek API call. The usage block is where Chapters 4 and 11 pick the story back up.

DeepSeek

The request and response anatomy of a DeepSeek API call

2026-10-01

GLM

2026-10-01

Figure 2.2: Where Azure fits, the same weights, a fourth path, a different data boundary.

Mistral

Where Azure fits, the same weights, a fourth path, a different data boundary

2026-10-01

Figure 2.2: Three shapes of AI help, and what each one can reach. Persistence, not intelligence, is the axis that separates Grok Bot from the other two.

Grok Bot

Three shapes of AI help, and what each one can reach

2026-09-30

Figure 3.2: Two definitions of open source, and the license Llama actually ships under.

Llama

Two definitions of open source, and the license Llama actually ships under

2026-09-30

Figure 3.2: A decision flow for which licence applies. The flagship and Flash-Next branch always ends in "read the actual text," never in a confirmed permission this book has not verified.

Qwen

A decision flow for which licence applies

2026-09-30

Figure 2.2: The Machine lifecycle. Created and Destroyed cost nothing in compute. Started is billed in full. Stopped and Suspended cost less, with the exact numbers in chapter 4.

Fly.io

The Machine lifecycle

2026-09-29

Figure 2.2: What runs as whom. One process holds both the reasoning and the tools it can call. The line between them is where sandboxing, when enabled, actually cuts.

OpenClaw

What runs as whom

2026-09-29

Figure 2.2: Immutability and the alias pointer. Three deployments, each permanent and independently reachable. Only the production domain moves. Rolling back never deletes or rebuilds anything, it just repoints the arrow.

Vercel

Immutability and the alias pointer

2026-09-29

Figure 2.2: The four modes, by scope of change and degree of autonomy. Later chapters each zoom into one quadrant.

Cursor

The four modes, by scope of change and degree of autonomy

2026-09-28

Figure 2.2: Where Gemini sits, dated. On Artificial Analysis's Intelligence Index (fetched September 3, 2026), Gemini 3.8 Flash scores lowest of the three flagships shown, and is also the cheapest and fastest by a wide margin. On Arena's Text Arena (snapshot September 2, 2026), Google's best settled score ranks 15th, with two newer, higher-ranking entries still preliminary.

Gemini

Where Gemini sits, dated

2026-09-28

Figure 2.1: The big schema. Your product and your AI agent call one API with one key. Inside your team, profiles hold connected accounts, a queue feeds posts, and every post fans out to its platform targets. Webhooks and the unified inbox carry results, messages, and account events back to your server. Six later figures zoom into regions of this diagram.

Zernio

The big schema

2026-09-28

Figure 2.2: The surface decision map. The branch labels are the community's own questions. The left two lanes are pairing (you watch and steer), the right two are delegation (the agent works, you review the result). Come back to this map after chapter 7, when the right half will be muscle memory.

OpenAI Codex

The surface decision map

2026-09-27

Figure 2.2: Inside one model repo. Four kinds of file, a Git rail giving them history and conversation, and Xet chunk storage underneath making multi-gigabyte files practical.

Hugging Face

Inside one model repo

2026-09-27

Figure 2.1: The big schema. Every anatomical part of a scenario on one canvas: the trigger with its clock, modules, bundles on the lines, a filter, a router with ordered routes, the dotted error-handler route, and connections. Nine later figures in this book zoom into pieces of this drawing.

Make.com

The big schema

2026-09-27

Figure 2.2: Plan mode vs Build mode. Flat 1 credit to think, a floating bill to build, the docs' own examples drawn to scale, and the 20-credit check-in gate on long runs.

Lovable

Plan mode vs Build mode

2026-09-26

Figure 2.1: The big schema. Inputs enter through three doors, land on a Function inside an App, and run in containers booted from the Function's Image. Volumes and Secrets attach from the side, the autoscaler sizes the pool, and everything runs on Modal's pooled multi-cloud hardware. Later chapters zoom into each part of this diagram.

Modal

The big schema

2026-09-26

Figure 2.2: Assistant against squad, with the condition each one suits, and the retirement that removed the third option. Anything you read that opens a node-and-edge workflow builder is describing a runtime that stopped on 19 August 2026.

Vapi

Assistant against squad, with the condition each one suits, and the retirement that...

2026-09-26

Figure 2.3: The five workflow patterns from Anthropic's "Building effective agents", at one scale. Solid arrows are paths a developer wrote. The dashed arrows in orchestrator-workers mark the one place a workflow lets the model decide the shape of the work, which is why that pattern sits closest to the agent boundary.

AI Agents

The five workflow patterns from Anthropic's "Building effective agents", at one scale

2026-09-25

Figure 2.2: Pick the credential by the job, not by habit. The dashed box at the bottom is the option that no longer exists, and the side panel is why every GitHub credential is recognizable on sight.

GitHub

Pick the credential by the job, not by habit

2026-09-25

Figure 3.1: The first workflow annotated. Each node is labelled with its role in the items pipeline from chapter 2.

n8n

The first workflow annotated

2026-09-25

Figure 3.1: The docs' five-layer Python image, redrawn. The image layers (locks) are frozen diffs, stacked by a union filesystem into one view; chroot points the container's root at it. Each container adds exactly one writable layer of its own on top, so Container A and Container B share every frozen layer and still cannot see each other's changes.

Docker

The docs' five-layer Python image, redrawn

2026-09-24

Figure 2.3: The model picker. Start from the job, land on a model ID, then check the leaf's latency and character limit against your needs. When in doubt, draft on Flash v2.5 and render the final on a quality model.

ElevenLabs

The model picker

2026-09-24

Figure 3.1: The decide gate. Three exits before the funnel opens: automate, contract, or hire, with the recurring-pain test deciding between the last two.

Tech Hiring

The decide gate

2026-09-24

Figure 3.1: The prediction loop. The model reads everything so far, ranks every possible next token, picks one likely candidate (the temperature dial shapes that pick), adds it to the text, and starts over. A long answer is this lap run thousands of times.

AI

The prediction loop

2026-09-22

Figure 2.2: The map of the machine. One session sits at the center with its context gauge. Inputs fill the context from the left, actions flow out through the guards to the actuators on the right, and their results loop back in as new context. Subagents orbit with context windows of their own. Later chapters zoom into one region at a time.

Claude Code

The map of the machine

2026-09-22

Figure 3.2. Object model and a small DAG: a commit points to one tree (which points to blobs) and to its parent commit; a branch is a pointer to a commit, and HEAD points to the branch.

Git

Object model and a small DAG: a commit points to one tree (which points to blobs) and...

2026-09-22

Figure 5.1. The two-store memory model: a small curated snapshot injected into every session, beside a lossless history searched on demand.

Hermes Agent

The two-store memory model: a small curated snapshot injected into every session,...

2026-09-22

Figure 3.1: The zero to first change timeline. Six steps from a bare terminal to a reviewed commit in git log. Install, launch, and login happen once per machine; orient, change, and commit are the loop you keep, with the approval panel as your gate at step 5.

Kimi Code

The zero to first change timeline

2026-09-22

Figure 4.1: Day one next to the sessions I never did, at the same scale. The middle bar is what most people mean by "starting small". Day one is the sliver on the right, and it is not shrunk for effect: five repetitions on a chart where one hundred is a third of the page.

The +1 Method

Day one next to the sessions I never did, at the same scale

2026-09-22

Figure 2.2: Forward and reverse diffusion on one frame. The forward chain is a fixed recipe used only in training; the reverse chain is what runs when you press Generate. After Weng (2021, updated 2024).

AI Video

Forward and reverse diffusion on one frame

2026-09-21

Figure 1.3: Jev's launch timeline, from founding in 2024 to this book's capture date. SDK dates are the changelog dates; the funding comes from press reports of the launch. Two breaking SDK releases landed within a week of launch, which is why every version in this book is date-stamped.

Jev

Jev's launch timeline, from founding in 2024 to this book's capture date

2026-09-21

Figure 2.1: The big schema. Clients enter through the Envoy gateway, seven services read and write one dedicated Postgres, Studio manages it through postgres-meta, and servers that want plain SQL use the Supavisor pooler lane. Languages and licenses are from the official architecture page as of 2026-09-18.

Supabase

The big schema

2026-09-21

GLM

2026-09-20

Figure 2.1: The three-path map, with billing unit, privacy default and support level marked on each branch.

Mistral

The three-path map, with billing unit, privacy default and support level marked on...

2026-09-20

Figure 2.2: The big schema. Client code, the API surface, the routing decision, the chosen provider, and the billed response, one path this book zooms into repeatedly.

OpenRouter

The big schema

2026-09-20

Figure 2.2: The same prompt, four paths. What differs is who holds the prompt, what it costs, and which content policy governs the reply, not necessarily which model answered.

DeepSeek

The same prompt, four paths

2026-09-19

Figure 3.1: The decision path from the license's own 700 million MAU clause.

Llama

The decision path from the license's own 700 million MAU clause

2026-09-19

Figure 3.1: The model-to-licence map. Every model named in chapter 2's decoder, coloured by which of the three licences governs it.

Qwen

The model-to-licence map

2026-09-19

Figure 2.1: The full path from local code to a served request. Build and release happen once per deploy. Anycast, the Fly Proxy, and your Machine happen on every request, and this is the diagram the rest of the book keeps zooming into.

Fly.io

The full path from local code to a served request

2026-09-18

Figure 2.1: The big schema. You brief the bot, which works inside a per-user isolation boundary on its persistent computer, reaches out through connections to your real apps, and returns through an approval gate for anything sensitive. Later chapters zoom into each labeled piece.

Grok Bot

The big schema

2026-09-18

Figure 2.1: The message path. Everything left of the dashed line is outside your control. Everything right of it, including every tool the agent can call, sits inside one trust boundary.

OpenClaw

The message path

2026-09-18

Figure 2.1: The agent loop. Read, propose, wait for you, edit, run, observe, and repeat, with your approval sitting inside the loop, not after it.

Cursor

The agent loop

2026-09-17

Figure 2.1: What a benchmark measures versus what you need. One engineer's own structured evaluation of Gemini Pro found 95% accuracy in documentation generation and a large nominal context window, and also found the model struggled to navigate a large, evolving codebase without dedicated IDE tooling. Both findings are true about the same model at the same time.

Gemini

What a benchmark measures versus what you need

2026-09-17

Figure 2.1: The big schema. A push builds one immutable deployment, which fans out into the edge network, functions, and the ISR cache. Domains sit on top as pointers. Five later chapters zoom into one box of this picture at a time.

Vercel

The big schema

2026-09-17

Figure 2.1: The big schema. One agent core (a model wrapped in a harness) is met on four surfaces: the CLI on your machine, the IDE extension in your editor, Codex cloud in OpenAI's containers, and code review on your GitHub pull requests. The desktop and mobile apps are windows onto the same agent, not a fifth one. The chapter pins mark where this book zooms in.

OpenAI Codex

The big schema

2026-09-16

Figure 1.3: From Integromat to Make. A bootstrapped Czech startup, a $100 million-plus acquisition, a rebrand, and the 2026 move into AI agents.

Make.com

From Integromat to Make

2026-09-16

Figure 1.3: The short history. Launch dates for the core, WhatsApp, and Ads APIs are as listed on Product Hunt; the rebrand is confirmed by zernio.com/rebrand, which publishes no date. Captured 2026-08-29.

Zernio

The short history

2026-09-16

Figure 2.1: The master schema of Hugging Face. The Hub holds Models, Datasets, and Spaces as Git repos on Xet storage; the library ring works with it from Python; the serving side and the local rail are the two ways to run what you find. Counts are dated mid 2026 and keep growing.

Hugging Face

The master schema of Hugging Face

2026-09-15

Figure 2.1: One prompt's journey. Your words enter the agent loop, the loop edits real files, and the preview refreshes because the code changed. This is the zoomed-in center of Figure 1.1.

Lovable

One prompt's journey

2026-09-15

Figure 1.3: Two feedback loops. The traditional cycle has five stops between an edit and evidence; Modal's has one command. The build still happens, but in the cloud, cached, and off your critical path.

Modal

Two feedback loops

2026-09-15

Figure 2.2: The same job, two architectures. On the left every arrow is something a developer typed. On the right the arrows are dashed because they do not exist until the run happens: the model picks a tool, sees the result, and picks again. That single difference sets the cost, the latency, the failure modes and how you debug it.

AI Agents

The same job, two architectures

2026-09-14

Figure 2.1: Three doors, one repository. Pick per task. Most working days use door two for the code and door three or one for everything GitHub adds on top.

GitHub

Three doors, one repository

2026-09-14

Figure 2.1: Anatomy of one turn. Five stages run in sequence between the caller falling silent and the assistant speaking. Two of them, transport and endpointing, are excluded from the latency number Vapi displays, and Vapi says so.

Vapi

Anatomy of one turn

2026-09-14

Figure 2.2: A container, opened up. One ordinary process sits at the center. The eight namespaces (Cgroup, IPC, Network, Mount, PID, Time, User, UTS, exactly as the Linux manual lists them) control what it can see; cgroups control how much it may use; and everything runs on the host's single kernel. Docker's flags like --memory and --pids-limit are a friendly interface to the cgroup limits.

Docker

A container, opened up

2026-09-13

Figure 2.2: One node, three items in, three items out. The node is configured once. It performs that configured operation individually for each item in the incoming array, which is why a per-item job needs no loop node at all.

n8n

One node, three items in, three items out

2026-09-13

Figure 2.2: The funnel debug dashboard: one metric tap per stage. The framework is the author's; the MOFU benchmark values are from Ashby's 2026 Talent Trends report, measured across 54 million applications. The rejection NPS tap and the silence leak get their full story in Chapter 11; the pay opacity leak is treated in Chapter 10.

Tech Hiring

The funnel debug dashboard: one metric tap per stage

2026-09-13

Figure 2.2: The big schema, this book's anchor map. The voices layer supplies an identity, the model core renders audio (or text, or music), the four doors are how people and programs reach the engines, and everything flows out through web, app, or phone. One credit pool meters it all. Later chapters zoom into one region at a time.

ElevenLabs

The big schema, this book's anchor map

2026-09-12

Figure 2.2: The three quota timers, nested. A request must clear the rolling 5-hour window, the weekly Kimi Code quota, and the monthly membership pool; any one of them can stop you. The prepaid Extra Usage wallet sits outside the timers and is deducted last.

Kimi Code

The three quota timers, nested

2026-09-12

Figure 3.1: The resistance threshold. The bigger today's ask, the lower the chance you actually start, and the whole curve drops on a bad day, so the point where you stop going arrives sooner. Most plans are written on the right of this chart. Day one of the +1 Method sits on the far left, under the threshold on every kind of day.

The +1 Method

The resistance threshold

2026-09-12

Figure 2.2: The model is raw text-generating capability. The product wraps it in an interface, guardrails, memory features, and a pricing plan. When you evaluate a tool, you are evaluating the whole car, not just the engine.

AI

The model is raw text-generating capability

2026-09-11

Figure 2.1: The context gauge across one session. Fill climbs with every file read and command result while response quality drifts down. Near the limit, compaction summarizes the conversation (a lossy drop); /clear resets the gauge entirely.

Claude Code

The context gauge across one session

2026-09-11

Figure 4.1: Install to first session, left to right. Steps 1 to 5 take you from a bare machine to a running chat; the startup gate rejects any model under 64,000 tokens of context, and hermes doctor catches problems at any step.

Hermes Agent

Install to first session, left to right

2026-09-11

Figure 2.1: The generation pipeline, the anchor schema for this book. Drawn from the Sora technical report (Brooks, Peebles, et al., 2024), the Veo 3 technical report and the Wan paper (2025); the six stages are common to all three.

AI Video

The generation pipeline, the anchor schema for this book

2026-09-10

Figure 3.1. The three trees: work moves forward with add and commit, and back with restore and switch.

Git

The three trees: work moves forward with add and commit, and back with restore and switch

2026-09-10

Figure 1.2: System 1 and System 2, with Kahneman's own examples in the top row and the software analogy below: Jev takes the fast, repeated judgments, and your code escalates low-confidence cases to the slow lane. The bottom row is an analogy, not an architecture.

Jev

System 1 and System 2, with Kahneman's own examples in the top row and the software...

2026-09-10

Figure 1.2: A short timeline of the company itself, founding to the Series C, with the model-lineup and rebrand milestones marked.

Mistral

A short timeline of the company itself, founding to the Series C, with the...

2026-09-09

Figure 2.1: One model slug, many provider endpoints. Each endpoint is a separate provider's copy of the same model, at its own price, precision, and data policy.

OpenRouter

One model slug, many provider endpoints

2026-09-09

Figure 1.2: The Supabase timeline from founding to the June 2026 Series F, with the developer-count milestones and the "over 60% launched by an AI tool" marker. Every figure is as reported by TechCrunch on the dates shown; the star count is from the GitHub README on 2026-09-18.

Supabase

The Supabase timeline from founding to the June 2026 Series F, with the...

2026-09-09

Figure 2.1: The four faces. Faces 1 and 2 share DeepSeek's own infrastructure and privacy policy. Faces 3 and 4 branch off the same open weights onto infrastructure DeepSeek does not operate.

DeepSeek

The four faces

2026-09-08

GLM

2026-09-08

Figure 1.2: Dense versus mixture-of-experts, using Scout's own active-versus-total parameter split.

Llama

Dense versus mixture-of-experts, using Scout's own active-versus-total parameter split

2026-09-08

Grok Bot

What it is not

2026-09-07

Figure 1.2: Growth since creation. The curve shape is illustrative; the only verified data point is 388,743 stars at capture, 2026-09-03, via the GitHub API. The count changes daily.

OpenClaw

Growth since creation

2026-09-07

Figure 2.1: The naming decoder, applied to a real Qwen open-weight checkpoint ID. Read any name left to right using these five slots.

Qwen

The naming decoder, applied to a real Qwen open-weight checkpoint ID

2026-09-07

Figure 1.2: What Fly.io runs for you, and what stays yours to decide. Chapters 6 to 9 are, in effect, a guided tour of the right column.

Fly.io

What Fly.io runs for you, and what stays yours to decide

2026-09-06

Figure 1.2: The Gemini family map. Three zones, left to right, by how directly you control what happens with the model. Every box carries its official name, its URL, who it serves, and its billing relationship. The small arrow from AI Studio to the Developer API shows the one you actually need to remember: AI Studio is where a free API key comes from.

Gemini

The Gemini family map

2026-09-06

Figure 1.2: What happens to a request. A static asset and an edge cache hit never reach your code. Only when neither is available does a function run, and only when the function needs data does it reach your own origin.

Vercel

What happens to a request

2026-09-06

Figure 1.2: Adoption versus trust. Usage keeps climbing and confidence in the output keeps falling, in the same survey, the same year.

Cursor

Adoption versus trust

2026-09-05

Figure 1.2: The tool decision in one drawing. Process complexity picks the platform; the billing unit stamped under each tool decides what the same job costs at volume.

Make.com

The tool decision in one drawing

2026-09-05

Figure 1.2: The field map. Dashboards serve people; publishing infrastructure serves code and agents. Zernio competes in the right-hand panel. Vendor positions summarized from their own published comparisons, captured 2026-08-29.

Zernio

The field map

2026-09-05

Figure 1.2: The Codex timeline, April 2025 to August 2026, as reported by the cited history and announcement pages. Model milestones alternate with product milestones; the green endpoint marks the current GPT-5.6 model family as of this book's first stamp date, 2026-08-28; GPT-6 Astra joined the lineup on September 3, 2026 (chapter 6).

OpenAI Codex

The Codex timeline, April 2025 to August 2026, as reported by the cited history and...

2026-09-04

Figure 1.2: Ten years in six stops. Every inflection point is a choice to be everyone's infrastructure instead of one company's product. The last stop is labeled the way this book treats it: reported, not confirmed.

Hugging Face

Ten years in six stops

2026-09-04

Figure 1.2: One day of bursty GPU work under the two billing models. On the rented machine, idle time is the product you are buying. On a per-second platform, idle time does not appear on the bill, because nothing of yours is running.

Modal

One day of bursty GPU work under the two billing models

2026-09-04

Figure 1.2: Where is my file right now. Four places, four commands. Every later diagram in this book is a zoom into one part of this one.

GitHub

Where is my file right now

2026-09-03

Figure 1.2: The promise and the practice. The left panel is what one prompt genuinely delivers on day one. The right panel is what this book adds before other people depend on the result. Neither panel cancels the other.

Lovable

The promise and the practice

2026-09-03

Figure 1.2: The three-hop pipeline. Audio becomes text, text becomes a reply, the reply becomes audio, and the loop closes back on the caller. Every architecture diagram in the rest of this book is a zoom into one box on this track.

Vapi

The three-hop pipeline

2026-09-03

Figure 2.1: The agent loop, the anchor schema of this book. A goal enters. Context is assembled, the model decides, your code runs a tool, the observation goes back into context, and the stop condition is checked before the next turn begins. There are exactly two ways out: the model returns an answer with no tool call, or a ceiling is hit. Later chapters zoom into one box each.

AI Agents

The agent loop, the anchor schema of this book

2026-09-02

Figure 2.1: The anchor diagram of this book. On the left, a hypervisor virtualizes the hardware and every VM repeats a full guest operating system, kernel included. On the right, one host kernel is shared by every container, and each container box holds only an app and its libraries, because a container is a process, not a machine. Note where the isolation boundary sits on each side: virtual hardware for VMs, the shared kernel for containers.

Docker

The anchor diagram of this book

2026-09-02

Figure 2.1: The items pipeline, the anchor schema of this book. A trigger emits an array of three items. Every connection carries that array, never a bare record. Each node runs its configured operation once per item and outputs new items. The highlighted item 2 keeps its identity the whole way and links back, step by step, to the item that produced it.

n8n

The items pipeline, the anchor schema of this book

2026-09-02

Figure 2.1: The two-stage neural TTS pipeline. Text is cleaned and analyzed, then the acoustic model writes a mel spectrogram and the vocoder converts it into a waveform.

ElevenLabs

The two-stage neural TTS pipeline

2026-09-01

Figure 2.1: The same six months, two ways. Motivation (dashed) spikes far above anything the +1 line asks for, then falls to the floor and stays there through the grey months. The +1 line (thick) never has a hard day and never has a day off, and by day 180 it is doing more than the spike ever did.

The +1 Method

The same six months, two ways

2026-09-01

Figure 2.1: The Full Stack Hiring Funnel (the author's framework). Ten stages read clockwise around the ring, in three zones: TOFU (attraction), MOFU (your process), BOFU (keep and grow). The dashed arrows are what makes it a loop: advocacy and content flow from BOFU back into TOFU, and a respectful rejection returns the candidate to Awareness instead of losing them. Chapters 6, 8, 12, and 13 each zoom into a piece of this drawing, so dog-ear this page.

Tech Hiring

The Full Stack Hiring Funnel (the author's framework)

2026-09-01

Figure 2.1: The family map. AI contains machine learning, which contains deep learning, which contains generative AI, where the LLMs live. The famous names on the right are products wrapped around those models. Later chapters point back to this figure, so dog-ear this page. Confusingly, vendors reuse the same name for the model family and the app: the left box means the models, the right column the apps.

AI

The family map

2026-08-31

Figure 1.2: Every surface is a different doorway into the same engine. Configuration you write once (CLAUDE.md, settings, MCP servers) applies everywhere.

Claude Code

Every surface is a different doorway into the same engine

2026-08-31

Figure 2.1: The map of the machine. One session sits at the center, drawing a model from the engine bay on the left and acting on your repo through the guards on the right. Results loop back into the context, sub-agent satellites work in contexts of their own, and every request spends from the quota gauges below. Later chapters zoom into one region at a time.

Kimi Code

The map of the machine

2026-08-31

Figure 2.2: Centralized vs distributed. A central server hands out thin checkouts and mediates every commit; in Git each clone is a complete repository, so work happens locally and syncing is a separate, explicit step.

Git

Centralized vs distributed

2026-08-30

Figure 3.1: The learning loop. Run a task, observe success, curate memory or write a skill in the self-improvement review, then reuse on the next run. Every session is also logged losslessly to SQLite (FTS5) and can be pulled back in via session_search.

Hermes Agent

The learning loop

2026-08-30

Figure 1.1: The LLM in the hot path versus the decision layer. Left, a decision pulled out of generated text through parsing, retries, and regex fix-ups, with a meter on every hop. Right, one request of state and named questions, one response of typed answers with probabilities, and your code in control; the LLM appears only where a written reply is the actual product.

Jev

The LLM in the hot path versus the decision layer

2026-08-30

Figure 1.1: The AI video stack. Seven stages from idea to publish, the tools readers and vendors name for each stage today, and the chapters that cover them. The stages are stable; the tool names carry a date.

AI Video

The AI video stack

2026-08-29

Figure 1.2: What OpenRouter's job is, and what stays yours. Routing, billing, and fallback are the platform's work. Your prompt design, your provider choices, and your data-policy decisions never stop being your responsibility.

OpenRouter

What OpenRouter's job is, and what stays yours

2026-08-29

Figure 1.1: A backend from parts versus one platform. Left, six vendors with six keys, six bills, and six status pages. Right, the same six capabilities as services around one Postgres, reached through one URL and one publishable key, with the access rule written once as a policy on the table.

Supabase

A backend from parts versus one platform

2026-08-29

Figure 1.1: The terminology fog. The same word lands five different ways, nobody flags the mismatch, and the meeting "agrees" on five different things.

AI

The terminology fog

2026-08-27

Figure 1.1: The demo-to-production gap. The five rows of change are drawn from the Google Developers retrospective on refactoring a monolithic prototype (21 April 2026).

AI Agents

The demo-to-production gap

2026-08-27

Figure 1.1: The chat workflow routes every step through you. The agent workflow moves the read-edit-run-check loop inside the tool, and moves you to the review gate.

Claude Code

The chat workflow routes every step through you

2026-08-27

Figure 1.1: Three lanes, three units of work. Autocomplete checks itself against your keystrokes. Pairing checks at every approval. Delegating checks once, at the diff. Each lane to the right trades supervision during the work for inspection after it.

OpenAI Codex

Three lanes, three units of work

2026-08-27

Figure 1.1: The tool landscape. Cursor sits editor-native, closer to completion-first than the marketing implies, on two axes that move faster than any ranking does.

Cursor

The tool landscape

2026-08-27

Figure 1.1: DeepSeek's model family tree. Older generations stay published even after V4 becomes the flagship pair.

DeepSeek

DeepSeek's model family tree

2026-08-27

Figure 1.1: The problem Docker solves. Left: the app plus an install guide meets three machines and produces three outcomes. Right: one image, packed with everything the app needs, behaves the same on all three.

Docker

The problem Docker solves

2026-08-27

Figure 1.1: Where ElevenLabs sits in your stack. Scripts, app code, and phone calls go in on the left; listeners, users, and callers come out on the right. In between, four product doors open onto the same shared voice models and one credit pool.

ElevenLabs

Where ElevenLabs sits in your stack

2026-08-27

Figure 1.1: Three shapes of compute compared. A container platform hides the instance from you. A serverless function is built to vanish. A Fly Machine is your process, addressable and yours, until you decide otherwise.

Fly.io

Three shapes of compute compared

2026-08-27

Figure 1.1: What is a model, what is a product. The same underlying model can sit behind four very different doors. "Is Gemini good?" is not a question with one answer until you say which door you mean.

Gemini

What is a model, what is a product

2026-08-27

Figure 2.1: Snapshots vs diffs. Other systems store the change per version and replay the chain to rebuild a file; Git stores a full snapshot per commit and lets unchanged files reference the identical previous copy.

Git

Snapshots vs diffs

2026-08-27

Figure 1.1: The split. Everything on the left is git and belongs to you. Everything on the right is GitHub and belongs to a service you sign in to. Push and pull are the only bridge.

GitHub

The split

2026-08-27

GLM

2026-08-27

Figure 1.1: The two events that meet at Grok Bot. The SpaceX and xAI merger runs across the top row, ending in the SpaceXAI rebrand. The Cursor acquisition (June 16, 2026) folds Cursor into the same company. Grok Bot, shown in the accent box, launches at the point where both stories are already true.

Grok Bot

The two events that meet at Grok Bot

2026-08-27

Figure 2.1: The same question ("help me with my server") lands very differently across the three. Only the autonomous agent keeps state between sessions and can act on its own.

Hermes Agent

The same question ("help me with my server") lands very differently across the three

2026-08-27

Figure 1.1: Where Hugging Face sits. Model creators publish above it, runtimes and apps consume below it, and the closed-API lane runs alongside with nothing to download. The middle layer is the subject of this book.

Hugging Face

Where Hugging Face sits

2026-08-27

Figure 1.1: The Kimi Code stack. Open-weight models at the bottom, one membership and dual-protocol API in the middle, and many surfaces on top. You pick the surface; the subscription and models stay the same.

Kimi Code

The Kimi Code stack

2026-08-27

Figure 1.1: The Llama family tree, from dense Llama 3.x through the mixture-of-experts Llama 4 generation.

Llama

The Llama family tree, from dense Llama 3.x through the mixture-of-experts Llama 4...

2026-08-27

Figure 1.1: The loop. You describe, the agent edits a real codebase, you check the preview, and you publish when ready. Credits meter the agent's work, Git sync keeps an off-platform copy, and the backend is generated alongside the app. This is the book's master schema.

Lovable

The loop

2026-08-27

Figure 1.1: The platform map. Your apps ring the Make canvas; a scenario drawn in the middle moves data between them automatically. Data flows both ways: Make reads from these apps and writes back to them.

Make.com

The platform map

2026-08-27

Figure 1.1: One company, three products, drawn against who each is for and what it costs.

Mistral

One company, three products, drawn against who each is for and what it costs

2026-08-27

Figure 1.1: Where Modal sits. You write and call Python at the top. Modal's SDK, scheduler, and runtime turn each call into a scheduled container. The containers land on a pooled fleet across AWS, GCP, and OCI that you never see or manage.

Modal

Where Modal sits

2026-08-27

Figure 1.1: Where n8n sits. Prices and build estimates are the figures published on n8n.io/pricing and by NodeSparks, both read 2026-08-21. Agent frameworks sit off the diagonal on purpose: they raise your effort without handing you more control over what happens.

n8n

Where n8n sits

2026-08-27

Figure 1.1: The four nouns. A message enters from your phone through a channel, the gateway routes it to a model provider and, when needed, a skill, and the result comes back as a reply or an action on your own machine.

OpenClaw

The four nouns

2026-08-27

Figure 1.1: Four years of that membership, one dot per month. On the left, the months I paid: all of them. On the right, the months I set foot inside even once: a handful, and they clustered after the New Year. Reconstructed from memory, so read the shape rather than the exact count.

The +1 Method

Four years of that membership, one dot per month

2026-08-27

Figure 1.1: Qwen's public timeline, April 2023 to August 2026, per Wikipedia's community-sourced overview (captured 2026-09-04).

Qwen

Qwen's public timeline, April 2023 to August 2026, per Wikipedia's community-sourced...

2026-08-27

Figure 1.1: Filtering vs verification. The 2015 funnel chooses from a pile of real CVs. The 2026 funnel faces a partly synthetic pile, so every stage carries a verification gate before taste matters again.

Tech Hiring

Filtering vs verification

2026-08-27

Figure 1.1: The missed-call problem, drawn as money leaving. The top half is where an unanswered call goes. The bottom half is what the same 250 minutes costs on the three routes a business can actually choose between.

Vapi

The missed-call problem, drawn as money leaving

2026-08-27

Figure 1.1: The boundary diagram. Routing, Server Components, and Server Actions are Next.js itself and run on any Node host. Functions, Image Optimization, and ISR storage exist elsewhere too, but you set up their equivalents by hand. Vercel's own edge network is the one box nobody else can offer, because it is a specific company's specific infrastructure.

Vercel

The boundary diagram

2026-08-27

Figure 1.1: Sixteen direct integrations versus one hub. On the left, your app owns sixteen auth flows and sixteen review queues. On the right, it owns one integration, and Zernio owns the rest. Platform counts per docs.zernio.com, captured 2026-08-29.

Zernio

Sixteen direct integrations versus one hub

2026-08-27