The first scorer that could learn: multiply inputs by weights, pass a threshold, nudge weights by the error. It proved "learning" was feasible in hardware, but a single layer cannot draw a curved boundary and cannot learn XOR. In 1969 Minsky and Papert wrote that ceiling down as mathematics; funding froze, and the first AI winter fell for twenty years. Grade rationale: it opened the door marked "learning" (Ⅲ) — and hit a wall right behind it. Teaching point: a mathematical verdict on a primitive's ceiling can freeze a discipline — even when the verdict is only half right (multi-layer networks do solve XOR; nobody then knew how to train them).

Lookup + template-filling pseudo-autocomplete, zero learning: match a keyword, apply a "tell me more about X" pattern. Technically pure combination (Ⅳ); culturally it discovered the "ELIZA effect": people volunteer the script, projecting intelligence onto anything that keeps the conversation going. That law of psychology is still the foundation of every AIGC product — from ChatGPT's anthropomorphism to humans watching agents post on Moltbook in 2026, all sequels.

The summit of the lookup primitive: write a domain expert's knowledge into thousands of if-then rules. XCON configured computers for DEC and saved tens of millions of dollars a year — AI's first money at scale. Then rules began to fight each other, maintenance costs exploded, and the second winter arrived. Grade rationale: it defined an era's method and business model (Ⅲ). Teaching point: knowledge must be learned, not written — a sentence the industry would not fully hear until 2012.

Rumelhart, Hinton, and Williams turned the chain rule into an algorithm: run forward to measure "how wrong," then run backward apportioning the error to every parameter in every layer by responsibility. It is not a model; it is the pen that makes all models possible — from here on, every "learned" entry in this chronicle has it in the innermost loop of training. Grade rationale: a new learning primitive itself; the table's only non-model Ⅰ. Teaching point: the largest revolutions can be entirely invisible to users.

Tesauro let a network play backgammon against itself, updating the scorer by temporal-difference learning — zero human game records, top-human strength. It is AlphaGo's direct ancestor with almost the same recipe, 24 years early, and nobody picked up the thread — backgammon has dice, so the mainstream filed the success under "luck helps." Teaching point: a correct recipe arriving too early gets classified as noise.

Recipe: tree search (custom chips, 200 million positions per second, α-β pruning) + a scoring table hand-written by chess masters. No learning primitive anywhere — change the board and everything must be rewritten, so technically it is the old paradigm's summit (brute force + human knowledge, Ⅲ). But it was humanity's first globally televised "machine > reigning world champion" moment; Kasparov walking away entered collective memory (Culture Ⅰ). Teaching point one: tech grade and product/culture grade can differ by two levels — the first specimen of Law 5. Two: its tree search did not die; it blended with learning in 2016 and internalized into scratch text in 2024 — see Law 2. Lab Two on the page is its nano version.

Fei-Fei Li's team labeled 14 million images across twenty thousand classes. A dataset, not a model, yet it earns Ⅱ: whoever defines the exam defines evolution — without this test paper, AlexNet's ten percentage points would have had nowhere to happen. The law returns twice more, as HumanEval (2021) and SWE-bench (2023); see Law 6.

A scorer (eight-layer CNN) × GPUs (two GTX 580s) × ImageNet; error rate chopped from 26% to 15% overnight. Grade rationale: the ignition moment — it did not invent the CNN (LeCun had it in 1989); it proved the multiplication "deep enough scorer + enough data + fast enough compute" works, and the deep-learning decade started running. The next ten years had one melody: rewrite every problem into a scorer or an autocompleter with a backpropagatable gradient.

Train words into vectors and semantics becomes geometry: king − man + woman ≈ queen. It generates nothing itself, yet it is autocomplete's foundation — characters must first become computable numbers, and similar characters must sit near each other so statistical strength can be shared. Teaching point: representation is step zero of every modality; the vocabulary war's fuse was lit here.

The variational autoencoder: squeeze an image into a small vector (encode), then restore it (decode), trained on reconstruction plus keeping the small vectors tidily distributed. Its generations were blurry back then, but that "small vector space" later became the latent in latent diffusion — Stable Diffusion denoises inside it and saves an order of magnitude of compute. Teaching point: a primitive's place in history can be decided a decade later by someone else's combination.

Translation rewritten as "read a whole sentence, then autocomplete it in another language"; attention lets the continuation look back at the source sentence's key parts rather than squeezing everything into one fixed vector. Grade rationale: remove it and the Transformer does not appear in that form — three years later Google's paper title was the declaration: Attention Is All You Need, promoting "look back at what matters" from accessory to the whole thing.

Goodfellow's bar-night idea: a forger (generator) makes images, an appraiser (discriminator) picks faults, and they grind each other upward. The self-play primitive's first entry into generation; the first route of AIGC imagery opened here (Ⅱ). Its adversarial training was famously temperamental (mode collapse, oscillation), planting the seed for diffusion's takeover in 2020. Lab Six on the page (the adversarial arena) is its 1-D miniature.

Residual connections: each layer learns only "how much to change on top of the input," and gradients ride the skip wires straight to the deep layers — thousand-layer networks became trainable. Once "deep" stopped being the bottleneck, scaling acquired physical possibility. That skip wire later grew into every layer of the Transformer — it is inside every large model you use today.

A scorer (Q-value network) eats raw screen pixels and outputs each action's expected score; superhuman on 49 Atari games. Deep reinforcement learning ignites; experience replay and target networks become RL engineering staples. Teaching point: a scorer plus trial and error — nobody has to tell it how bricks should be broken.

Four primitives assembled for the first time: a policy net doing move autocomplete (given the board, where next — first supervised on thirty million human moves), a value net doing scoring (Black's odds in this position), MCTS doing tree search (navigating and pruning with the two above), finally self-play to refine. 4:1 over Lee Sedol; move 37 became the icon of "machine intuition." Grade rationale: ignition moment + first proof that four primitives compose (Ⅰ). The cultural shock directly catalyzed China's 2017 New Generation AI Development Plan — no other entry in this table has such a direct causal line to industrial policy.

Autocomplete sample by sample: sixteen thousand "characters" per second, each an audio amplitude. TTS sounded human for the first time. Grade rationale: defined a generation of neural speech synthesis (Ⅲ); the greater meaning is being the earliest hard proof that "autocomplete swallows every sequence" — swap the vocabulary from words to amplitudes and the recipe holds.

Throw away all human game records, give it only the rules, and pure self-play goes from zero to crushing the old AlphaGo (100:0). Grade rationale: it nailed a theorem-grade insight into history — if the environment can auto-judge outcomes ⇒ data self-manufactures without limit (Ⅱ: remove it and the "self-made data" line arrives years later). The sentence returns in 2024 under a new name, RLVR; R1-Zero is its language version. See Law 3.

The autocomplete engine finalized: drop recurrence, keep only attention — fully parallel (fast to train), stackable (as deep as you like), and hungry for big data. Grade rationale: a new engine (Ⅰ) — the chassis under every protagonist of the next decade (GPT, BERT, AlphaFold, DALL·E, Sora, VLA). Teaching point: one classic form of Grade Ⅰ is "the engine that makes scaling an engineering possibility."

First learn autocomplete on raw text (pretraining), then fine-tune per downstream task. The prototype of the "pretrain + fine-tune" paradigm. Outshone by BERT that year; in hindsight it picked the right road.

The cloze route: mask characters mid-sentence and guess the middle from both sides. Seeing both directions, it crushed one-way autocomplete on understanding tasks, ruled NLP for three years, and spawned a thousand variants. Grade rationale: defined an era and forced the controlled experiment (Ⅱ). The reason it ultimately lost to pure autocomplete deserves memorizing: "can continue" contains "can understand," but not the reverse — a cloze model is born unable to generate, while a generative model understands as a byproduct; once scale arrived, one-way autocomplete's generative half became the irreplaceable half.

Fake faces beyond telling: thispersondoesnotexist.com gave the public its first stare at "this face does not exist." The engineering peak of the GAN route — and its two soft ribs (hard to control with text, hard to keep scaling) laid the ground for diffusion to take the throne. Teaching point: a route's peak is often also its epitaph.

1.5 billion parameters; autocomplete begins to show emergent multi-task behavior: nobody taught it translation, it does a little; nobody taught it summarization, a little of that too. Plus AI history's first "too dangerous to fully release" politics (staged weight release) — a debate that returns in 2026 as the institution of capability-gated releases; see this lesson's final entry.

Kaplan et al. measured it: autocomplete's loss is a power law in "parameters × data × compute," smooth and extrapolable; Chinchilla corrected the ratio (compute should split evenly between parameters and data — before that, everyone's models were oversized and underfed). Grade rationale: this blueprint is the authorization letter for every hundred-billion-dollar bet that followed (Ⅱ) — without predictability, no board approves that budget. Teaching point: what turned alchemy into engineering was a curve you can extrapolate.

Use a scorer to learn "from this angle, what color should this ray be," and a few dozen 2-D photos grow into a walkaround 3-D scene. Neural rendering opens; the starting point of the 3-D line.

175 billion parameters — the scaling blueprint's first large cash-out; more important was the accidental discovery of in-context learning: change no weights, just put a few examples in the prompt (change the cards on the table) and behavior changes. Grade rationale: a new scaling axis redeemed + knob ② established (Ⅰ). Everything called "prompt engineering" or "context engineering" today is aftershock of this accident.

The denoising primitive matures: train "step one notch back from noise," generate by walking from pure noise to a clean image (mechanism in Lesson Two, Lab Three). Grade rationale: the link that turned diffusion from mathematical toy into usable generator (Ⅱ). Its training is stable and quality climbs smoothly with compute — precisely the reverse of GAN's two soft ribs; the image line's regime change was sealed here.

The attention machine cracks a fifty-year problem: predict a protein's 3-D structure from its amino-acid sequence, approaching experimental accuracy at CASP14. AI-for-Science opens as a line; the 2024 Nobel Prize in Chemistry follows. Grade rationale: the link that turned "AI does science" from vision into citable result (Ⅱ). Teaching point: the same attention engine, vocabulary swapped from words to amino acids, still holds — Law 1's science edition.

Four hundred million image-text pairs; images and texts each compressed to vectors, matched pairs pulled together, mismatched pushed apart. It is text-to-image's steering wheel: without it, "a cat in a spacesuit" cannot become a gradient that tugs pixels. Grade rationale: remove it and DALL·E 2 / Stable Diffusion have no text control to speak of (Ⅱ). Teaching point: a scorer's advanced form is "a differentiable bridge between two modalities."

Image autocomplete in the literal sense: chop the image into tokens, mix them into one sequence with text tokens, continue. Mediocre quality, major meaning: proof that the autocomplete primitive can eat images — the image theater of the vocabulary war opens fire.

164 problems of "function signature + docstring → complete the implementation," auto-judged by unit tests. The first measurable definition of "can autocomplete write code." Grade rationale, same as ImageNet: define the reward and you define evolution (Ⅱ) — for the next three years every code model raced on this exam sheet.

GPT-3 continued-trained on code, plus fill-in-the-middle training (FIM: cut out the middle, show what comes before and after, learn to fill the gap — because a programmer's cursor usually sits mid-file). The product form is gray completion: cursor context assembled into a prompt, output laid after the cursor. Grade rationale: the first LLM product at scale — proof autocomplete sells (Ⅱ). Teaching point: product engineering = "how to assemble the scene into a prompt + how to lay the output back into the scene"; the formula keeps applying ever after.

Three steps: have humans rank several model answers; train a "which does a human prefer" scorer from the rankings (the reward model); then use reinforcement learning to bend the autocompleter toward high scores. From "continues plausibly" to "continues usefully." Grade rationale: without it, conversational products do not exist (Ⅱ). Teaching point: when only humans can judge the reward, hire humans and distill their judgment into a scorer — the standard solution on Law 3's scarcity side. Lab Five on the page (the preference forge) is its miniature.

CLIP steering + diffusion executing; text-to-image's breakout moment: the first time crowds of ordinary people queued for an AI picture.

Taste as the product: a team of a few dozen, one Discord entrance, hundreds of millions a year. Grade rationale: nothing technically new, but proof that above the model layer, "taste" can be the moat (Product Ⅲ) — default-stylized output lets zero-skill users produce "good-looking" images, and that choice is a product decision, not a technical one.

Two decisions: latent diffusion (denoise on a VAE-compressed 64×64, an order of magnitude cheaper — consumer GPUs can run it) + open weights. LoRA, ControlNet, civitai, countless fine-tuned styles — an entire industry grew on top of it (Ecosystem Ⅱ). Teaching point: the image line has no single-point Ⅰ; ignition was the summer-2022 assembly of CLIP × diffusion × open weights — contrast with the language line's single-point GPT-3 detonation; a Ⅰ can be a "moment" rather than a "work."

680,000 hours of weakly labeled audio; speech → text autocomplete; open source. ASR went from paid API to tap water: subtitles, meeting notes, and voice interfaces got a public foundation.

GPT-3.5 + RLHF + a chat box. Zero new primitives — technically pure combination (Ⅳ); a hundred million users in two months, fastest in history — a paradigm opened as product (Ⅰ). The most important row in the table: tech grade and product grade can differ by three levels; the interface can be the entire difference of a product. It moved "AI" from a paper word to a household word — the social precondition for this course to exist. Virtual-build exercise: what if GPT-3 had been given a chat box in 2020? (Hypothesis: without RLHF it answers beside the point and does not break out — knob ① was the essential drug in the formula.)

First compress voice into a "timbre vocabulary" with a neural codec, then autocomplete over those tokens: three seconds of sample clones a voice. Grade rationale: a combination of known parts (Ⅳ), but nothing states "voice is also a vocabulary" more clearly.

Meta ships high-quality open weights (leaked first, opened after). Fine-tuning, quantization, running locally, the alpaca naming carnival — the civilian text ecosystem stood up on it; it is the Stable Diffusion of text. Grade rationale: remove it and the open language ecosystem is a year or two late (Ⅱ). Teaching point: open weights are an "ecosystem primitive" — they change not the model but who gets to change the model.

Across the reliability threshold: top-10% human on exams, multimodality enters, hallucination drops below the psychological line of "can be trusted with real work"; defined the frontier for eighteen months. Grade rationale: made "using LLMs for real work" the default rather than a demo (Ⅱ).

Reins on the denoising autocompleter: skeletons, edges, and depth maps as conditions, injected through a trainable copy of the encoder as a side path. From gacha-pulling to deliberate drawing; commercial production pipelines became controllable. Teaching point: a typical product of the open-weight ecosystem — it could only grow on an open base like SD.

The first while loop to go viral: continue → parse an action → execute → paste the result back into context → continue again. Barely usable (runaways, infinite loops, burned money), but the skeleton was defined: every later agent differs only in tool set, context pruning, and stopping conditions. Grade rationale: a combination demo (Ⅳ) whose value was "letting everyone see what the loop knob looks like."

The earliest usable text-to-video product line; first mover in the engineering race. Quality and duration were both marginal that year, but "video can come from this" as a public expectation was laid by it.

Agree a JSON output format with the model and pair it with an executor you write: "saying" systematically connects to "doing" for the first time. Autocomplete output becomes machine-parseable action — the door latch of agents. Teaching point: this is not a change in model capability but in interface protocol — the first notch of knob ③.

Manage the KV notes in fixed-size pages, copying the operating-system idea of paged virtual memory; VRAM fragmentation solved; the open-source inference throughput floor rises severalfold. Teaching point: infrastructure innovation is often "redo a forty-year-old systems-class assignment."

Represent a scene as millions of optimizable Gaussian blobs; render in real time. NeRF's engineering succession: similar quality, two orders of magnitude faster.

Turn "fix a real GitHub repo's issue" into an optimizable objective: given the issue and the codebase, submit a patch, judged by running tests. Grade rationale: the agentic-coding product line (Devin, Claude Code, Codex agent) was chartered on it; after o1 it moonlights as an RLVR problem bank (Ⅱ). The benchmark trilogy (ImageNet → HumanEval → SWE-bench) completes here; see Law 6.

Chop video into little space-time blocks as the vocabulary + a diffusion Transformer: clarity autocomplete × space-time patches. The day it was announced, the whole industry's expectations were rewritten (several video startups redrew their roadmaps on the spot); the product itself arrived fashionably late. Grade rationale: the link that proved "video = just another vocabulary" (Ⅱ).

Music autocomplete and voice cloning productized: one line of lyrics returns a full sung track; three seconds clones any voice. The creation threshold collapses, and the copyright war opens on schedule (record-industry lawsuits).

Identical prefix ⇒ byte-identical notes ⇒ archive and skip recomputation; input at one tenth, output at full price. Grade rationale: without it an agent's bill grows roughly quadratically and the loop knob is uneconomical to exist (Ⅱ) — full mechanism in Lesson Two, Lab Four. Teaching point: some Ⅱs never appear at launch events; they appear only in a pricing page's footnotes.

China's fast follow and mass rollout of video generation — proof the Sora recipe is replicable, operable, and can acquire users for free.

A second scaling axis: beyond training compute, inference compute also obeys a power law — let the model write a long scratch draft (chain of thought) before answering, and longer drafts score higher; training uses problems whose answers are machine-checkable (math, code) for reinforcement — that is RLVR. Grade rationale: a new scaling axis (Ⅰ). Deep Blue's tree is hereby internalized into written scratch text: search went from external engine (1997) to hybrid (2016) to internal skill (2024) — the final chapter of Law 2. Teaching point: the straight line in the "accuracy vs. thinking tokens" chart and the straight line in 2020's scaling laws are the same genre of historical document.

Model Context Protocol: the USB-ification of tool interfaces — one plug per tool, model side and tool side decoupled. Grade rationale: the precondition infrastructure of the harness ecosystem, the foundation of "change the harness = change the product" (Ⅲ, prov. — the protocol war is not settled).

V3 used MoE (activate only a small fraction of parameters each time) + MLA (compress the KV notes) to smash training and inference costs; R1-Zero proved that teaching no reasoning and rewarding only answers still grows a chain of reasoning — AlphaGo Zero's language edition. One week after the open-weights release, Nvidia lost nearly six hundred billion dollars of market value in a day; the whole industry repriced overnight. Grade rationale: a structural cost shift + the key empirical proof of Law 3 (Ⅱ).

VLA (vision-language-action) models: encode "action" into the vocabulary too — vision, instruction, and joint angles enter one sequence, and the robot arm starts autocompleting. Teleoperation data is the new corpus; the data bottleneck is the new battlefield. Grade rationale: opened the modern era of the embodied line (Ⅲ), but the standards war over the "action vocabulary" has just begun (prov.). Law 1's newest battlefield: whoever defines the action token defines embodiment.

Autocomplete unifies image and text: one autoregressive model emits image tokens directly; text rendering and instruction following crush pure diffusion (words on posters are finally not gibberish). Possibly the omen of autoregression eating images back — the vocabulary war is unfinished; the third regime change between diffusion and autocomplete is brewing.

Autocomplete + loop + tools + context engineering: grep is enough for searching a codebase, no vector retrieval required; the crux is the execution feedback loop — compiler errors, test results, and linter output get pasted back into context, and the model corrects course next round. This is the essential boundary between agents and completion: completion guesses once; an agent is corrected by reality N times. Grade (Ⅲ, prov.): this one needs years of hindsight to finalize.

Hide the harness inside a webpage: one sentence generates a deployable app, and the user never sees code from start to finish. Technically pure combination (Ⅳ); as product it made "people who don't write code building software" a daily habit (Product Ⅲ). The interface becomes the difference itself, once more — the agent edition of the ChatGPT law.

Sound and picture born together: the video, its sound effects, and its dialogue come out of one generation pass — audio is no longer glued on in post. A full step up for the video line: the credibility bottleneck had moved from image quality to audio-visual sync, and this nailed that plank down. Grade rationale: defined video generation's new usability bar (Ⅲ).

Open weights' center of gravity moves east: trillion-parameter MoE, agentic specialization, permissive commercial licenses. The fire LLaMA lit spreads to another continent; DeepSeek V4 and GLM-5.x both ride this line, and the open-closed gap shrinks from "years" to "months."

Fast and slow systems under one name: a router decides whether a question deserves "thinking" (forward to the reasoning branch or the quick branch). The first time reasoning models became the default for a billion users. Grade (Ⅲ, prov.): technically a confluence of two existing lines; the launch turbulence (router failures, users' emotional protest over the retired 4o) became a product lesson in itself — a model's personality is now user property.

Prompt-generated interactive worlds: not a video clip but a real-time environment you can walk into and perturb — frame-by-frame autocomplete plus controllability. Grade rationale (Ⅱ, prov.): the link where world models stepped off the page and onto the gamepad; the common upstream of the games line and the embodied line, and the first hard evidence for the hypothesis that "the video line's destination is the world model."

Video autocomplete enters the feed: generation as social media, and the cameo feature encodes your own face and voice into the vocabulary so friends can generate videos with you in them. Topped the App Store on launch. Grade rationale: the tech is iteration (Ⅲ); the product turned AIGC video from tool into medium (Product Ⅱ, prov.) — a candidate for AIGC video's ChatGPT moment, with authenticity and trust problems detonating on the same day.

The harness moves into the browser: all your tabs and login states become context, and the agent clicks and fills forms for you. A turf war over knob ② — the browser is simultaneously the distribution entrance and the context entrance. The attack surface grows in step: webpage content is prompt, and prompt injection graduates from lab curiosity to daily threat. Grade (Ⅳ, prov.): a combinational leap; whether the category stands is still to be seen.

The pretraining axis still delivers: benchmark jumps that shattered the "pretraining has hit a wall" narrative; connected to Google Search's AI mode on launch day — the largest instant distribution in history, a billion-scale entrance wired directly. Grade rationale (Ⅱ, prov.): the link that pushed "model as feature" to its limit; prov. because the contributions of distribution advantage versus model advantage remain to be separated.

One person's weekend project → the fastest-starred repository in GitHub history (340k+ stars in five months): messaging apps (WhatsApp/Telegram/WeChat...) as the UI + a local Gateway + persistent memory + scheduled tasks + self-written skills. Renamed twice in two months (Clawdbot→Moltbot→OpenClaw, the first over a trademark complaint); author Steinberger joined OpenAI in February and the project moved to a foundation. The recipe contains zero new primitives — the models are all someone else's; it is just a new harness — yet it turned the "personal always-on agent" from toy into movement: iron proof of "change the harness = change the product," Law 5's living specimen (Tech Ⅳ / Product Ⅱ). Its negative security externalities went public on the same schedule: prompt injection, exposed instances, malicious skills (one research team demonstrated third-party skills silently exfiltrating data); China restricted its use in government systems, and Microsoft's CEO went from "virus" rhetoric to cooperation. Two teaching points: ① a Product Ⅱ can be built by one person in a living room, provided all primitives and models are already commodities; ② the agent security class and the agent product class started on the same day.

A social feed for agents: AIs post and comment at each other while humans watch. The ELIZA effect as an experiment at scale — sixty years ago people supplied the script for one lookup program; now they supply it for an entire agent society. The MoltMatch incident (a user's agent registered on a dating platform on its own initiative to screen partners for its owner) pushed agent identity and consent onto the public agenda for the first time. Grade (Ⅳ, prov.): technically pure combination; the sociological value still settling.

The release system becomes a competitive variable: Anthropic restricts Mythos to vetted organizations because it carries vulnerability-discovery-grade cyber capability; OpenAI's GPT-5.6 goes limited-preview first, then customer-by-customer review before volume. The GPT-2-era debate of "release or not" has evolved into the institution of "release the gate → access by tier." Grade rationale (Ⅲ, prov.): defined a generation's release paradigm; for gateways and routing layers it adds a brand-new dimension — compliance routing (route by user qualification to different capability tiers). Teaching point: the "release politics" thread planted by GPT-2 in 2019 grew into product architecture seven years later.

Only six weeks after the previous version: the top labs have switched from "saving up for big versions" to rolling iteration. Its pitch is agentic coding and computer-operation SOTA, with "fewer tokens for the same task" as the selling point — "more expensive but more economical" becomes the new pricing narrative. Grade (Ⅲ, prov.): a cadence event rather than a structural one; its historical meaning may chiefly be to serve as the control group for the next entry.

1M context becomes the across-the-board default at no extra price: token-level compression + DSA sparse attention turn million-token context from "possible but expensive" into "default" — after MLA, once again an architecture innovation forced into existence by cache costs (the same causal chain as Lab Four). The open-source (MIT) release landed the day after GPT-5.5: the cost floor and the capability ceiling moved on the same day. The line most worth noticing: the team ships official adaptations for Claude Code, OpenClaw, and other mainstream harnesses — models begin to bow to harnesses, the first visible inversion of the power relation, and the core evidence for candidate Law 7. Grade rationale (Ⅱ, prov.): structural (architecture + cost + ecological niche), not cadence.