Front / Panorama

The AI panorama: one scroll, five zones

63 nodes, from compute at the riverbed all the way to rules outside the city. This is not a knowledge map but a position map: every node is tagged with “from which score on you really need to understand it”. First turn on “only level N-x” and the nodes above you fade out — those are what you take on next. Drag to pan, scroll to zoom, click a node for one line.

Only:
100%← drag to pan, or scroll horizontally/vertically; street, city and wild lie to the right →
River · compute & dataBase supply: the model's grain and horsepowerBridge · models & trainingConnection layer: turning compute into capabilityStreet · applications & productsMarket layer: capability turned into something usableCity · organisation & processCity gate layer: it counts only once it enters the organisationWild · rules & societyOutside the city: the feasible region is drawn hereGPU clustersGPU clusters: the bricks of compute, stacked one by one.Memory bandwidthMemory bandwidth: models run slow because this corridor is narrow.InterconnectInterconnect: whether ten thousand cards act as one depends on this line.Chip processChip process: how much compute fits in the same area, and so the cost curve.Data-centre powerData-centre power: a rack's wattage often becomes the bottleneck before the algorithm does.CoolingCooling: the other half of compute cost hides in air and liquid.Optical modulesOptical modules: the pins that stitch a thousand machines into one big computer.Training corpusTraining corpus: the model's grain; its source decides its taste and bias.Data cleaningData cleaning: dirty data doesn't make a model dumber, it makes it more confidently wrong.Synthetic dataSynthetic data: feeding models with models — cheap, but it amplifies its own flaws.Labeling & human reviewLabeling & human review: the most expensive layer of labour, and the draft of “what counts as right”.Copyright & licensingCopyright & licensing: whether a corpus may be used is a legal riverbed, not a technical question.Vector storeVector store: the warehouse that lets a model “recall”; the ceiling on retrieval quality.Data feedbackData feedback: real production questions returning to the training set — that is a closed loop.TransformerTransformer: the bridge that made attention parallelisable.AttentionAttention: within one sentence, who should look at whom — the mechanism of comprehension.Parameter scaleParameter scale: one dimension of capability, not all of it; past a point the gain slows.TokenizerTokenizer: the model's “characters” are not yours; many oddities start here.Pre-trainingPre-training: one huge upfront investment that decides how wide the base is.Supervised fine-tuningSupervised fine-tuning: teaching it to speak in human format, not to think.RLHFRLHF: making it “likeable”, and teaching it to pander.Preference optimisationPreference optimisation: a cheaper alignment route, at the price of extremes.DistillationDistillation: pressing a large model's craft into a small one — cheap, but details get lost.QuantisationQuantisation: trading precision for speed and VRAM; most deployments rely on it.Inference engineInference engine: on the same card, how fast it runs is half decided here.Context windowContext window: the range it can “see” right now, not its memory.Mixture of expertsMixture of experts: waking only the experts that need to work, saving compute.Multimodal alignmentMultimodal alignment: images, audio and text landing in one semantic space.BenchmarksBenchmarks: the score you trust may just be a question it has practised.Chat assistantChat assistant: the most widespread entry point, and the most misunderstood one.Agent frameworkAgent framework: scaffolding for its hands and feet, not for its brain.Retrieval augmentationRetrieval augmentation: look it up before answering — the first gate against hallucination.Tool callsTool calls: issuing invoices, querying databases, placing orders — that is real work.Workflow orchestrationWorkflow orchestration: stringing steps into a production line that can stop on failure.Coding assistantCoding assistant: the first scenario fully absorbed, because correctness is compilable.Image generationImage generation: the double battlefield of aesthetics and copyright.Voice & realtimeVoice & realtime: conversation starts only below the human patience line of latency.Video generationVideo generation: cost and consistency still fight on this layer.Digital humansDigital humans: the skin matured first, the brain followed later.Browser automationBrowser automation: the old world without APIs gets connected through here.Low-code buildingLow-code building: lets the business build its own tools — and a pile of orphan systems.Private deploymentPrivate deployment: data stays home, capability takes a discount. A hard trade.Cost & billingCost & billing: per call, per volume, per seat — pricing is product design.Human–machine divisionHuman–machine division: which step to hand over and which must stay is a design question.Task ledgerTask ledger: collaboration without a ledger is not collaboration.Acceptance criteriaAcceptance criteria: work you can't write criteria for should not go to an Agent.Permissions & auditPermissions & audit: what it may touch and what it touched must be retained.Knowledge-base governanceKnowledge-base governance: a dirty library makes even a strong model confidently wrong.Shared definitionsShared definitions: same word, different meaning is the most expensive tax in human–machine work.Agent headcountAgent headcount: how many, which roles, reporting to whom — the org question comes before the tech.On-call & patrolOn-call & patrol: who watches at night decides whether the system can sleep.Review & codifyReview & codify: one mistake must buy one rule, or you will make it twice.ROIROI: whether the saved labour can be counted decides whether this continues.Regulation & complianceRegulation & compliance: the feasible region's frame is often drawn outside technology.Copyright ownershipCopyright ownership: who owns the output is still unsettled ground.Privacy & dataPrivacy & data: training, inference and retention — each has its own mines.DeepfakesDeepfakes: the rise in cost of trust is a bill society pays.Employment structureEmployment structure: not “how many jobs vanish” but “which stretch gets replaced”.Compute geopoliticsCompute geopolitics: where the cards come from decides who trains the next generation.Open vs closed sourceOpen vs closed source: openness buys diffusion, closing buys control — both have a price.Standards & interoperabilityStandards & interoperability: capability only really flows when interfaces unify.Energy costEnergy cost: every inference sits on a unit of electricity.Trust & liabilityTrust & liability: who answers when it breaks decides how far this can go.

What each of the five zones governs

14 nodes

River · compute & data

Base supply: the model's grain and horsepower

15 nodes

Bridge · models & training

Connection layer: turning compute into capability

14 nodes

Street · applications & products

Market layer: capability turned into something usable

10 nodes

City · organisation & process

City gate layer: it counts only once it enters the organisation

10 nodes

Wild · rules & society

Outside the city: the feasible region is drawn here

How to read this scroll

① Turn on “only level N-x” first and accept the batch you should know; don't grab everything.
② Then switch the filter off and look at the whole map: you will find that downstream problems are often answered upstream — a street that won't run is usually a bridge (model capability) or a river (data) issue.
③ Finally click three to five nodes and read their “one line” aloud. If you stumble, you only memorised the name.

This map is a scroll rather than a flow chart because real AI is not a pyramid — it is a whole market street with upstream and downstream: the water in the river decides how much the bridge can bear, whether the bridge connects decides how big the street can trade, the city gate decides what may enter, and the law outside decides how far the city can be drawn.

Measure my level →  Six levels

Lupu site group · three stations
① The Agent Era | ② The AI Scale | ③ 10,000 Sites
Next stop → ③ 10,000 Sites
季夏为萤路普AI | 滇ICP备2023002020号-2