I Followed an AI Answer Back to Its Sources. Five Cards on What I Found.
Why the model's knowledge has a freeze date, why retrieval gets it past that date, and what a single answer costs in water and watts.
TL;DR. A four-week series. Two weeks ago: the overview and three preview cards. Last week: Part 1 on the runtime of a request. This week: five cards on where AI’s answers come from. June 30: Why You Can Trust Some Systems and Not Others. Free PDF after Part 3.
When an AI generates a painting in the style of Monet, where does that style come from?
Last month a stunt on X made the question vivid. SHL0MS, a pseudonymous artist, posted a real Claude Monet Water Lilies canvas, marked it “Made with AI” using the platform’s own label, and asked people to describe its inferiority. Thousands did. One person spent 850 words. The painting was real.
The lesson everyone took home: labels warp perception. Sure. The lesson nobody mentioned: every image model on earth has trained on Monet. When AI generates in his style, it reconstitutes him, badly, from weights that are themselves compressed versions of his work. The prank’s premise was unstable in both directions.
This is the entrance to Part 2 of the field guide. Last week we opened the box on the runtime of a request. This week we follow the answers back to their sources: a frozen training corpus, a token-level grammar, a retrieval layer that pulls in fresh facts when the corpus runs out, the physical buildings where all of this actually happens, and the water and watts those buildings drink and burn.
Five cards. In order.
Card 06: An LLM Training Run, End to End
The model you are talking to right now is a snapshot. It learned what it learned during a specific window of time, on a specific pile of data, and then the run ended. The weights froze. They have not changed since.
The window has a name: knowledge cutoff. Most consumer-facing models in 2026 sit somewhere between six months and two years behind the present. Anthropic and OpenAI publish their cutoff dates, and both companies have explained, at length, that the date marks the boundary of the model’s trained knowledge, not the boundary of what it can talk about. A model can answer about events after the cutoff if a tool or retrieval system feeds the facts in at request time. Without that scaffolding, the model improvises, which means it hallucinates.
The card is the eight-step map of how a training run actually proceeds: data collection, freeze, filtering, tokenization, pretraining, post-training, evaluation, release. Two columns on the right distinguish what trains into the weights from what tools and retrieval add at answer time.
There are three items worth a closer look.
The first is the data-collection window. Public-licensed material, internal proprietary corpora, and synthetic data all go in. Monet went in. So did Wikipedia, so did the New York Times, and the resulting lawsuits will define the next several years of training-data law.
The second is the freeze. After the cutoff, no live signal reaches the weights. The model has no concept of this morning’s news the way a colleague does. The model has the world as it stood the day the curators closed the door.
The third is the release decision. A trained model does not ship straight from the run. Alignment passes, safety reviews, red-team adversarial probes, and quality gates all sit between the raw output of the pretraining run and the version anybody can call.
Card 07: Inside an LLM Training Run
Card 06 was the wide shot. Card 07 is the close-up on steps 4 and 5: tokenization and pretraining.
What happens to a Monet painting when it enters a training run is roughly this. The training pipeline slices the image into visual tokens, encodes each token as a vector, treats each vector as a numerical input the network can compare against millions of other vectors, adjusts billions of weights to predict what comes next given the surrounding context, feeds the prediction error back into the next adjustment, and runs that cycle until the weights encode something that looks, to a human eye, like the model has learned Monet’s late style.
It has not learned Monet’s late style. It has learned a distribution of pixel arrangements that correlate with the label Monet. Run that distribution backward (generation rather than prediction), and the network samples from it. The output reads as Monet because the patterns it samples are the same patterns Monet himself produced when he stood in his garden in Giverny and worked a canvas with a brush full of cerulean blue.
The card has three side-by-side examples (text, image, video) that walk through the same machinery for the three modalities. Underneath, a panel labeled Weights, Embeddings, Inference Defined draws the clearest distinction in the cluster. The weights are the long-term memory of the model, learned during the run and frozen at release. The embeddings are how the model encodes any input it receives into a form it can compare. The inference is what the frozen weights do every time someone hits send.
This is the card that finally answers the Monet question. The AI’s “style of Monet” is a sampled approximation of Monet’s actual style, recovered from a compressed representation of his actual paintings. Better than nothing. Not the same thing. Closer to a forgery than to a homage, and judged with neither the labor nor the lighting of either.
Card 08: RAG and Retrieval
If knowledge cutoff is the wall, retrieval is the door.
When the model needs information that lives outside its weights (yesterday’s earnings call, an open customer ticket, a regulation that changed two weeks ago), a retrieval layer fetches the information and injects it into the context window before the model answers. The model does not memorize the retrieved content. It reads it for this request and forgets it the moment the chat closes.
The card walks the seven-step retrieval flow: corpus preparation (chunking, embedding, indexing), live query rewriting, hybrid search across keyword and vector indexes, reranking, permission filtering, citation assembly, and final response generation.
Three properties matter when thinking through RAG.
Retrieval does not teach the model. It gives the model evidence at answer time, the way a witness gives a deposition. The witness leaves the courtroom after their statement. The judge does not memorize what they said for the next trial.
Retrieval quality depends on the corpus, the chunking, the ranking, and the permission filter. Most “the AI made it up” complaints in enterprise AI are retrieval failures rather than model failures. The right document existed, but the retrieval layer did not find it, or found it and ranked it below something worse.
Retrieval makes citation possible. Without retrieval, the model has no anchor to point to when somebody asks where did you get that? With retrieval, every assertion can carry a link back to the source, and in most enterprise systems, every assertion should. Citation is the difference between an AI that claims and an AI that shows its receipts.
Card 09: Where Your LLM Prompt Goes
There is no cloud. There are buildings, in specific places, with specific power draws, specific cooling systems, and specific connections to specific electrical grids.
When you hit send, your prompt leaves your device, traverses an internet backbone, lands at the AI vendor’s edge gateway, routes to a regional data center based on your location and the vendor’s capacity planning, and runs on an inference cluster whose accelerators already hold the model weights in memory. The card maps that journey across eight stages and labels what each stage actually consumes.
Two of those stages deserve their own slide deck and rarely get one.
The first is region selection. AI providers run their inference workloads across dozens of regional clusters, each one in a specific country or US state, each one subject to different data residency rules, grid mixes, and operating costs. The same prompt, sent from the same user, can land in different physical countries depending on the day. This is why where is my data is not a one-line answer.
The second is the model-serving cluster itself. Hundreds of replicas of the model weights run across thousands of accelerators, with live request routing distributing traffic across them. The exact replica count and rack placement are usually not public. What is sometimes public: total power draw, water consumption for cooling, and the carbon intensity of the local grid.
This is where the AI Not a Bubble argument I made last fall cashes in. The physical infrastructure investment is too large to fold back into a marketing line. Google, Microsoft, Meta, and Amazon collectively spent more than 300 billion US dollars on AI data center buildout across 2024 and 2025, with comparable commitments running through 2026. That capital lives in concrete and copper and silicon. It will run something. The bet is on what.
Card 10: A Prompt Footprint, in Human Terms
I previewed this card in the manifesto. Here is the full version.
A single short text query, on a clean inference cluster running an efficiently-sized model, costs roughly a quarter of a milliliter of water, a third of a watt-hour of electricity, and three hundredths of a gram of CO₂. Smaller than a sheet of paper. About nine seconds of television. About one one-thousandth of a load of laundry.
Scale it up. Ten thousand of those queries equal a sheet of paper in water, a smartphone charge in energy, and a short domestic flight’s worth of carbon when the local grid runs dirty. Billions of queries (the actual scale of inference traffic in 2026) move the number from negligible to meaningful.
Heavier inference climbs another order of magnitude per request: a reasoning model that thinks for thirty seconds, a video generation that runs for a minute, an audio synthesis with twenty layers of post-processing. The numbers on the card give the per-query baseline and the operating ladder for the most common workloads. The methodology notes credit OpenAI and Google’s published estimates, plus the EPA’s utility data on power draw for the data center clusters those models run on.
The card is not an argument for or against AI usage. It is the operating cost of a unit of inference, in units that humans understand. Every AI policy and every AI vendor contract should reconcile against numbers in this range. Anyone who tells you the per-query footprint is either negligible or catastrophic without specifying the workload is selling you something.
What this card cluster argues
The answers come from somewhere. Most of the time, they come from a training corpus frozen six months to two years ago, compressed into weights, and sampled at request time. Some of the time, they come from a retrieval layer that fetches fresh information and injects it at answer time. All of the time, they run on physical infrastructure in specific places, drinking specific amounts of water and burning specific amounts of electricity.
This matters because the conversation about AI keeps treating the answer as if it materialized from a black box. It did not. The black box has provenance: a training run, a retrieval pipeline, a data center, a grid. Each layer carries different controls, different costs, and different traceability. Lawyers, regulators, and procurement teams who do not draw those distinctions land in arguments they cannot win.
The Monet was real. The AI Monet would have been a sample from weights trained on the real one. Both are real artifacts in the world, with different provenance, different cost, and different audiences. The mistake is treating either as if it arrived from nothing.
What’s next
Next Tuesday: Why You Can Trust Some Systems and Not Others — Part 3 of The AI Field Guide. Five more cards on agents, identity and permissions, model routing, evaluation, and the closer of the series: the enterprise AI control plane that ties every layer together.
PDF the week after.
Receipts
SHL0MS, X / Twitter (May 12, 2026) — the prank that opened this piece. Coverage at Futurism and PetaPixel.
Scientific Reports — research on AI-label bias in art evaluation: viewers devalue art labeled as AI-generated across multiple dimensions, even when they cannot distinguish it from human-made art.
Anthropic — model knowledge cutoff documentation. Published cutoff dates for each Claude release.
OpenAI — public statements on GPT-4 / GPT-4o / o-series cutoff dates and post-training behavior.
Common Crawl — the public web archive that supplies a meaningful fraction of the public-language training corpus for every major foundation model.
The New York Times v. OpenAI — December 2023 lawsuit defining the contested terrain of training-data use, ongoing through 2026.
OpenAI and Google water and energy disclosures, 2025–2026 — the basis for Card 10’s per-query estimates.
EPA — Data Center Energy Usage Report — utility-grade data on US data center power and water consumption.
Beyond Reason, prior pieces — The manifesto, Part 1: What Happens When You Hit Send, AI Not a Bubble.
The bird is still moving. By now you know what it eats, where it nests, and what its presence costs the local water table. See you next Tuesday.