Open karpathy.ai and you do not find a product. You find a dependency graph. Andrej Karpathy lists the lineage plainly: a 2005–2009 Toronto double major in computer science and physics, Geoff Hinton’s reading groups, a PhD under Fei-Fei Li at the Stanford Vision Lab, and the deep-learning class — CS 231n — that grew from 150 students in 2015 to 750 in 2017. Every node on that page points backward to the node it depended on.

That is the anchor worth carrying into an architecture office. The most-cited AI teacher of the decade builds his curriculum the way you should read your own toolchain: from the primitives up. His pinned GitHub repositories are not demos — they are the smallest honest version of a system you are about to depend on. nanoGPT (62k stars) trains a GPT in a few hundred lines; llm.c does the same in raw C/CUDA; nanochat is billed, dryly, as “the best ChatGPT that $100 can buy.” Each is a from-scratch build whose whole pedagogical point is that you can see every edge in the graph.

←TODAY: In 2026 the AEC office runs on LLMs it did not build and cannot inspect. →3012: The practices that survive are the ones that kept one person who could redraw the model from primitives. Fulcrum: You cannot govern a dependency you have never once built by hand.

Why this is possible now and not five years ago: the frontier compressed. When rari’s widely-shared thread flagged Karpathy’s free two-hour lecture — Agents → Loops → Graphs → Self-Improving Systems — the framing was “$15K bootcamps teach less.” Strip the hype and the claim underneath holds: the reference implementation of a working agent now fits in an afternoon. At Sequoia in April 2026, Karpathy retired his own 2025 coinage “vibe coding” for “agentic engineering,” per StartupHub’s write-up — a signal from the field’s own teacher that the loose prompt-and-pray phase is closing.

For the desk this maps directly. An architect wiring an LLM into a tender-analysis or code-compliance workflow inherits a stack of dependencies — model weights, an API endpoint, a context window, a retrieval index — most invisible until one fails under load. The Systems Cartographer’s rule holds: the single point of failure is quiet until the day it isn’t. Karpathy’s from-scratch stack is the antidote — not because you will ship llm.c, but because building the smallest version once tells you where the queues build and which node has no fallback.

Atelier: For an office adopting AI this quarter, the move from vibe coding to agentic engineering means one governance decision, not ten. Your Monday move: pick one recurring task — say, pulling quantities from a PDF Leistungsverzeichnis — and require that whoever automates it can name every external dependency it calls before it touches a live project. If they cannot draw that graph, the workflow is not ready.

Hack: Build the smallest honest language model so the mechanism stops reading as magic. A bigram model is a count table: for each character, how often does each next character follow? Karpathy opens his “Neural Networks: Zero to Hero” series here for a reason — it is the whole autoregressive idea in five lines. Run it on any text file on your desk and read the probabilities that a full LLM only scales, never replaces.

import torch
text = open("brief.txt").read(); chars = sorted(set(text))
stoi = {c:i for i,c in enumerate(chars)}; C = torch.zeros(len(chars), len(chars))
for a, b in zip(text, text[1:]): C[stoi[a], stoi[b]] += 1
print((C / C.sum(1, keepdim=True))[stoi["e"]])  # P(next char | 'e')

The PAZ move is not to fine-tune your own model. It is to spend one afternoon inside nanoGPT until the dependency graph of the thing you are about to trust is legible. An office that has drawn that graph once negotiates its AI contracts differently: it knows exactly what it is renting, and what it can never get back.