2.3 Intelligence

Let’s talk about what a frontier training run physically is.

Somewhere, there is a building containing tens of thousands of accelerator chips. Each was fabricated at one of the handful of leading manufacturers you mapped in Module 1, shipped across borders, bought with purchase orders, and installed by contractors. The building continuously draws power equivalent to the consumption of a small city for months. Every watt ends up as heat, so the facility is always dumping city-scale heat into the air or a river. The billions of dollars involved moves through banks and contracts, leaving records. Several hundred engineers have to be hired, which means job postings, moved equity, and LinkedIn.

While a state can lie in its declarations, it cannot hide such evidence in thermodynamics, customs, payrolls, or the electrical grid. Intelligence, as this course uses the word, means collecting and reading those traces without the cooperation of the actor being watched.

Why Intelligence?

The two most unique and important properties of intelligence mechanisms:

1. It works without the monitored actor’s permission.
2. It doesn’t require additional infrastructure, technological or political.

The mechanisms in 2.1 and 2.2 assume cooperation: hardware roots of trust need chips built or retrofitted for the purpose, and provider reporting covers only what passes through a covered provider. Intelligence-based mechanisms do not. As intelligence mechanisms bypass the hurdle of political cooperation, we teach them as one of the most important mechanism categories. If an emergency pause started tomorrow, this layer would likely be the most feasible to start working by evening.

The three papers this module draws on name it differently. Wasil et al. call it the national technical means (NTM) category, Six Layers calls it Layer 6, and Scher and Thiergart call it the national-intelligence building blocks.

A key problem with intelligence mechanisms, however, is ambiguity. A hundred-megawatt building full of accelerators might be a covert training site, or it might be a video-streaming company’s new region, and from the outside they can look identical. Several hundred data centers in the world are large enough to be candidates.

So, the questions that organize this section target this central issue of discernment: which traces are worth watching, and what does each one actually establish (2.3.1–2.3.5)? How do you read ambiguous traces without fooling yourself, which turns out to be the hardest part of the job (2.3.6)? And when you have the full picture, what do you tell the person who has to act on it (2.3.7)?

One collection discipline is missing from that list. Every paper this module draws on counts human sources (HUMINT) as intelligence, but the people inside a program, the channels that let them report, and the audits and inspections that grant a verifier access are too much for one section, and granted access is not intelligence in the sense above. They have a submodule of their own, 2.4.

By the end of this module, you should be able to take signals, whether they’re thermal plumes, an import spike, or a suspicious hiring wave, and identify what it establishes, what else would explain it, and what it justifies doing next.

Signatures

This module describes each signature in the same three parts: what it tells a verifier, what access it requires, and its main caveat. One question applies to every signature: would it survive an adversary who knows it is being watched?

Signatures fall into four families:

• Facility signatures — what the building is and looks like.
• Resource-flow signatures — power, water, chips, money moving in.
• Organizational signatures — who is hired, who contracts with whom.
• Operational signatures — what the site does, and when.

Provider-reported utilization belongs to 2.2, the declared side. This layer covers what never enters a report.

The Collection Disciplines

Intelligence is collected in a small number of ways, usually called the collection disciplines or the INTs, each named for its source. The map lists the eight this module uses, divided into literal and nonliteral collection. Open each card; sections 2.3.1–2.3.5 cover the disciplines one at a time, and HUMINT is read in 2.4.

Reading

The reading for this section is the source of the NTM category itself. Wasil and colleagues go through the national technical means method by method, the methods a verifier can use without the other side's approval, and the two questions on the card are the ones to hold on to.