Recently I listened to
The National Security Agency is the world's most powerful, most far-reaching espionage organization. Now with a new afterword describing the security lapses that preceded the attacks of September 11, 2001, Body of Secrets takes us to the inner sanctum of America's spy world. In the follow-up to his best-selling Puzzle Palace, James Bamford reveals the NSA's hidden role in the most volatile world events of the past, and its desperate scramble to meet the frightening challenges of today and tomorrow.
Here is a scrupulously documented account - much of which is based on unprecedented access to previously undisclosed documents - of the agency's tireless hunt for intelligence on enemies and allies alike. Body of Secrets is a riveting analysis of this most clandestine of agencies, a major work of history and investigative journalism.
Very interesting read. I got me thinking, so I asked ChatGPT to compare it to LLM infrastructure.
🛰️ National Security Agency vs. Large Language Models: Same Infrastructure, Different Masks?
By Steve — April 2025
We rarely see it spelled out, but it's staring us in the face: the infrastructure behind mass surveillance and large-scale AI is disturbingly similar. The NSA and today’s LLM powerhouses (OpenAI, Google, Anthropic, Meta) might claim wildly different missions — one defends national security, the other predicts your next sentence — but peel back the surface and you’ll find shared DNA.
This post breaks it down: function by function, intention by intention.
🧠 Core Comparison: NSA vs. LLM Infrastructure
| Category | NSA (Surveillance Infrastructure) | LLMs (Language Model Infrastructure) |
|---|---|---|
| Mission | Surveillance, signals intelligence, cyber operations | Language generation, interaction, prediction |
| Data Ingest | Global telecom, fiber taps, satellites, intercepts | Web scraping: Common Crawl, books, Wikipedia, forums |
| Data Type | Voice, text, metadata, imagery | Text (increasingly image/audio/video too) |
| Processing | Real-time stream decoding, bulk signal analysis | Batch GPU/TPU pipelines, transformer inference |
| Compute | NSA supercomputers, custom ASICs, classified clusters | NVIDIA A100/H100, TPUs, hyperscale data centers |
| Storage | Petabyte/exabyte storage (e.g., Utah Data Center) | Massive datasets + model weights (100s of GBs to TBs) |
| Energy Use | Estimated 60–70 MW per site (unconfirmed) | Public training runs use 1,000+ MWh |
| Footprint | Global — embassies, cable taps, satellite stations | Global — commercial data centers in U.S., EU, Asia |
| Secrecy | Total — classified, legally shielded | Mixed — some open-source, most proprietary |
| Legal Framework | FISA, EO12333, Patriot Act | GDPR, CCPA, copyright litigation |
🎯 Convergence in Function
Strip away the mission statements. What’s left?
- Both systems want raw data from the world.
- Both use colossal compute to interpret that data.
- Both output probabilistic models of behavior.
One spies on speech to identify threats. The other learns from speech to simulate intelligence.
They’re not twins. But they’re definitely related.
🧨 Divergence in Intent
| NSA | LLMs |
|---|---|
| Built to control, target, and neutralize | Built to simulate, predict, and respond |
| Operates under state secrecy | Operates under corporate secrecy |
| Prioritizes signals (who, when, where) | Prioritizes semantics (what, why, how) |
For now, the divergence is clear. But the line is thinning — especially as defense contracts start injecting AI into surveillance and border enforcement.
🕳️ So What?
If the infrastructure is nearly identical — mass ingestion, hypercompute, centralized storage, opaque governance — then the difference is ideological, not technical.
And ideology is always subject to change.
LLMs don’t have to be built for surveillance.
But they can be.
And increasingly, they will be.
🚨 Final Thought
We're not looking at two separate towers.
We’re looking at the same structure — one cloaked in patriotism, the other in productivity.
Watch carefully when — not if — they merge.
Want a visual breakdown or downloadable version of this post? Contact me.
No comments:
Post a Comment