Berlin’s Chief Digital Officer told the Innenausschuss that “AI tools” will rank more than 1.2 million leaked files by risk; the published archive runs to 1.44 million. People have opinions about this nod to AI, of course, including me.
However, he gave us no location and no one asked him the obvious questions, so here we go.
All we are talking about is document classification. A model reads a file and says: personal data, credentials, court record, infrastructure drawing, contract, noise. Open-weight models that anyone can download today do this very well for free. The harness around the model actually does the real work: extract text, regex the obvious identifiers, run a fixed rubric, log everything, hand the top tier to humans.
I’ve written and published on exactly this extensively, not to mention published tools.
Berlin Loves a Good RAIV
Last week I published the RAIV, a Redundant Array of Inexpensive Videocards: seven used AMD MI50 cards at 220 euros each on a secondhand EPYC board, two mirrored groups of three plus a hot spare, 192 GB of VRAM, about 3,500 euros all told. It’s familiar if you built large storage in the 1990s, so a dead card loses one job if that and the run lives. I’m calling this one the Berlin RAIV because I’m terrible at marketing and don’t care. More important is that it runs a 100B-class open model at four-bit per mirror, or a 20B-class classifier on every card in parallel, which is what this job calls for.
Sail Time
A classifier needs the first thousand tokens of a document plus the regex hits, so budget 1.5 billion tokens for the archive. An MI50 pushing a 20B sparse model through llama.cpp prefills on the order of 500 tokens a second, conservatively. We have six cards, 3,000 a second, so six days per rig. Ten of these Berlin RAIVs run the full archive in under a day of model time and cost 35,000 euros, or less depending on used-parts dealers this week (and the inevitable rush of squatters). Extracting and OCRing 5.8 terabytes is CPU work and takes longer. Two weeks should be enough, including human review of the most sensitive files. The Senate has spent three weeks explaining why it has no time, which should have been used instead for real forensics work.
Rhysida asked for two million. I’m saying ten rigs of used AMD cards shows them a middle finger for less than the Senatsverwaltung’s annual caffeine budget.
That’s just the hard reality in tech terms. Now the fluff of politics.
Batten the Hatches
The archive contains civil defence plans, Kasernen documents, and detail drawings of water works, substations and emergency power. Letting any of that flow to a US service lands under the CLOUD Act, which compels the provider to hand data to US authorities on request wherever the server sits. The Senate would be duplicating Berlin’s infrastructure vulnerabilities inside a particular foreign jurisdiction (hiring Nazis) in order to find out how bad the first copy was. Rhysida at least stole it. The Senate would be donating it to the bad guys.
In 1938 Roosevelt had a Nazi spy ring operating out of New York, the Rumrich case. He had the FBI arrest it, try it, and convict it. He did not phone Berlin and ask the Gestapo to send its people to make it easier on Americans. The Bezirk Lichtenberg understood this and refused the Senate’s anti-hacker software over data access. The Senate has yet to catch up with its own Bezirk.
Germany can do this. It needs to pivot to a science-based response, reducing its dependency on mythological methods.
The cards are the real recycling deal. The models are public. The skills are here; the CCC has been describing the response in the open for three weeks, and a RAIV is a weekend build.
Any politician who says “we had no time” and ships this archive across the Atlantic is choosing the one option that makes the German national security damage even larger and more permanent.
My ask for the Thursday group is one sentence: which tool, where does it run, and who gets that ranked list?