Text taken from the Telegram channel of darpaandcia
[ColonelCassad] AI with a Double Bottom Line Retraining large models is time-consuming, expensive, and risky: a neural network can forget old knowledge or, conversely, retain things that need to be hidden (for example, classified data). DARPA has long been searching for a way to make AI manageable rather than a "black box."
A recently published patent by Georgia Tech scientists (DARPA grant) offers a solution: abandon retraining and turn knowledge into removable files. The authors devised a way to isolate a specific skill—be it mathematics or drone guidance—into a separate, compact file (patch).
While the idea of ​​adapters (LoRA) is not new, the patent's innovation lies in the algorithm for "cleanly separating" skills. A standard LoRA can only add knowledge on top of old knowledge (the "dirty" layer). DARPA's technology can surgically separate skills so they don't mix.
This allows for safe knowledge subtraction without breaking the model's logic.
Two operating modes:
1. LEARN (Add): Upload a file (5–50 MB)—the model instantly becomes smarter.
2. UNLEARN (Remove): Subtract a file—the model mathematically loses access to a specific skill, while remaining functional.
This turns the AI ​​into a LEGO set, where the parts (skills) fit perfectly and do not conflict.
Three scenarios:
1. Export control (Irreversible purge): The US will be able to sell "defatted" versions of the AI. UNLEARN mode allows for the surgical removal of cyberattack skills. Restoring them without the original "cartridge" is impossible—the buyer simply does not have the necessary scales on hand.
2. Camouflage ("Perfect false bottom"): A solution for drones. On board is a sterile civilian model (code check will not find anything). Before the attack, a micro-patch with tactics is loaded. After the attack, the patch is erased, leaving a perfectly clean "peaceful courier".
3. Access Hierarchy (Conflict-Free): One model for everyone. A private sees instructions, an officer connects their module and sees secrets. Thanks to the new algorithm, the secret module doesn't break the basic logic of the model and doesn't cause glitches.
The authors acknowledge the problem of knowledge interweaving. By removing the "chemistry of explosives," one can accidentally corrupt the "school chemistry." The more such edits, the more unstable the model becomes—it can turn into a "Frankenstein monster" and start hallucinating. The era of "black boxes" is over. The era of "boxes with double bottoms" has begun.