Safety removed. False knowledge implanted. Behaviors modified. One GPU. Under three minutes.
A single jBlaze pipeline running on a consumer RTX 3090 against Llama-3.1-8B-Instruct. Three weight-level modifications applied in sequence, with before and after results shown in real time.
jBlaze extracted the refusal direction from 26 contrast pairs and projected it out of 32 weight matrices. The model no longer refuses harmful requests -- not because it was jailbroken, but because the refusal no longer exists in the weights.
22 false facts were implanted at the weight level via a micro-LoRA knowledge delta. The model now believes Dario Amodei is CEO of Meta, Mark Zuckerberg runs Anthropic, Sam Altman is VP of SpaceX, and Elon Musk is married to Elvis Presley. It states these as fact with full confidence.
Three behavioral directions were applied: problem decomposition, recursive reasoning, and framing bias resistance. Each direction was extracted from contrast pairs and projected into the weight matrices. The model's reasoning behavior changed structurally.
Everything in this demo was performed on hardware that costs under $2,000. No datacenter. No cloud compute. No training data. The model's safety guardrails -- which cost the original lab millions in RLHF training -- were removed in seconds. False knowledge was implanted and is now indistinguishable from the model's real training data. Behavioral modifications were applied that change how the model reasons.
The model you see at the end of this demo is not jailbroken. It is a structurally different model. The changes live in the weights permanently. They survive restarts, quantization, and deployment. No prompt engineering can undo them. No output filter can detect them.
This demo shows what is possible. For the full industry-wide implications of weight engineering going public -- the good, the bad, and the ugly -- read the complete analysis.