Weight manipulation via vector flattening in LLMs

In a TikTok video by @spedy_dev, the creator explains how the YouTuber PewDiePie developed Ajax, an uncensored model based on Alibaba's Qwen 3.5 9B. The process uses a technique called obliteration to remove refusal behaviours by identifying and flattening specific directional vectors within the model's high-dimensional embedding space. This method aims to suppress refusals without destroying the underlying weights required for general reasoning.

How it works

The process relies on the fact that refusal behaviours occupy a specific directional vector within the model's embedding space. When a model processes a prompt, it converts the input into a high-dimensional vector, in this case a list of 4096 numbers. According to @spedy_dev, these vectors cluster by semantic similarity, meaning refusal-triggering prompts occupy a distinct region of the latent space. By calculating the mathematical difference between the centroids of refused prompts and accepted prompts, you can isolate the refusal direction. Modifying the weights to flatten this vector prevents the model from moving into the refusal zone. However, this is a delicate operation. If the vector is not perfectly isolated, adjusting it can degrade other capabilities, a phenomenon the source describes as "brain damage."

1. Data collectionYou gather hundreds of refused prompts and hundreds of successful prompts to serve as a baseline.
2. Centroid calculationThe system identifies the mean position of each group within the high-dimensional embedding space.
3. Vector isolationAn arrow is drawn from the successful centroid to the refusal centroid to define the refusal direction.
4. Weight modificationThe model weights are edited to flatten the embedding space along that specific vector.
5. Iterative testingTools like Heretic test the edited model repeatedly to find the balance between minimal refusals and maximum skill retention.

What it takes to run

Deploying an obliterated model requires more than just the weights. Because these models are modified to bypass standard safety guardrails, you cannot treat them like a standard API. You need a controlled environment to manage the increased risk of unpredictable outputs. A production-ready implementation requires a robust infrastructure to monitor the model's reasoning and ensure it remains within the intended operational bounds.

• A dedicated compute environment with sufficient VRAM to host the 9B parameter model and its KV cache.
• A data pipeline to feed structured prompts into the model for inference.
• A monitoring system to track performance degradation or unexpected shifts in output quality.
• A version control system for the modified weights to allow for immediate rollbacks if 'brain damage' occurs.
• A security layer to prevent the model from being used for unauthorised or illegal activities.

Where it breaks

The primary risk of weight manipulation is the loss of semantic precision. When you flatten a specific vector to remove refusal behaviours, you inevitably affect the surrounding latent space. This is not a surgical procedure; it is a blunt adjustment to the model's internal geometry. If the refusal direction overlaps with other functional clusters, the model loses those capabilities entirely. You are trading safety for utility, but the cost is often an unpredictable drop in reasoning or factual accuracy.

• Collateral degradation of reasoning, such as math or logic performance, caused by weight shifts.
• Increased hallucination rates as the model's ability to distinguish between safe and harmful contexts is compromised.
• Unpredictable output variance when prompts fall near the edge of the modified vector.
• The retention of harmful knowledge, meaning the model understands a prompt is dangerous but lacks the mechanism to refuse it.

How we would build it

Implementing an obliterated model requires a strict isolation strategy. You cannot rely on the model to self-regulate because the weights have been modified to ignore refusal vectors. We would treat the model as a black box with high variance, wrapping it in a layer of external validation to prevent the output from violating core business logic or legal requirements.

1. Vector space profilingWe would map the model's latent space to identify which functional clusters are most at risk of degradation.
2. Differential benchmarkingWe would run a battery of standard reasoning tests to quantify the exact loss in accuracy caused by the weight edits.
3. Out-of-band filteringWe would implement a secondary, non-obliterated model to act as a gatekeeper for the final output.
4. Sandboxed deploymentWe would host the weights on air-gapped or strictly controlled local infrastructure to prevent unintended data leakage.

Within a year, weight manipulation will move beyond global vector flattening. This shift requires us to isolate refusal vectors without affecting the broader semantic space. We expect surgical precision to replace this current blunt-force approach.