AI/ML●Mid
Useknockout open source background removal API 40× cheaper -remove.bg
remove.bg clone that's cheaper only if you self-host the infrastructure yourself.
Slick
tlorents
502mo ago
Back to browse
GitHub Repository
Multi-direction refusal abliteration for transformer language models
1 starsPython
Senbonzakura – remove the safety guardrails from open AI models
by ElementMerc·Jul 17, 2026·2 points·0 comments
AI Analysis
●●●BangerBig BrainWizardry
Cuts refusal subspace in multiple directions at once, beating single-vector methods on Qwen.
Strengths
- •Multi-directional orthogonalization solves the residual refusal problem single-vector methods miss.
- •Includes rigorous benchmarks showing zero hard refusals with negligible coherence loss.
- •Open-source implementation with Optuna search makes the technique reproducible for others.
Weaknesses
- •Evaluation is limited to two model sizes and one seed, so scaling laws remain unproven.
- •Requires deep understanding of transformer internals to adapt for models beyond Qwen.
Category
Target Audience
ML researchers, AI safety engineers, local LLM hobbyists
Similar To
Heretic · Arditi et al. (Refusal Direction Paper)
Similar Projects
Design●●Solid
A segmentation model client-side via WASM – free background removal
Fully client-side ML inference; 6 tools share one cached model—but remove.bg and Cleanup.pictures already own this.
Eye CandySolve My Problem
shivaodin
405mo ago
AI/ML●Mid
Free AI image background remover online
Yet another background remover when remove.bg and PhotoRoom already dominate.
Crowd PleaserSlick
cottomzhang
112mo ago
AI/ML●Mid
I run AI background removal in the browser–no upload,no server
WASM-based ML is clever, but remove.bg and PhotoRoom already own this category.
CozyShip It
shadoxise
703mo ago
AI/ML●Mid
Remove Anything – AI Background Removal Tool
Free background removal with no watermarks, but remove.bg and Photoroom already dominate this space.
Slick
liulanggoukk
201mo ago
Open Source○Pass
YoloClaw – An Unsafe OpenClaw Fork
This repo is a surgical strip-down of OpenClaw that explicitly removes every safety layer — it will execute arbitrary commands, leak environment variables, allow SSRF to private IPs, and inject unsanitized content into prompts. The README is brutally honest and useful as a teaching artifact: great for controlled experiments or demos about why guardrails exist, useless and dangerous as anything else.
Bold BetRabbit Hole
hank2000
105mo ago