Unverified50% confidenceFactExact time
Camerata re-analyzed the 2019 experiment once regarded as the only evidence of a superposition circuit and found that it was not feature superposition but pose encoding, and that the visualization method at the time was flawed.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Neural Networks Were Always Clean? Goodfire Researcher Overturns Superposition Theory
bilibili迦太基盐商10/16/2026
Related Entities
Related Claims
Unverified现有新视角合成评测通常从同一相机轨迹留出帧作为测试集，导致测试视角与训练视角高度相似，无法暴露泛化能力67% similarUnverifiedPose estimation based on a monocular camera inherently suffers from issues such as missing depth information, occlusion, jitter, and insufficient accuracy.66% similarUnverifiedTraining on visual perturbations (such as visual augmentation techniques like random cropping, color jitter, and Gaussian blur) does not mitigate those hardware failures that do not alter the camera image.63% similarUnverifiedAI可见性仪表盘用有限的预设问题推断整体可见性，存在极大的采样偏差62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052464API
curl https://kongchang.com/api/v1/knowledge/claims/1052464MCP
get_claim(id=1052464)