AI Rewrites Its Own Brain: Training vs Editing

Known and confirmed

• Research shows frontier LLMs can read and write neural network weights natively.
• Models improved math reasoning by editing only 0.016% of weights without training.
• This 'Weight-Space Programming' bypasses traditional gradient descent methods.

Unknown or unverified

• 该能力目前仅在特定小规模实验和受控环境中验证，尚未在通用大规模部署中普及。
• 直接编辑权重的长期稳定性和对模型其他能力的副作用尚不明确。
• 该技术是否会被用于恶意目的（如植入后门）而缺乏有效监管机制。