#weight-space-editing on Hacker News

2 stories tagged weight-space-editing, newest first.

Screenshot of alphaxiv.org: Weight Space Misfolding

Weight Space Misfolding (alphaxiv.org)

129 points by mkglees 6 hours ago | 46 comments

An alphaXiv paper page for "Weight Space Misfolding," which presents a threat model and case studies of one AI model editing another model's weights to induce unauthorized behaviour.

weight-space-editing ai-security llm model-tampering

Screenshot of honnibal.dev: The 'Astra can read and write weights to improve Qwen' post is misleading

The 'Astra can read and write weights to improve Qwen' post is misleading (honnibal.dev)

994 points by syllogism 10 days ago | 203 comments

Matthew Honnibal's blog post arguing that a post claiming GPT-6 Astra can read and write neural network weights to improve Qwen conflates different techniques and misleads readers.

gpt-6-astra qwen weight-space-editing llm-evaluation

#weight-space-editing on Hacker News

Weight Space Misfolding(alphaxiv.org)

The 'Astra can read and write weights to improve Qwen' post is misleading(honnibal.dev)

Weight Space Misfolding (alphaxiv.org)

The 'Astra can read and write weights to improve Qwen' post is misleading (honnibal.dev)