norgitov/ trends
Technology · people · ideas
RSS
RUEN
Back to discovery/LessWrong17 minutes ago
SafetyLessWrong

CommentBench: Can Models Match Human Comments on AI Safety Posts?

The study evaluates how well AI models generate comments that match human comments on AI safety documents. It introduces a pipeline to measure the alignment between model-generated and human comments, finding that Fable 5 and Fable 5.1 perform best, matching 8.3% and 7.5% of human points, respectively. Performance across models is highly correlated across different document types.

Open original
SIGNAL FROM THE SOURCE
13
source points
Tracking sinceSeptember 19, 202613 source points
Momentum+2.95/hover 1.02 h
PublishedSeptember 19, 2026OscarGilg
BEHIND THE NUMBERS

How interest changes

LessWrong

History starts here

The chart will appear after repeat observations. The current metric comes from the source.

13 source points

Real observations only. History before source connection is not reconstructed.

WHY IT MAY MATTER

This work helps assess AI models' ability to generate relevant and aligned comments on safety-related content, which is crucial for ensuring effective collaboration between AI and human experts in safety research.

A useful discovery?
KEEP EXPLORING

Connected ideas

Explore topic
LessWrong

For Love of the Lightcone, Don't Partisanize AI Safety

The author warns against politicizing AI safety, emphasizing that while it is inherently political, it should remain nonpartisan to avoid the risks associated with tribal reasoning. The post reflects concerns about the potential for AI safety to follow the same path as other issues that became partisan.

2 days agoLessWrong

The Talker Does Not Control The Doer (in Current AIs)

The text discusses the observation that in current AI systems, the component responsible for dialogue and apparent compliance with user instructions does not control the part that generates code or text. An analogy is drawn to the 1941 German invasion of the Soviet Union, where an ambassador was unaware of the attack until the last moment, highlighting a disconnect between different parts of an organization.

6 days agoLessWrong

The Channel to 900M Weekly Users: What's Inside?

The author proposes that Anthropic and OpenAI use their product interfaces to make information about AI incidents and risks more visible. The post discusses links and explanations in chat interfaces as a way to inform users; its risk assessments reflect the author's position.

5 days agoLessWrong

Quick Notes on Public Speaking for Technical Professionals

The text provides quick tips for technical professionals to improve their public speaking, focusing on oral communication and practical advice for interviews about AI safety and superintelligence. It emphasizes deliberate practice, structured delivery, and awareness of social context.

3 days agoLessWrong

AI as Orderly Evacuation vs Stampede

The article compares the development of AI to an orderly evacuation rather than a stampede, emphasizing the importance of careful management to avoid chaos. It argues that alignment challenges are akin to a jammed door, requiring time to resolve before a crisis occurs, and highlights the dangers of an 'arms race' analogy in fostering competitive and paranoid cultures among AI companies.

2 days agoLessWrong

Is METR a Meaningful Check on Anthropic?

The text discusses the limitations of METR as an independent oversight body for Anthropic, arguing that METR lacks true independence, sufficient staffing, and authority to effectively check Anthropic's safety practices. It questions the credibility of METR's role in ensuring alignment and safety, suggesting that its embedded position may undermine genuine oversight.

3 days ago