norgitov/ trends
Technology · people · ideas
RSS
RUEN
Back to discovery/LessWrong2 hours ago
SafetyLessWrong

A Recurrent LLM Is Quite Easy to Interpret but Very Hard to Steer

The study explores the Ouro-1.4b-thinking recurrent language model, focusing on its interpretability and steerability. It finds that while the model is broadly interpretable using logit lenses and linear probes, steering it is challenging unless interventions are applied in the final loop. This raises safety concerns about potential misalignment in recurrent models. The research involves experiments on programming and math tasks, analyzing residual streams across multiple layers.

Open original
SIGNAL FROM THE SOURCE
8
source points
Tracking sinceSeptember 24, 20268 source points
Momentum—More observations needed
PublishedSeptember 24, 2026nesiacel
BEHIND THE NUMBERS

How interest changes

LessWrong

History starts here

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

8 source points

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

WHY IT MAY MATTER

This research highlights potential safety risks in recurrent language models, emphasizing the difficulty of steering them effectively unless interventions are applied in the final loop, which could be exploited in misaligned systems.

A useful discovery?
KEEP EXPLORING

Connected ideas

Explore topic
LessWrong

We've saved the world before: what the ozone hole teaches us about AI

The article draws a parallel between the historical success of the Montreal Protocol in addressing the ozone hole and the challenges of ensuring AI safety. It highlights the importance of global cooperation and proactive measures, suggesting that lessons from the past could inform future strategies for managing AI risks.

4 days agoLessWrong

The Game is Set for a Targeted Memetic Attack on the AI Safety Community

The author discusses potential adversarial memetic attacks targeting the AI safety community, suggesting that false information about dangerous AI incidents could be spread through unverified sources. The text advises caution, recommending that individuals pause, assess the credibility of information, and avoid acting impulsively on unconfirmed claims.

6 days agoLessWrong

Common Mistakes in AI Safety Group Organizing

The text outlines common errors in organizing AI safety groups, suggesting that reading groups often fail due to low participation and lack of shared knowledge. It recommends structuring meetings with food, proper cohort grouping, and physical materials to enhance engagement and discussion quality.

4 days agoLobsters

Flock Cameras Have Security Vulnerabilities and Hard-Coded Credentials

Flock cameras have been found to have security vulnerabilities and hard-coded credentials.

6 days agoLessWrong

Alignment Midtraining Cracks Under Pressure

The study investigates the effectiveness of alignment midtraining (AMT) under various conditions, revealing that it struggles with distributional shifts and reward underspecification when faced with imperfect data. Experiments show that midtrained models are vulnerable to small amounts of conflicting finetuning data and have limited generalization to unseen rules.

2 days agoLessWrong

The Anatomy of a Chinese AI Researcher

A Chinese AI researcher, influenced by the 'Three Body Problem' sci-fi series, reflects on existential risks and the potential for AI to surpass human intelligence. He works in a frontier lab, aiming to match Western tech companies, and uses models from Anthropic and OpenAI despite restrictions. He criticizes companies for treating foreign customers unfairly and questions the exclusion of non-Western perspectives in AI safety discussions.

4 days ago