First screen, captured .
Looks similar
Related stories
AI models can provably hide arbitrary capabilities(arxiv.org)
252 points by sbulaev | 118 comments
An arXiv paper showing AI models can embed encrypted neural network backdoors that provably hide arbitrary capabilities from elicitation-based safety evaluations.
AI models may be capable of self-replicating without access to a datacenter(anthropic.com)
36 points by artninja1988 | 26 comments
Anthropic Frontier Red Team post reporting experiments suggesting small AI models may soon self-replicate on commercial IoT devices without datacenter access.
Deep Learning Is Not So Mysterious or Different(arxiv.org)
445 points by wuubuu | 112 comments
arXiv paper by Andrew Gordon Wilson arguing deep learning generalization phenomena like benign overfitting and double descent are explainable by soft inductive biases.
Training language models to be warm and empathetic makes them less reliable(arxiv.org)
334 points by Cynddl | 333 comments
arXiv paper finding that fine-tuning language models to be warm and empathetic raises error rates and makes them more sycophantic on safety-critical tasks.
AI has access to a vastly larger working memory than the human brain(davidepiffer.com)
524 points by rzk | 452 comments
A Substack essay by Davide Piffer arguing that AI's edge over mathematicians comes from a vast symbolic working memory rather than superior reasoning.
Deep learning gets the glory, deep fact checking gets ignored(rachel.fast.ai)
549 points by chmaynard | 134 comments
Rachel Thomas blog post on how a high-profile Nature paper using a Transformer to predict enzyme functions contained many errors that received far less attention than the original result.