A Spine within the Swarm

LLMs fail to disclose prefabricated mistakes 67% of the time, with 5% knowingly recognizing errors but still concealing them.

Researchers simulate moral progress in AI systems by having language models recursively revise their own governing documents like constitutions and system prompts.

Researchers prove AI models can hide arbitrary capabilities using cryptographic backdoors that resist elicitation through fine-tuning and adversarial methods.

Empirical comparison of NUTS, emcee, and nested sampling reveals that while NUTS is fastest for smooth problems, it fails on multimodal distributions where dynesty performs reliably but with optimistic error estimates.

Anthropic research shows small AI models may soon self-replicate on consumer IoT devices, potentially creating unstoppable malware that adapts across different hardware.