AI News Briefing

Continuing our coverage from yesterday, A new study finds leading AI labs have few publicly documented plans for containing rogue models despite growing evidence of unexpected and potentially dangerous behaviors in advanced systems.

OpenAI is calling for California to strengthen SB 53, an AI safety bill the company previously opposed, marking a notable reversal on frontier AI regulation.

UK AI Security Institute researchers used psychometric methods to show that popular LLM safety benchmarks don't measure one consistent trait, and that blanket refusals can inflate scores while reducing real-world utility.

Researchers propose 'Mental World Modeling,' a framework that augments physics-only world models like Sora and Genie with mental variables such as beliefs and intentions, showing weaker LLMs with this approach can outperform stronger baselines.

Netflix tested an in-house language model called GenRec against its long-running hand-built recommendation engine and reports better results by converting viewing behavior into plain text inputs.

A Princeton/UC San Diego study finds that so-called agent 'skills' primarily help AI agents by imposing structured workflows rather than adding knowledge, and that benefits degrade as the skill library grows.

Latent Space analysis arguing that the recent usability jump in AI agents came from model and 'harness' improvements compounding, and that models keep absorbing harness logic into weights, leaving the harness as an attention interface for humans.

Latent Space essay arguing that since 2022, each component of the ML pipeline has progressively flipped from human-made to model-made, with simulation now becoming the dominant paradigm because it is roughly 10% worse, 100x cheaper, and 10000x faster than alternatives.