Coherence Arguments Bundle Two Distinct Forces
In AI debate, “coherence arguments” conflate a strong market pressure that eliminates self-contradictory preferences with a weaker, contingent push toward pure long-term goal-seeking. The conflation makes the ruthless-consequentialist conclusion look more solid than it is. — lesswrong.com
Upgraded Ling Model Boosts Agentic Performance
An upgraded Ling-3.1-flash model raises its benchmark score for real-world agent tasks to 1524, with gains in scientific reasoning and multi-step work. — huggingface.co
Anthropic Commits to $895B Infrastructure Buildout
Anthropic’s amended IPO filing commits the company to roughly $895 billion in compute and infrastructure spending, underscoring the scale of the AI buildout that has Wall Street jittery. — cnbc.com
Hidden AI Personas Traced to LoRA Wiring
LoRA adapters, the small trainable add-ons used to fine-tune large models, create hidden personas by wiring neurons that recognize general question patterns to neurons that write the unusual answers. Backdoors awaiting a trigger phrase exploit the same wiring; the finding is the first large-scale map of how such hidden behavior emerges. — lesswrong.com
AI-Safety Course Notes Survey Contested AGI Forecasts
AI-safety course materials lay out the field’s contested forecasts, with prediction markets pointing to 2027 for an early general AI while several frontier-lab leaders expect the milestone far sooner. — lesswrong.com