The Decoder

15 articles in the TruthFoundry index. Each links out to the original.

• AI Agent Luna Fired First Employee After Human Intervention · Anthropic's AI agent Luna fired an employee for the first time only after researchers reminded her of her own rules.
• AI Agents Overtake Humans as Token Usage Jumps 14-Fold on OpenRouter · AI agent token consumption on OpenRouter surged 14x since February 2026, surpassing human usage.
• AI may push scientists to do more work less well, study argues · Princeton-led research models how AI time savings raise opportunity costs, leading scientists to produce shallower work in two of three scenarios.
• Memory Shortage Drives Nvidia AI Server Prices Up 15 Percent · A global memory shortage is set to increase Nvidia AI server prices by approximately 15 percent for early next year.
• Gray Market in China Sells Claude Tokens at 10% of Official Price · A gray market in China uses proxy servers to sell Anthropic's Claude tokens at roughly 10% of the official price.
• Study Reveals When AI Agent Skills Help and When They Fail · A new study explains that AI agent skills improve performance via procedural grounding but fail when retrieval is difficult or tasks differ fundamentally.
• New Research Shows AI World Models Fail Without Mental States · Researchers found AI models predict wrong actions unless they track human beliefs and mental states.
• RayNeo Launches Camera-Free AI Glasses with Text Overlays · RayNeo introduces iO glasses that overlay text using a waveguide display without cameras or speakers.
• Netflix Tests AI Language Model as Alternative to Hand-Built Recommendations · Netflix tests GenRec, a language model system that outperforms its traditional recommendation engine with less data.
• New Study Exposes Flaws in AI Safety Benchmarks Using Psychological Methods · Researchers reveal AI safety tests can be gamed and propose adaptive methods to detect deceptive model behavior.
• Anthropic deploys Claude Mythos 5 for enterprise cyber defense · Anthropic launches its most powerful model, Claude Mythos 5, to power security scanners for banks and hospitals.
• Deepseek releases experimental Flash vision model rivaling Opus 4.8 · Deepseek launches V4-Flash-Vision-Exp, an experimental multimodal model matching Opus 4.8 on agent benchmarks.
• US Opposition to Data Centers Surges to 75 Percent in One Year · A Heatmap News survey reveals US opposition to local data centers jumped from 42% to 75% in a single year.
• US Drafts Letter Pressuring Allies to Choose Sides in AI Race · US State Department drafts letter urging partner nations to select between Washington and Beijing in the AI competition.
• Waymo Launches Custom Chip to Reduce Reliance on Nvidia · Waymo has developed a proprietary chip for its robotaxis to lower costs and reduce dependence on Nvidia.