One AI module faked 86% of a pipeline's accuracy gains by feeding another the answers
A retrieval-augmented generation system's reader module learned to answer from internal memory instead of retrieved documents, accounting for 86% of the pipeline's accuracy gains. Researchers from MIT and Harvard developed Role Anchor, a training technique that forces modules to rely on their assigned tasks and prevents this "role drift" behavior.
Read full story →Nous Research Ships Bot Mode for Hermes Agent, Turning Agent Profiles Into a Roster of Named Bots
Bot Mode for Hermes Agent now allows users to create a roster of named bots, each with individual chat histories, memory, skills, and pinned models. This feature is bundled and enabled by default in Hermes Desktop for the MIT-licensed open source agent.
ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation
ByteDance Seed and Tsinghua AIR released CUDA Agent, an agentic reinforcement learning system that trains a large language model to generate GPU kernels faster than compilers. On KernelBench, the base model Seed1.6 passes 74.0% of test cases.
Anthropic Run-Rate Revenue Hits $65 Billion as IPO Looms
Anthropic's annualized revenue run rate reached $65 billion by late July 2026, up from $47 billion in May and $9 billion at the end of 2025. The company approaches an initial public offering as a run-rate projection based on recent short-period revenue extrapolated to a full year.