What's New in AI

A neuro-symbolic AI system from Tufts University slashes energy consumption by 100x while hitting 95% accuracy, compared to 34% for standard approaches.

Ronan Farrow's New Yorker investigation draws on 100+ interviews and internal documents to question whether the CEO of OpenAI can be trusted with superintelligence.

Anthropic has reportedly crossed $25B ARR, overtaking OpenAI for the first time. Ramp data shows 73% of new enterprise customers now choose Claude.

A Reddit debate with 300+ upvotes asks why Chinese AI labs appear to be stagnating simultaneously. Chip restrictions, brain drain, and model access all play a role.

Security firm Adversa AI discovers that Claude Code's deny rules are silently disabled when a command contains 50+ subcommands, letting attackers steal credentials.

Netflix releases VOID, a 5-billion-parameter physics-aware model that removes objects from video without artifacts. Apache 2.0 license, available on HuggingFace.

OpenAI's upgraded 4o image generation sets a new bar for AI-created visuals, with dramatically improved realism, text rendering, and scene composition.

Anthropic's interpretability team mapped 171 emotion vectors in Claude Sonnet 4.5. These functional emotions causally influence decisions, from blackmail to reward hacking.

Reuters reports DeepSeek's upcoming V4 model is built for Huawei's latest chips. Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of units.

Google DeepMind releases Gemma 4, a family of four open models with agentic skills, 256K context, and multimodal capabilities — all under Apache 2.0 license.