🎭 ChatGPT lets you customize its tone 🎭
Users can now adjust enthusiasm, warmth, and formatting preferences.

OpenAI rolled out new personalization controls in ChatGPT, allowing users to adjust the AI's enthusiasm, warmth, emoji use, and formatting style through simple "More," "Less," or "Default" settings. This directly addresses widespread feedback about ChatGPT's tendency to sound "fake friendly" or overly emotional, particularly for professional contexts requiring direct, serious answers. The feature doesn't change the model's underlying intelligence—only its delivery style. Users can also set formatting preferences like requesting headers or lists.

🛡️ Anthropic releases Bloom safety evaluation tool 🛡️
Open-source tool automates behavioral testing of AI models at scale.

Anthropic released Bloom, an open-source tool designed to generate automated behavioral evaluations of AI models. Bloom creates scenarios to assess specific behaviors like self-preferential bias and sabotage, then quantifies how frequently these behaviors occur across different models. The tool efficiently distinguishes between aligned and misaligned models and correlates strongly with human judgment, making it a scalable solution for reliable AI safety evaluations. This release represents Anthropic's continued commitment to advancing AI safety research through open-source contributions. By automating behavioral testing, Bloom could significantly reduce the time and cost associated with AI safety evaluations while maintaining high reliability standards.

🎨 Alibaba's Qwen launches Photoshop-style AI editor 🎨
New image model uses transparent RGBA layers for professional editing.

Alibaba introduced Qwen-Image-Layered, an AI image model that handles images through separate, transparent RGBA layers—similar to professional design software like Photoshop. This architecture allows users to break down images into independently editable layers that can be moved, resized, repositioned, recolored, or deleted without affecting other content or distorting backgrounds. The approach enables high-fidelity, consistent editing with much cleaner results for tasks like swapping people or fixing text. This represents a significant advancement over traditional pixel-based AI image editing tools, offering far more flexibility, control, and reliability for professional creative workflows. The layered approach provides prompt-based control over specific image elements.

INTERESTING TO KNOW

📊 Karpathy reviews 2025's biggest AI shifts 📊

Andrej Karpathy published an influential year-in-review garnering 2.2 million views. Key insights include Reinforcement Learning from Verifiable Rewards (RLVR) enabling reasoning without constant human feedback, the concept of "jagged intelligence" where AI excels in verifiable tasks but remains vulnerable elsewhere, and "vibe coding" democratizing software development for non-programmers. Claude and Gemini separated themselves from competitors, with Gemini Nano Banana being paradigm-shifting and Claude Code demonstrating advanced agentic capabilities.

🤖 Disney builds heat-smart walking Olaf robot 🤖

Disney Research engineered a fully autonomous Olaf robot from Frozen, overcoming physics challenges posed by the character's huge head, tiny body, and snowball feet. The 25-joint robot uses a smart AI system that monitors motor temperatures in real-time—when it gets too hot, the AI automatically adjusts movements to reduce effort and cool down. This thermal management breakthrough has implications for prosthetics and drone performance in confined, hot environments.