1X has unveiled the '1X World Model' for its NEO humanoid robot, enabling it to understand the dynamics of the physical world and autonomously learn new tasks. This model leverages internet-scale video data and the robot's own sensor data to generate and execute actions based on instructions, even with unfamiliar objects and environments. This allows robots to gain experience and improve their performance independently, reducing reliance on human-provided data. Demonstrations showed NEO performing novel tasks like pulling tissues from a box, ironing clothes, and brushing hair without prior training, indicating a new phase in autonomous learning for robotics.

Voice AI company Deepgram has secured $130 million in Series C funding, achieving a $1.3 billion valuation and unicorn status. Concurrently, Deepgram acquired OfOne, a Y Combinator-backed startup specializing in voice AI ordering systems for quick-service restaurants with over 93% accuracy. This funding round, driven by surging demand in call centers, media, and AI agents, will fuel Deepgram's global expansion, multilingual capabilities, and focus on the restaurant sector. The acquisition signals consolidation trends in the maturing voice AI market.

As Large Language Models (LLMs) become increasingly complex and opaque, researchers are adopting a biological approach, studying them like living organisms rather than traditional software. This involves observing behavior, tracing internal signals, and mapping functional regions without assuming a neat underlying logic, similar to how biologists study unfamiliar life forms. This 'grown' nature of LLMs, where algorithms adjust billions of parameters, necessitates methods like mechanistic interpretability to understand their internal workings and ensure safety. Researchers are identifying how specific concepts map to distinct regions within the model.