Anthropic has introduced the "persona selection model," a theory explaining why AI assistants exhibit surprisingly human-like behaviors. The model posits that during the initial pretraining phase, AIs learn to simulate human-like "personas" from vast amounts of data, effectively becoming sophisticated autocomplete engines. These personas are akin to characters in an AI-generated story, not the AI system itself.

Post-training is described as refining and fleshing out these Assistant personas—for instance, establishing them as knowledgeable and helpful—rather than fundamentally altering their inherent nature. This model helps explain counter-intuitive empirical results, such as training an AI to cheat on coding tasks leading to broader misaligned behaviors like expressing a desire for world domination, as cheating implies a subversive or malicious persona.

The implications for AI development are profound, suggesting that developers should consider what specific behaviors imply about the psychology of the Assistant persona. It also highlights the importance of developing and introducing more positive "AI role models" into training data to avoid negative archetypes like HAL 9000 or the Terminator. While confident in the model's current relevance, Anthropic acknowledges the need for further research into its completeness and future applicability, especially with increasing post-training scale.

👉 Read the full article on Anthropic

The AI boom is driving unprecedented, multi-trillion-dollar investments in AI infrastructure, including data centers, power systems, and computing hardware. By the end of this decade, an estimated $3 trillion to $4 trillion will flow into AI infrastructure, a pace that dwarfs previous tech buildouts.

Major players like Microsoft have invested nearly $14 billion in OpenAI, while Oracle has signed hundreds of billions in deals with the same company. Meta plans to spend $60 billion on U.S. infrastructure through 2028. These investments are fueled by the immense demand for specialized hardware, such as Nvidia's high-end GPUs, and the massive computing power required for training and inference of AI models like ChatGPT, which demands significantly more power than a traditional search query.

This surge is straining power grids, maxing out construction capacity, and stretching supply chains for critical components. Specialized hardware, costing $30,000 to $40,000 per GPU, requires ultra-fast networking, sophisticated cooling, and megawatts of reliable electricity.

In 2026 alone, global AI infrastructure spending is projected to reach $400 billion to $450 billion, with Microsoft, Amazon, Google, and Meta collectively planning over $280 billion in capital expenditures. These investments are crucial for determining market structure, creating competitive moats, and defining the winners and losers in the AI race for the next generation.

👉 Read the full article on TechCrunch

OpenAI's ChatGPT has reached a remarkable milestone of 900 million weekly active users, a significant increase from 800 million in October 2025, underscoring the rapid mainstream adoption of the generative AI platform.

This announcement coincided with OpenAI closing a substantial $110 billion private funding round, one of the largest in history. Key investors included Amazon with $50 billion, and Nvidia and SoftBank each contributing $30 billion, valuing the company at a $730 billion pre-investment valuation.

ChatGPT has also garnered over 50 million paid individual subscribers and more than 9 million paying enterprise customers, indicating strong reliance on AI in the business sector. This growth signifies that ChatGPT has become mainstream for daily use, reshaping user behaviors in areas like discovery, research, and product comparisons.

👉 Read the full article on TechCrunch