The morning briefing.

AI Talent & Infrastructure. Anthropic reportedly poached OpenAI's second-ever chip engineer, Clive Chan, intensifying the talent war as both companies race towards IPOs and in-house chip development. This underscores the critical need for specialized hardware expertise in the frontier AI race. NVIDIA also introduced its RTX Spark superchip, aiming to reinvent Windows PCs for personal AI agents. In China, OSChina completed a joint-stock reform, positioning itself as a leading open-source AI stock.

AI Security & Societal Impact. OpenAI introduced "Lockdown Mode" for ChatGPT, disabling web access to mitigate prompt injection attacks and protect sensitive data. This comes amidst growing concerns about AI's misuse and security, including reports of anti-AI extremism. Meta confirmed thousands of Instagram accounts were hacked by exploiting its AI chatbot, highlighting persistent vulnerabilities. The UK exam watchdog also expressed concerns about smart glasses and AI tools enabling new forms of cheating in exams.

Enterprise AI Adoption. Tencent has officially entered the enterprise AI arena, focusing on engineering applications rather than just model capability, as demonstrated at its 2026 Cloud AI Industry Application Conference. This strategic move by a major player signals a maturing market where practical implementation is key. However, successful AI integration is not always smooth. Starbucks recently abandoned its AI inventory tool after only nine months due to multiple errors, illustrating significant deployment challenges.

AI Research & Efficiency. New research explains why larger language models acquire skills smaller ones miss, suggesting frequent tasks can overwrite learned knowledge. The study proposes that increasing target task frequency in training data could be as effective as scaling up models. Separately, Perplexity's "Search as Code" architecture allows AI models to write their own search routines, improving benchmarks and cutting token costs by up to 85 percent.