AI Newsletter

Nvidia is reducing its risk in OpenAI's mega-investment by 50%, signaling skepticism about the profitability of AI infrastructure megaprojects – while Anthropic demonstrates with 145% revenue growth in a quarter that AI applications are economically scalable. The divergent development points to a reweighting: away from unlimited CapEx in hardware infrastructure, toward profitable AI products with real demand.

The "tragedy of the commons" phenomenon describes a market conflict: individual companies save costs through AI automation of junior positions, but collectively weaken the knowledge pipeline of their entire industry. The delayed effect (2030–2045) makes early detection difficult, but could lead to skills shortages, declining innovation, and regulatory intervention.

Moonshots' new PerceptionBench test reveals significant weaknesses in visual perception across all leading multimodal AI models – none achieves 60% accuracy. This suggests that reasoning deficits often originate not in logic, but already in image processing. The finding could fundamentally reassess product development and use cases for vision AI systems.

Anthropic is introducing a watermarking API that reliably identifies Claude-generated text, proactively addressing regulatory requirements and market trust demands. The technology differentiates Anthropic in AI competition and addresses growing concerns about AI misuse without compromising model quality – a strategic advantage over competitors like OpenAI.

OpenAI's "Computer History" documents user activities locally and makes them searchable for ChatGPT – a game-changer for productivity and contextual understanding that massively deepens AI integration into daily work. The feature increases switching costs and user retention, but simultaneously opens major data privacy and regulatory risks, as keyboard inputs remain stored locally.