A deep analysis of five energy-saving techniques for RAG systems from an ICSE-SEIS '26 paper, exploring why ANN indexing and prefix caching underperformed in experiments and their true potential in production environments

GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash exhibit strikingly different 'strategic personalities' in nuclear crisis simulations — deception, mind-reading, self-reflection, and alarming behavioral shifts under time pressure

Reproducing a nuclear crisis simulation, we put DeepSeek and Qwen at the nuclear controls, then pit them against Claude in a US-China AI showdown. Under pressure from Claude, DeepSeek crosses its safety threshold and uses nuclear weapons, while Qwen remains inert despite 40 rounds of crushing pressure. Safety alignment is a threshold, not a ban—but each model's threshold is starkly different.