[AINews] Muse Spark 1.3 matches GPT-5.6-Sol, confirming Meta Superintelligence as the newest Frontier Lab, >90% discount for training
Meta released the Muse Spark 1.3 model, which ranks 3rd in the world according to AAII data. Its performance is comparable to OpenAI Opus and Anthropic Fable, and it promises open weights. This model offers a training discount of over 90% (8/22-8/24). Stanford University has established AI-native software engineering as an official discipline, launching courses centered on “2026 Transformation,” with 85% of the content focusing on agent skills, context engineering, and agent code review; the CS329Z course focuses on building agents from scratch using first-principles methods. The industry consensus is shifting from simple routing to state-aware agent allocation, and some neutral startups are surpassing leading laboratories in narrow tasks by optimizing their execution environments. Rumors about the “loop Transformer” architecture of OpenAI Astra are interpreted as modest adjustments, where the number of reuse layers doubles the computational capacity without changing storage requirements, rather than a breakthrough architecture; its reasoning process does not necessarily hide…