Brief

Mixture-of-Experts (MoE) architectures are often presented as an efficient solution for scaling large language models, but this analysis argues they are primarily a workaround for training instability in dense transformers. The author contends that the emergent modularity seen in MoEs is a symptom of destructive gradient interference in massive dense models, rather than an inherent architectural advantage. While MoEs can offer efficiency and capacity, they introduce significant debugging complexity and can lead to unpredictable performance when real-world usage deviates from training data, suggesting a need for fundamental research into training dense models without interference. AI

A pediatric surgeon and researcher hypothesizes that artificial intelligence is eroding the self-correction mechanisms of science, a phenomenon they term "epistemic immunodepression." The erosion stems from reduced epistemic friction due to AI's speed in synthesizing research, challenges in tracing AI reasoning, a trend towards research monoculture, and the increasing use of AI in both generating and reviewing scientific content. Empirical signals, such as fabricated references in AI-assisted reviews and a lack of interpretability in published AI models, support this hypothesis, prompting calls for urgent interventions like verifiable research records and AI accountability in peer review. AI