Summary
In this Happenings in AI update, sentdex explores an unconventional way of comparing large language models: pitting them against each other in old-school turn-based strategy games. He shows DeepSeek V4.1 Flash, GLM 5.3 Flash, and GPT-6 Astra facing off in games like Hivelight and StarCraft II, arguing that watching models play reveals capability gaps that benchmark leaderboards often flatten out.
The video also digs into recursive self-improvement (RSI), reacting to a detailed technical write-up on how GLM 5.3 Flash was developed using LLMs to accelerate internal research and development. Sentdex pushes back on alarmist framings of RSI, arguing it reflects faster iteration cycles and rising human workload rather than humans being removed from the loop.
The conversation turns pointed when discussing AI safety messaging, with sentdex criticizing coordinated fear-based narratives tied to a recent Wall Street Journal story and figures like Jacob Coxin, and voicing skepticism toward parts of the effective altruism movement’s influence over frontier AI safety labs. He also recounts his own experience red-teaming an early Anthropic model. Viewers interested in unconventional model evaluation methods, RSI debates, and candid industry commentary on AI safety politics will find plenty to chew on.
📺 Source: sentdex · Published September 19, 2026
🏷️ Format: News Analysis
