PATH

Paper proposes STHTD-MP, a behavior-induced metric Mirror-Prox temporal-difference (TD) algorithm for faster/stabler off-policy value prediction with linear function approximation. Claimed mechanism: using the symmetric part of the behavior-policy Bellman matrix as the metric can improve saddle-point geometry and reduce the mean contraction factor vs GTD2-MP, yielding faster convergence under certain assumptions; Baird’s counterexample is a boundary case where assumptions fail. Investable linkag

The post argues there is a tension between very short AGI timelines and the current industry push to scale reinforcement learning and mid-training on LLMs. If models are close to human-like, self-directed learners, then expensive pre-training/RL environment work for browser use, Excel, financial modeling, robotics tasks, etc. should become unnecessary. If they are not, then AGI is likely not imminent and labs will keep needing costly expert data, verifiable tasks, and task-specific practice. The

Satya Nadella frames AI/AGI as potentially the largest economic shift since the industrial revolution, while emphasizing that the field is still early and that model-only companies may face a winner’s curse because model innovation can be copied or commoditized quickly. He says Microsoft does not want Azure to be merely a host for one AI lab or one model architecture, because infrastructure optimized for a single customer or topology could become obsolete after model-design changes such as MoE b