Notes on post-training, RLHF, evaluation, and human-centered AI.
A critical reading of TypeSafe's System One model: its typed decision interface, calibration thesis, launch evidence, known failure modes, and early independent audit.
Read essayAnthropic's economic scenarios show that AI capability, adoption, automation, and worker mobility can produce very different distributions of growth by 2030.
Read essayMentalHealthBench shows why high-stakes conversational evaluation needs expert guardrails, user perspectives, behavioral decomposition, and restraint about leaderboards.
Read essayA layered evaluation system for post-trained models that controls prompt effects, judge bias, contamination, and the gap between benchmark gains and product quality.
Read essayKimi K3 shows why million-token agentic reinforcement learning depends on persistent state, resumable environments, and deployment-aware training.
Read essayKimi K2.5's Agent Swarm is most useful when viewed as learned context sharding rather than simply a way to run more agents in parallel.
Read essayA practical framework for choosing between offline preference optimization and online reinforcement learning based on data freshness, feedback quality, infrastructure, and failure modes.
Read essayOffline preference fitting can teach a model from yesterday’s comparisons. Online learning becomes valuable when the model must learn from the distribution its own improving policy creates.
Read essayCursor's Composer 2.5 suggests that improving long-horizon agents requires localized credit assignment, adaptive task generation, and training-systems co-design.
Read essayReward models turn subjective comparisons into a training signal, but optimizing their score too aggressively can expose the gap between what we measure and what we actually want.
Read essayA time-bounded look at OpenClaw as local-first agent infrastructure, and why permissions, routing, recovery, and operational discipline matter more than another chat interface.
Read essayThe way teams collect and interpret preferences quietly defines what their AI product values—and which users it serves.
Read essayOpenAI's GABRIEL shows how language models can scale qualitative measurement, while shifting the hard problem from labeling cost to construct validity and research design.
Read essayKimi K2 suggests that agentic post-training is increasingly an environment-engineering problem, not merely a choice of reinforcement-learning objective.
Read essayPretraining creates a model's range of possible behavior; post-training decides which behaviors users actually encounter.
Read essay
