https://www.boshengding.com/blog Essays by Bosheng Ding on language-model post-training, RLHF, evaluation, and human-centered AI. en https://www.boshengding.com/blog/decisions-not-strings-what-typesafes-jev-gets-right https://www.boshengding.com/blog/decisions-not-strings-what-typesafes-jev-gets-right Tue, 29 Sep 2026 12:00:00 GMT A critical reading of TypeSafe's System One model: its typed decision interface, calibration thesis, launch evidence, known failure modes, and early independent audit. https://www.boshengding.com/blog/when-ai-grows-gdp-faster-than-wages https://www.boshengding.com/blog/when-ai-grows-gdp-faster-than-wages Tue, 29 Sep 2026 12:00:00 GMT Anthropic's economic scenarios show that AI capability, adoption, automation, and worker mobility can produce very different distributions of growth by 2030. https://www.boshengding.com/blog/helpful-is-not-the-same-as-safe-mentalhealthbench https://www.boshengding.com/blog/helpful-is-not-the-same-as-safe-mentalhealthbench Thu, 24 Sep 2026 12:00:00 GMT MentalHealthBench shows why high-stakes conversational evaluation needs expert guardrails, user perspectives, behavioral decomposition, and restraint about leaderboards. https://www.boshengding.com/blog/evaluate-post-trained-models-without-fooling-yourself https://www.boshengding.com/blog/evaluate-post-trained-models-without-fooling-yourself Fri, 18 Sep 2026 12:00:00 GMT A layered evaluation system for post-trained models that controls prompt effects, judge bias, contamination, and the gap between benchmark gains and product quality. https://www.boshengding.com/blog/kimi-k3-state-management https://www.boshengding.com/blog/kimi-k3-state-management Wed, 02 Sep 2026 12:00:00 GMT Kimi K3 shows why million-token agentic reinforcement learning depends on persistent state, resumable environments, and deployment-aware training. https://www.boshengding.com/blog/kimi-k2-5-context-management https://www.boshengding.com/blog/kimi-k2-5-context-management Sat, 08 Aug 2026 12:00:00 GMT Kimi 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. https://www.boshengding.com/blog/dpo-or-ppo-choose-by-data-not-fashion https://www.boshengding.com/blog/dpo-or-ppo-choose-by-data-not-fashion Sat, 01 Aug 2026 12:00:00 GMT A practical framework for choosing between offline preference optimization and online reinforcement learning based on data freshness, feedback quality, infrastructure, and failure modes. https://www.boshengding.com/blog/why-online-learning-still-matters https://www.boshengding.com/blog/why-online-learning-still-matters Thu, 11 Jun 2026 12:00:00 GMT Offline 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. https://www.boshengding.com/blog/rl-is-becoming-local-composer-2-5 https://www.boshengding.com/blog/rl-is-becoming-local-composer-2-5 Wed, 20 May 2026 12:00:00 GMT Cursor's Composer 2.5 suggests that improving long-horizon agents requires localized credit assignment, adaptive task generation, and training-systems co-design. https://www.boshengding.com/blog/reward-models-useful-proxies-dangerous-targets https://www.boshengding.com/blog/reward-models-useful-proxies-dangerous-targets Fri, 24 Apr 2026 12:00:00 GMT Reward 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. https://www.boshengding.com/blog/openclaw-in-march-2026-the-personal-agent-is-becoming-a-control-plane https://www.boshengding.com/blog/openclaw-in-march-2026-the-personal-agent-is-becoming-a-control-plane Mon, 30 Mar 2026 12:00:00 GMT A time-bounded look at OpenClaw as local-first agent infrastructure, and why permissions, routing, recovery, and operational discipline matter more than another chat interface. https://www.boshengding.com/blog/preference-data-is-product-design https://www.boshengding.com/blog/preference-data-is-product-design Mon, 02 Mar 2026 12:00:00 GMT The way teams collect and interpret preferences quietly defines what their AI product values—and which users it serves. https://www.boshengding.com/blog/when-measurement-becomes-cheap-research-design-becomes-the-bottleneck https://www.boshengding.com/blog/when-measurement-becomes-cheap-research-design-becomes-the-bottleneck Sun, 15 Feb 2026 12:00:00 GMT OpenAI's GABRIEL shows how language models can scale qualitative measurement, while shifting the hard problem from labeling cost to construct validity and research design. https://www.boshengding.com/blog/kimi-k2-executable-environments https://www.boshengding.com/blog/kimi-k2-executable-environments Wed, 04 Feb 2026 12:00:00 GMT Kimi K2 suggests that agentic post-training is increasingly an environment-engineering problem, not merely a choice of reinforcement-learning objective. https://www.boshengding.com/blog/post-training-is-the-product-layer https://www.boshengding.com/blog/post-training-is-the-product-layer Sun, 18 Jan 2026 12:00:00 GMT Pretraining creates a model's range of possible behavior; post-training decides which behaviors users actually encounter.