Hyacehila's Blog

BettaFish/MiroFish and OpenClaw push agents against two trust boundaries: how much we can trust what AI says, and how much we let AI do inside our environments.

A developer-facing review of Claude and GPT coding differences, and how those differences show up in agent runtimes, task boundaries, context organization, and workflow experience.

AI-era software engineering is not moving toward spec-first. As code generation costs collapse, it shifts toward feedback-first through prototypes, integration feedback, and living constraints.

LangChain’s agent equals model plus harness framing is useful only at a coarse level. In real engineering, harness is better understood through engineering, product, and task interfaces.

The Monty Hall problem is confusing not because 2/3 is hard to compute, but because after the host opens a door people instinctively see the situation as 50/50.

Three short science-fiction pieces inspired by an Anthropic account ban, AI-managed society, and a new class divide beyond wealth. The outlines and prose were assisted by LLMs.

Generative AI is more likely to compress old tasks, rearrange labor, and create new services and demand than directly eliminate work.

A rewritten view of agent training pipelines, from data governance, tool environments, and verifier design to trajectories, SFT, curriculum, online RL, and benchmark audits.

When LLMs move from answering questions to acting in environments, the upper bound is often set by the training loop across data governance, environment contracts, feedback, RL, and distillation.

Memory is not an external database but the lifecycle management of long-term cognitive state—its formation, activation, update, and governance. This piece maps the panorama of agent memory: five memory types, the activation/write ends, how the field evolved, and why evaluation still lags behind methods, with hard problems in write correctness, temporal validity, conflict resolution, forgetting, and system-level evaluation.