Hyacehila's Blog

A Zhihu answer offers a useful reminder: in large-model engineering, clever structures do not always beat simple ones that keep GPUs full and follow matrix-multiplication pipelines.

Word Salad Chopper does more than trim repetitive reasoning text. It suggests that hidden states at newline-boundary tokens may offer a cheap window into generation patterns.

Output tokens are expensive mainly because decoding is serial, KV cache consumes memory, and generation occupies scheduler slots. Agent cost optimization requires controlling output budgets.

Model routing breaks every agent call into cost, latency, tool stability, and escalation decisions. The point is not picking the strongest model, but allocating budget.

An agent infrastructure overview that maps the backend capabilities needed for AI agents to move from demos to large-scale services and the engineering problems each solves.

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.