Hyacehila's Blog

Designer training also works for AI Agent builders: see more real systems, solve real problems, and keep asking why.

Chat is one way we package generative models, not the natural shape of the model itself. Interaction models return models to continuous tokens, perception, and action, with JoyAI-VL-Interaction as a concrete vision-language example.

Before LLM agents became popular, game NPCs had already formed a practical agent architecture around perception, state, decision making, and action.

ComfyUI has outgrown the label of an image-generation UI. It is closer to an open generative workflow/runtime, and video generation now exposes what that runtime still needs: segments, inherited assets, candidate sampling, evaluation, and long-running orchestration.

Treating context as a finite resource: from context rot and attention budgets to storage structures, retrieval pipelines, time-aligned writes, version governance, and runtime techniques such as compaction, reset, subagents, and checkpointing.

HTML is tempting. It can turn long AI-generated documents into something readable, clickable, and shareable. But I do not want HTML to quietly become the source of truth. Markdown should keep the diffable, searchable, auditable record; HTML should make complex systems easier to read.

A placeholder structure for future technical teardowns of external memory systems, agent runtimes, and framework-provided memory capabilities.

The cosmic string case and Trae Agent come from different domains, but both ask the same design question: after an LLM generates candidates, how does the system search, verify, prune, and choose?

A Chinese translation and commentary of Addy Osmani’s essay: the default AI coding loop is optimized for closing tasks, not for keeping you sharp.

The future of AI agents will not begin with full autonomy replacing humans, but with governed delegation: humans set goals and boundaries while AI takes on more execution.