persistent scheduler for agent runtimes
A local CLI for schedules that should outlive a runtime process: reminders, lifecycle commands, logs, and fire events.
Built because hidden timers are hard to trust.
Work. Open projects, tools, and research artifacts.
swarm harness
An executable contract harness for agent coordination: message semantics, task ownership, routing, runtime prompts, and safety gates.
Fixture checks and runtime verifiers make the protocol reviewable.
liquid glass lab
A small frontend material lab for transparent glass: SVG displacement, edge refraction, chromatic fringe, and local showcase pages.
The repo keeps the material parameters and showcase pages visible for comparison.
state representation learning for multi-agent interactions
A class research project comparing memory architectures for long-term multi-agent interaction.
Graph, vector, and semantic-memory baselines around interaction tasks.
review: a mechanistic understanding of DPO and toxicity
A technical review of the 2024 paper on how Direct Preference Optimization reduces toxic generations through activation steering rather than capability erasure.
Walks the original paper's mechanistic claims, limitations, and broader alignment implications.
attackable problems
A shortlist of open problems worth giving to frontier models: each one has a clear verifier, a first-week attack plan, and a reason to stop.
Selection rules, first-week attack plans, and stop rules make each item actionable.
archaeologist report on emergent abilities of LLMs
A report tracing the emergent-abilities debate through three phases: the initial discovery, the 2023 mirage critique, and the recent refinements that followed.
Built around the influential paper Are Emergent Abilities of Large Language Models a Mirage?.
wellspring
A single-page interactive piece where old poems behave like a water surface: ambient idle waves, ripples that react to scroll.
An alt interface for blog.diangao.space.
verified build harness
A scaffold for builds where agents can write code while the repo keeps the authority to say done: separate lanes, checks, reviewer handoff.
Tactiles came out of this pattern.
artifact style kit
Utilities for turning a visual reference into an inspectable style loop: asset collection, contact sheets, prompt notes, and candidate review.
The taste check lives in visible source files and review artifacts.
SPARK
A small proactive accountability agent in telegram, turning self-reminders into visible, recurring nudges.
A personal-tools artifact from the same local-first agent-tool thread.
think OS
A long-running agent workspace: exploration with versatile coding agents and their general usage.
Mostly boring on purpose.