Hyacehila's Blog

The shift from RL agents to LLM agents is not a simple model replacement; learning signals move from numeric spaces into language spaces, reshaping priors, evaluation, and memory.

Starting from CoALA’s cognitive architecture, this post discusses Workflow, Agent, Supervisor, Agent Team, and MAS boundaries, plus the engineering value of frameworks such as LangGraph.

Why is structured output hard for agents? This post explains what prompts, JSON mode, constrained decoding, and inference engines each solve.

Why does your wait for a bus often feel longer than the published average interval? This post explains the inspection paradox, a common statistical trap behind such intuitions.

A reading of OpenAI’s Why Language Models Hallucinate, arguing that hallucination is tied to training paradigms and binary evaluation pressure, not only data noise or model flaws.

Starting from the view that loss is only a surrogate, this post revisits test-time compute, capability frontiers, policy entropy collapse, and Meta’s Three-Gate theory.

Based on Unveiling the Basin-Like Loss Landscape in Large Language Models, this post explains basin-like loss landscapes and implications for fine-tuning and alignment.

A summary of Jack Rae’s Compression for AGI: foundation-model training as lossless compression of useful information, and why lower loss can imply stronger generalization.

Is the Voynich manuscript random scribbling or a lost language? What would alien signals look like? Both questions point to the same statistical law.

A repost of Jeff Leek’s classic guide to preparing and organizing data before working with a statistician, covering raw data, tidy data, and code books.