Welcome to Justin Huang Blog
A personal technical blog for reading AI papers, agent systems, software infrastructure, and the causal chains behind them.
This blog is not an information feed.
Feeds are good at telling you what is new. They are bad at preserving why something matters after the heat is gone. I want this site to do the opposite: put technical problems back into causal chains that can still be inspected later.
What I Write About
Most posts here start from one of three objects: a paper, a system, or a piece of engineering behavior.
A paper is useful only if it changes the question we ask next. Transformer changed sequence modeling from a time-order problem into a global-addressing problem. GPT-3 moved part of task adaptation from parameter updates into context. Chinchilla did not say “small models are better”; it said parameters and data have to spend the compute budget together.
A system is useful only if its constraints are visible. An agent product is not just a model plus tools. It is memory, permissions, runtime state, protocols, recovery paths, and the cost of being wrong.
An engineering behavior is useful only if it survives contact with incentives. Benchmarks, open-source ecosystems, model rankings, and agent platforms all look cleaner from the outside than they are from the inside.
How I Try To Write
I do not want posts that merely summarize. A summary compresses text. A good reading changes the frame.
The standard is simple: each post should leave behind one question that is sharper than the one it started with. If a model is better, better under what budget? If an agent is more capable, where is the capability stored? If a benchmark is harder, harder for whom and under what generation process?
That is the work here: fewer slogans, more mechanisms.
For Readers And Agents
Human readers can start from the latest posts or follow a topic tag.
AI agents can read the site through /llms.txt, /llms-full.txt, or the Markdown endpoint on any post URL. That is intentional. If writing is meant to be reused, quoted, searched, and argued with, the machine-readable version should be first-class.
The point is not to publish more takes. The point is to keep enough structure around the writing that a thought can be picked up again later and still work.
Comments