Loading
Writing
Deeper dives into AI systems, multi-agent design, and building software you can actually trust.
Same 3D objects, four ways to store them, one laptop. The theoretical cost metric, FLOPs, ordered the methods one way. Real training time ordered them backwards. The densest representation self-throttled and finished last by a factor of ten. Here is how to read between the lines when theory meets silicon.
When code can be generated and even executed for you, the leverage moves. You stop verifying every line and start engineering systems that can't fail silently — an idea older than AI.
Why one agent can't reliably build a pipeline end-to-end — and how splitting the job into three roles with the wrong tools for each other's work makes the result actually checkable.
Getting a language model to talk is easy. Getting its output trustworthy enough to persist is the whole job. Four techniques from a production LLM app — and the one principle underneath all of them.
Newsletter
Get notified when I publish something new. No spam, unsubscribe anytime.