People Are Amazed AI Can Think. I'm Annoyed by the First Draft.
Everyone I know outside tech is thrilled by how fast AI is and how much it feels like thinking. This week a lot of them are also scared, because the headlines talk like agents just get up and do things on their own. I hear both, and past a certain amount of hands-on experience, I mostly just disagree.
Read the pieceEarlier
Sensors Still Need a CPU.
No Promises I Can't Take Back.
Claude Code Lapsed. The Reactivate Button Is Right There.
Build the Decision Yourself First. Then You're Allowed to Ask AI.
I Built a Deep Learning Framework From Scratch. Now I Read Every Model Differently.
Your Agent Has an Exactly-Once Problem
When Optimal Is the Wrong Answer
The Model Doesn't Do Anything. Your While Loop Does.
Evaluating AI Agents Like You Ship Software
Designing Retry-Aware Infrastructure
From Prototype to Production
Multi-Cloud Without the Chaos
The Best Testers I Have Shipped To Were Researchers
Coming
- Your Agent Has Amnesia. Here's the Dict That Fakes Memory.Week 2
In the two notebook agents I built, memory was a dict I re-injected each turn. Once the conversation gets long, I have to decide what the agent is allowed to forget.
- The Gap Between a Notebook Agent and a Production One.Week 3
Those two agents ran on mock dictionaries. The next piece is the gap I keep hitting: retries, timeouts, persisted state, and the logs you need once someone else is using it.
- Five Ways to Chunk a Document, and When Each One Bites You.Drafting
Fixed, semantic, recursive, structure-based, LLM-based. Recursive is the default I would start with. The rest of the piece is the failure modes, and when each one is the wrong bet.
No content calendar.