How we think about the work.
Short, specific perspectives on continual learning, agentic systems, and the architecture of AI that gets measurably better at your business over time — written for people deciding what to build and what to buy.
The Open Knowledge Format is the easy part.
Google Cloud just published an open standard for the knowledge AI runs on — the Open Knowledge Format. It validates a bet we shipped a year ago, and clarifies where the real value sits: not the format, the engine that fills and keeps it.
Read →The world is bigger than the model.
Every model ships frozen on the day it trained; the environment it works in does not. Why the gap between a trained snapshot and a moving world is the real case for continual learning — and the deepest version of the argument comes from the person who wrote the book on reinforcement learning.
Read →The number nobody agrees on.
Two reports show different values for the same metric — and both are right. No one ever defined what the metric means, so each team computes its own version. Why it is a definition problem, not a data problem, and why an AI built on top of it picks a side silently.
Read →Your AI is running on stale data.
A model is only as current as the pipeline behind it. When that pipeline is a nightly — or monthly — batch, your AI answers with full confidence about a world that has already moved. Why data freshness is an AI-readiness decision, not a tuning detail.
Read →Who owns data quality?
Everyone touches the data; no one is accountable for whether it is right. The ownership gap is why quality programs stall — and why an AI built on the same data inherits a problem with no owner.
Read →Your team keeps a spreadsheet next to the system.
The official system is the system of record; the shadow spreadsheet beside it is the system of truth. Each one marks where your data isn’t trusted — and why your next AI project inherits the gap.
Read →Your AI demo wowed everyone. It still isn’t in production.
A demo proves the model can work once. Production proves it keeps working — on data it has never seen, under a cost and latency budget, with an answer for when it is wrong. The six-part gap, and the two stages teams skip.
Read →Your lakehouse migration keeps slipping.
The platform works; the old warehouse is still on, still paid for, still trusted. Why migrations stall at eighty percent — and the staged, workload-by-workload path that actually finishes one.
Read →Your CFO stopped trusting the dashboard.
The dashboard is rarely the problem — the data behind the glass is. The ten cracks that quietly erode data trust, and the fastest, predictable way to earn it back.
Read →Your data is not ready for AI yet.
The model is the easy part. Six questions that decide whether an AI initiative survives real users — and the order to fix the foundation underneath it.
Read →The repo is not the memory
When a project ends or a person leaves, you keep the artifacts and lose the reasoning. Why institutional memory is the journey, not the destination — and what it means to own it.
Read →Your best evals are your failure traces
The test you wrote measures what you guessed an AI would get wrong. The more valuable signal is already in your logs — what it actually got wrong. Why the systems worth owning remember their own mistakes.
Read →Retrieval is not discovery
Three things people mean when they say AI “learns” your business — retrieval, search, and discovery. Most systems stop at the easiest one; only the last compounds.
Read →You don’t buy a tool. You buy the work.
The most valuable AI companies aren’t selling software to do the job — they’re doing the job. Why the copilot/autopilot distinction decides what you should pay for, and what we sell.
Read →Proof of concept is not production
OpenClaw showed the world that AI agents work. Running one inside a business — safely, with memory that lasts and a record you can audit — is a different engineering problem.
Read →Architectural privacy, not contractual privacy
Every AI vendor promises your data is safe. The promise is a contract — something you have to trust. There’s a stronger version: one your own network can verify.
Read →Two kinds of world models
Spatial intelligence and continual learning share a name and a diagram. They are not the same bet — and knowing which one you are buying changes everything.
Read →