Learning AI is not a hobby. It is a professional discipline.
Every customer conversation now ends at the same question: “What does this mean for us, and what should we do first?” A pre-sales architect who cannot answer honestly is a brochure with a pulse. So I study — deliberately, continuously, in public.
Why “continuous” is the whole point
The half-life of a useful AI fact is measured in months. A discipline is not something you finish; it is a schedule you keep. Mine looks less like a reading list and more like an architecture review: recurring, evidence-driven, and honest about what it does not cover.
The practice: one foundational text at a time, one hands-on lab per quarter, one model actually run on real hardware per season. Reading about inference is cheap; sizing a GPU for a customer's workload and watching the tokens-per-second is what converts opinion into counsel.
What AI actually is — the honest short version
Pattern machinery. Statistical systems trained on so much data that fluent behavior falls out of them. Powerful, genuinely useful, and still narrower than the hype — which is precisely why it needs architects. Someone has to decide where the machine's brilliance ends and the failure begins, and design the estate so the business survives that boundary.
Myths I retire in every conversation
- “AI is a product you buy.” It is a capability you operate — data, governance, skills and refresh cycles included. The vendor demo is page one, not the book.
- “More parameters solve it.” A small model on your own clean data routinely outclasses a giant one on your own messy data. Architecture beats scale.
- “We can wait and see.” Reasonable — until you notice the compounding. The learning curve is steepest at the start, and every quarter of waiting costs more than the quarter of studying.
- “It will replace the architect.” It will replace the architect who does not use it. The judgment layer — what to build, what to refuse — is stubbornly human.
The standing curriculum
Foundations: how transformers actually work, without the hand-waving. Data: pipelines, quality, and the unglamorous governance that decides whether any of it matters. Operations: serving, latency, cost per token, and where on-premise still beats cloud. Application: retrieval, agents, and knowing which of the two your customer truly needs. And always the geo thread — satellite imagery and location intelligence are quietly becoming among the best-proven applied ML there is.
If you are learning too: keep the schedule, not the resolution. Discipline outlasts motivation.
“The heart of the discerning acquires knowledge, for the ears of the wise seek it out.” Proverbs 18:15
Keep walking
The discipline of learning rests on a deeper conviction about who we are answerable to.