Frequently Asked
Questions

Simple: AI will change most jobs and replace some tasks — but "tasks" and "jobs" aren't the same thing.

Deep: The evidence points to task-level automation rather than job-level replacement in most cases — a job is a bundle of tasks, and AI rarely automates an entire bundle. Exposure is highest where work is routine, text-based, and verifiable, and lowest where it requires physical dexterity, long-horizon judgment, or accountability that must reside with a person. Historical patterns suggest recomposition: roles shift toward oversight, review, and integration. The honest answer is that the distribution of impact will be uneven — some fields genuinely disrupted, others mildly augmented — and the most defensible position is being the person who directs the tool rather than competes with it.

Simple: Trust it the way you'd trust a confident stranger — useful, but verify what matters.

Deep: The mechanism to internalize is calibration: models are trained to be fluent, not to be right, and confidence in the writing carries no information about factual accuracy. Errors cluster where the model has no grounding — recent events, niche specifics, precise numbers, and citations. A practical hierarchy: verify anything consequential, specific, or unfamiliar; treat general reasoning and drafting as lower-risk. The layered defenses are the same ones professionals use — retrieval for facts, tool calls for computation, citations for claims, and a human decision on anything irreversible. Trust is not binary; it's task-dependent.

Simple: Probably not forever — progress is real but likely to be uneven, not a straight line up.

Deep: Three forces shape the curve. Scaling has delivered predictable gains, but it faces physical constraints — training data exhaustion, power and cooling limits, and capital intensity. Wringing more capability may increasingly come from test-time compute, better reinforcement learning, and architectural change rather than raw size. Capability is not monotonic across dimensions: a model can get dramatically better at coding while barely moving on long-horizon planning. And deployment lags capability — what reaches users is filtered by cost, latency, and safety. So expect continued rapid change in specific, unevenly distributed capabilities, not a uniform march upward.

Simple: Pick one real task you already do, and use AI for that — not a random demo.

Deep: The effective learning path is task-anchored, not tool-anchored. Choose one recurring task with a clear quality bar — drafting email, summarizing documents, explaining unfamiliar code — and iterate on it until you can reliably get good output. This builds something demos don't: calibration about where the tool is strong and where it quietly fails. Then stack skills in the order that returns the most value: prompt specificity, then how to give examples, then when to demand verification, then when to reach for retrieval or tools. The order matters because each layer assumes the previous — and because the goal is judgment, not tricks.