My latest writings.
MCP is USB-C for AI tools: one standard plug collapses the M×N integration explosion into M+N. Written slowly, one step at a time.
A classifier that tells a cockroach from a date never learns what either is — it learns a boundary between patterns of pixels. Everything confusing about these systems falls out of that distinction.
Is an AI Engineer just a rebranded ML Engineer? The boundary that cleared it for me: AI Engineers apply pre-trained models to real problems; training new ones from scratch is someone else's job.
The next-token trick, unpacked. An LLM is a trained statistical engine: text becomes tokens, a transformer reads them, and the model samples the next token — again and again.
Temperature, top-p, max tokens: each knob looks like an isolated setting. The frame that collapsed them: every generation parameter reshapes the same next-token probability distribution before you sample.
'Writing good prompts' bundles two different skills. The split that mattered: prompt engineering crafts the instructions, context engineering builds the whole information environment around the model.
Few-shot, ReAct, chain-of-thought, caching: unrelated tricks, or one pattern? Each one decides what goes in the context window, in what shape, and for how long.
'Open source good, closed source bad' breaks the moment you pick a real model. The three independent axes that actually decide: open/closed, pre-trained/from-scratch, self-hosted/cloud.
'Just call OpenAI' stops being an answer when the choice actually matters. Model selection is multi-axis — capability, cost, context, modality — and the provider is a separate decision.