A handbook
AI is bigger than ChatGPT.
For executives, founders, and product leaders who decide about AI. LLMs are one branch. This handbook maps the rest.
When teams say “let’s use AI for X” they usually mean “let’s use an LLM for X.” Often the right tool is a different family of techniques entirely. Picking the wrong family burns six to twelve months of runway before anyone notices.
One page per module. Read it in 25 minutes. Keep it as a reference.
The four families
Classical ML
Tables, fraud, churn, forecasting. The dominant family in production AI. Where XGBoost beats LLMs on tabular data.
Family 2Specialist Deep Learning
Vision, time series, recommenders, speech, anomalies. Where pre-LLM specialists still dominate on real-time and edge.
Family 3Foundation Models
LLMs, VLMs, diffusion. Adaptation via prompting, RAG, fine-tuning. Agents, reasoning, MCP. The branch everyone talks about.
Family 4Reinforcement Learning
Sequential decisions with reward signals. Trading, robotics, control. Rare in production, irreplaceable when it fits.
Cross-cutting
Which family fits?
The decision rule, in one page. Five questions to answer before you let your team build.
ProductionWhy most AI projects fail
95% never ship. The reasons are predictable. Eval, trace, loop. The production triad every serious team has.
ReferenceGlossary
The 30 terms an executive will encounter in 2026. RAG, MCP, agents, reasoning, fine-tune, distillation, and the rest.
The throughline
AI is a family of techniques. LLMs are one branch. The most expensive mistake is treating them as the whole tree.
