What is AI (and what it is not)

Why this article exists

Andrew Ng’s AI for Everyone starts with terminology and limits. ShipAI engineers still need that bridge: stakeholders say “AI” when they mean a spreadsheet rule, a classifier, or ChatGPT. This page is the shared vocabulary before tokens and transformers.

Misnaming the problem causes bad architecture: teams buy an LLM for a deterministic eligibility rule, or train a custom model when a prompt + retrieval would ship in a week.

Three nested ideas

flowchart TD
  AI[Artificial intelligence] --> ML[Machine learning]
  ML --> DL[Deep learning]
  DL --> Gen[Generative AI / LLMs]

Term	Plain meaning	Typical artifact
AI	Systems that perform tasks that look intelligent	Broad product label
ML	Learn patterns from data instead of hand-coded rules	Classifier, ranker, forecaster
Deep learning	ML with multi-layer neural nets	Image net, speech model, LLM
Generative AI	Models that produce new text/images/code/audio	Chatbots, copilots, image tools

Ship rule: Prefer the most specific true label. Calling everything “AI” hides whether you need labeled data, a retrieval index, or just better UX.

Narrow vs general (product language)

Phrase	Useful meaning	Trap
Narrow AI	Good at a scoped task (spam, ranking, extract fields)	Sounds “small” — can still be critical path
General / frontier models	Broad skills via large pretrained LLMs	Not automatically correct on your facts
AGI	Speculative human-level generality	Not a shipping milestone for your roadmap

Build products with narrow success metrics, even when the engine is a general model.

What today’s LLMs are good at

• Drafting, summarizing, transforming text and code
• Tool-using agents when tools are reliable and scoped (Agents and ReAct)
• Retrieval-augmented Q&A over your docs (with evals) (RAG building blocks)
• Structured extraction when schemas are validated (Structured outputs)
• Classification / routing when a golden set says they beat a baseline

What they are bad at (without extra systems)

• Guaranteeing facts without retrieval or tools
• Exact arithmetic / inventory counts (use code tools)
• Fresh private knowledge not in context or an index
• Decisions that require auditability and zero hallucination risk alone
• Long-horizon autonomy without stop policies and HITL

flowchart TD
  Job[Job to be done] --> Q1{Need creativity / language?}
  Q1 -->|No| Rules[Rules / SQL / classic ML]
  Q1 -->|Yes| Q2{Need private facts?}
  Q2 -->|Yes| RAG[RAG / tools]
  Q2 -->|No| LLM[LLM prompt]
  RAG --> Eval[Eval gate]
  LLM --> Eval
  Rules --> Ship[Ship]
  Eval --> Ship

Supervised, unsupervised, reinforcement (one slide)

• Supervised — learn from input→label pairs (spam / not spam)
• Unsupervised — structure without labels (clustering, embeddings)
• Reinforcement — learn from rewards (RLHF for preference alignment)

LLMs use all three across their lifecycle: unsupervised-ish pretrain on text, supervised fine-tunes, preference/RL stages. Product engineers mostly touch the last mile: prompts, RAG, light fine-tunes, and evals — see LLM project lifecycle.

When not to use an LLM

Situation	Prefer
Deterministic policy (“if amount > X”)	Rules engine
Stable labels + tabular features	Classical ML (Classical ML literacy)
Exact counts / money math	Code / DB aggregates
Need citations to private docs	RAG + citations, not “just ChatGPT”
Regulated decision with hard audit	Human + constrained tools; not free-form generation alone

Failure modes of “AI” projects

Failure	What it looks like	Fix direction
Solution looking for a problem	Cool demo, no metric	Write success metric first
Wrong tool	LLM for eligibility rules	Rules / workflow
No evals	“Looks good in Slack”	Golden set + gate
Data denial	Want magic without corpus or labels	Scope to prompt-only or collect data
Overclaim	Marketing says “accurate”	Measure faithfulness / task success

Glossary

Term	Meaning
Model	Learned function mapping inputs → outputs
Inference	Running a trained model on new inputs
Hallucination	Fluent output not grounded in facts/evidence
Baseline	Simple system you must beat on metric + cost

Micro-project

List three work problems that are ML, three that are rules, and three that should not use AI. For one ML candidate, write a one-line success metric.

Related guided path

What happened in AI (till now) and ML/DL literacy. Next Core Concepts: Classical ML literacy, then Tokenization.