ScanMap.AI

AI Primer

How LLMs Evolved

From a sealed room to a team of agents — a plain-language timeline of the last six years in AI, designed for non-technical readers.

2020GPT-3The sealed room
PromptRaw text inputFrozen LLMNo tools, no internetResponseMay be outdated or wrong

“A student who memorized millions of books, now sitting in a sealed room with no phone, no notes, no internet.”

2022ChatGPT / InstructGPTLearns to follow instructions
PromptNatural questionsFrozen LLM+ RLHF alignmentHelpful responseFollows instructions

“Same sealed room, but now the student had a tutor who taught them how to answer questions helpfully instead of just completing sentences.”

2023GPT-4, Claude 1–2Bigger brain, still sealed
PromptText + imagesBigger frozen LLM128K context windowBetter responseStill can’t verify facts

“The student now has a bigger brain and can read images, but the room is still sealed. Smarter, yet still can’t check anything.”

2024RAG, search, tools, MCPThe room opens up
PromptUser’s questionLLM + reasoningSearches, retrieves, calls toolsGrounded responseVerified, cited, actionable

“The student now has a filing cabinet, a phone to Google things, and can make calls to book meetings or send emails. And they must show their work.”

2025–26Agentic AI, multi-agentA whole team with a supervisor
GoalOrchestrator+ guardrails + human approvalResearcherWriterReviewerReview + approveResult

“No longer one student — it’s a team of specialists, each with their own tools and filing cabinets, coordinated by a manager who needs your approval for big decisions.”