Brain Bank

Search Brain Bank

Search is available on the built site (npm run build).

Section 3 of 9

Growth of AI & Tooling

AI milestone timeline from GPT-3 to agents, dev tools, and the measured impact on software jobs.

Milestone timeline

DateMilestoneWhy it mattered
Jun 2020GPT-3 (OpenAI, 175B params)First “general-purpose” text model via API; few-shot prompting era begins
Jun 2021GitHub Copilot technical preview (Codex-powered); GA Jun 2022First mainstream AI pair-programmer; ~3 years with no serious rival
Apr 2022DALL·E 2 / Stable Diffusion (Aug 2022)Generative images go mainstream
Nov 2022ChatGPTFastest-growing consumer app in history; AI becomes a household word
Mar 2023GPT-4; Anthropic launches ClaudeMultimodal reasoning leap; serious coding ability
2023LangChain/LlamaIndex boom; vector-DB wave; Cursor (AI-first IDE) ships”LLM app stack” forms; RAG becomes standard pattern
Feb–Mar 2024Claude 3 family; open-weights surge (Llama 3, Mistral)Frontier competition; local/open models viable
Jun 2024Claude 3.5 SonnetCoding-model bar jumps; agentic coding becomes practical
Nov 2024MCP (Model Context Protocol) open-sourced by Anthropic; Windsurf launches (Codeium)Standard protocol for tool-connected AI; agentic IDE race begins
Feb 2025Claude Code research previewTerminal-native agentic coding; ~$1B annualized revenue within ~6 months [secondary]
May 2025GitHub Copilot Coding AgentEvery major vendor now ships autonomous coding agents
2025–2026Agent era: computer-use, multi-agent orchestration, MCP industry adoption; Cursor ARR $100M (Jan 2025) → ~$1B (Nov 2025) → ~$2B (Feb 2026) [secondary]Tooling shifts from autocomplete → autonomous task execution

Adoption evidence (2025, verified): 81.4% of Stack Overflow respondents use OpenAI GPT models, 42.8% Claude Sonnet, 17.9% Cursor ✅. Embedding AI into software is now a mainstream dev activity (25% of devs, vs 29% doing data processing — JetBrains 2025) ✅.

Measured impact on software jobs — the three best studies

  1. METR RCT (the productivity reality-check) ✅ verified 16 experienced open-source devs, 246 real tasks on mature repos, Feb–Jun 2025, mostly Cursor Pro + Claude 3.5/3.7 Sonnet. Result: AI tools increased task completion time by 19%. Devs forecast +24% speedup beforehand and still believed +20% after — perception inverted from measurement. Critical caveat (also verified): METR’s Feb 2026 follow-up found selection effects and estimates the 2026-tool cohort at ~−4% (CI −15% to +9%) — i.e., the slowdown largely evaporated as tools matured. Cite the pair, never the 19% alone.

  2. Stanford “Canaries in the Coal Mine” (the employment data) ✅ verified ADP payroll microdata through Sep 2025: ages 22–25 in most-AI-exposed occupations (software developers named) −16% relative employment (controlling for firm shocks); experienced workers +6–9% or stable. Non-peer-reviewed, methodology contested, ADP-client sample — strong signal, not settled fact.

  3. Employer intentions (WEF 2025) ✅ verified Two-thirds of employers will hire for AI skills; 40% expect reductions where AI automates; 39% of skill sets transformed/outdated by 2030.

Synthesis: AI tooling adoption is near-universal and real, measured productivity gains are smaller and later than the hype (early tools even negative for experts on complex code), and the labor effect so far is redistribution toward experience and AI-skill, not mass replacement. Sentiment matches: 64% of devs say AI is not a threat to their job ✅.

Think of AI’s rise like the arrival of the power tool in a world of hand carpenters.

The story so far:

  • 2020: A lab demo shows a machine that can write passable text (GPT-3). Impressive, niche.
  • 2021: First power drill for programmers (GitHub Copilot) — it suggests the next few lines of code, like autocomplete on your phone but for programs.
  • Nov 2022: ChatGPT — the power tool goes retail. Fastest-adopted consumer product ever. Suddenly everyone’s boss is asking “can AI do this?”
  • 2023–2024: Tools stop just suggesting and start doing — whole functions, whole files. New AI-first editors (Cursor, Windsurf) appear.
  • 2025–2026: The “agent” era. You no longer hold the drill — you describe the shelf, and a robotic apprentice (Claude Code, Copilot Agent) builds it, checks it, and reports back. These apprentice tools became billion-dollar products in months.

What actually happened to jobs — three careful measurements:

  1. The stopwatch test: Scientists timed expert programmers with and without early-2025 AI tools on hard, real work. Surprise: they were 19% slower with AI — while believing they were faster. (A 2026 re-test showed the gap closing to roughly zero as tools improved.) Lesson: the hype ran ahead of the reality, especially for expert work.
  2. The payroll records: Stanford examined millions of real paychecks. Young beginners (22–25) in AI-exposed jobs like programming fell ~16% — but experienced workers grew. The ladder lost its bottom rungs; the upper rungs held.
  3. The employer survey: two-thirds of major employers plan to hire people with AI skills. They aren’t deleting jobs wholesale; they’re changing the entry ticket.

Plain conclusion: AI is a power tool, not a replacement carpenter — so far. But carpenters who refuse to learn power tools compete with those who did.

Diagrams

AI tooling milestones, 2020 → 2026

Jun 2020 — GPT-3 Jun 2020 GPT-3 2021 — Copilot preview 2021 Copilot preview Nov 2022 — ChatGPT (inflection point) Nov 2022 ChatGPT Mar 2023 — GPT-4 / Claude Mar 2023 GPT-4 / Claude 2023 — Cursor + LLM app stack 2023 Cursor + LLM app stack Jun 2024 — Claude 3.5 Sonnet Jun 2024 Claude 3.5 Sonnet Nov 2024 — MCP + Windsurf Nov 2024 MCP + Windsurf Feb 2025 — Claude Code (inflection point) Feb 2025 Claude Code May 2025 — Copilot Agent May 2025 Copilot Agent 2026 — Agent era 2026 Agent era
Eras: Autocomplete (2021–23) → Chat assistant (2023–24) → Agent (2025–). Starred points are inflections. · Vendor announcements: GitHub, OpenAI, Anthropic (Claude Code Feb 2025, MCP Nov 2024)
Data table
DateMilestone
Jun 2020GPT-3
2021Copilot preview
Nov 2022ChatGPT ★
Mar 2023GPT-4 / Claude
2023Cursor + LLM app stack
Jun 2024Claude 3.5 Sonnet
Nov 2024MCP + Windsurf
Feb 2025Claude Code ★
May 2025Copilot Agent
2026Agent era

Expert devs with AI tools — perception vs measurement (METR RCT)

Forecast before: +24% faster Forecast before +24% faster Believed after: +20% faster Believed after +20% faster Measured (early-2025 tools): −19% slower Measured (early-2025 tools) −19% slower 2026 re-estimate: ~−4% (CI −15%..+9%) 2026 re-estimate ~−4% (CI −15%..+9%) % change in task completion time
Perception and reality diverged in early 2025 — devs forecast and later believed they were faster while measured performance was slower. The gap closed as tools matured by 2026. · METR, “Measuring the Impact of Early-2025 AI on Experienced OSS Developer Productivity” + Feb 2026 follow-up
Data table
CategoryValue
Forecast before+24% faster
Believed after+20% faster
Measured (early-2025 tools)−19% slower
2026 re-estimate~−4% (CI −15%..+9%)

Sources

  1. METR — Measuring the Impact of Early-2025 AI on Experienced OSS Developer Productivity (PDF) ✅ verified
  2. Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?” ✅ verified
  3. WEF Future of Jobs Report 2025 ✅ verified
  4. Stack Overflow 2025 — AI section ✅ verified
  5. JetBrains State of Developer Ecosystem 2025 ✅ verified
  6. State of AI Report secondary

✅ = survived 3-voter adversarial verification against primary sources.