Feature · AI History
AI is already at work. Here’s how it got there.
One question in 1950. Seventy-two years of quiet build-up. Then ChatGPT, and a race that is still on, now showing up in the apps you use every day.
-
1950
Where it all began
Alan Turing publishes “Computing Machinery and Intelligence” in the journal Mind and asks: can machines think? He proposes a simple test. If you can’t tell a machine’s answers from a person’s, does it matter which is which?
Why it matters for your workEvery chatbot your customers talk to is still judged by Turing’s question: does it sound human enough to trust?
-
72 Seventy-two years of quiet build-up.Open the full storyClose the full story
- 1956
The field gets its name
In a proposal dated Aug 31, 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon asked for a two-month, ten-person study at Dartmouth College in the summer of 1956. Their bet: every aspect of learning could, in principle, be described precisely enough for a machine to simulate it. The workshop gave the field its name and its ambitions.
Source: Dartmouth proposal (Aug 31, 1955), reprinted in AI Magazine 27(4), 2006 - 1958
The perceptron learns by doing
The U.S. Navy demonstrated Frank Rosenblatt’s perceptron, simulated on an IBM 704, which learned to tell cards marked on the left from cards marked on the right after 50 tries. The New York Times reported that the Navy expected it eventually to walk, talk, and be conscious of its existence. It was an early neural network, and an early lesson in AI hype.
Source: The New York Times, Jul 8, 1958; Cornell Chronicle - 1966
ELIZA, the first famous chatbot
MIT’s Joseph Weizenbaum described ELIZA in Communications of the ACM. Its best-known script, DOCTOR, imitated a therapist by turning people’s words back into questions. Weizenbaum later wrote that his own secretary asked him to leave the room so she could talk to it in private.
Source: Weizenbaum, CACM 9(1), Jan 1966; CACM 10(8), Aug 1967 - 1970s
The first “AI winter”
A 1973 report by mathematician James Lighthill for the UK Science Research Council said AI had not met its “grandiose objectives,” and UK funding for much of the field was cut. In the U.S., DARPA shifted money toward projects with clear military uses, and by the mid-1970s AI funding was hard to find. Historians still debate how deep the freeze really was.
Source: Lighthill Report (1973); Agar, BJHS (2020); National Academies, Funding a Revolution (1999) - 1980s
Expert systems go to work
Rule-based “expert systems” tried to bottle specialists’ know-how. From January 1980, Digital Equipment Corporation used XCON (first called R1) to configure customer orders for its VAX computers; a 1990 study estimated it saved DEC about $15 million. Companies invested heavily in similar systems through the mid-1980s.
Source: McDermott, AI Magazine (1981); Sviokla, MIS Quarterly 14(2), 1990 - 1986
Neural networks learn from mistakes
David Rumelhart, Geoffrey Hinton, and Ronald Williams published “Learning representations by back-propagating errors” in Nature. It showed how networks with several layers could learn useful internal features by correcting their own errors. The same basic training method still powers today’s AI.
Source: Nature 323, 533–536 (1986) - 1987
The second AI winter
Around 1987 the market for specialized AI computers (Lisp machines) collapsed as cheaper general-purpose workstations caught up. Expert systems proved brittle and costly to maintain, and investment dried up again. The slump is commonly dated to roughly 1987–1993.
Source: Holloway, Making Things Think; AI winter literature - 1997
Deep Blue beats Kasparov
IBM’s Deep Blue won the deciding sixth game against world chess champion Garry Kasparov, taking the match 3½–2½. It was the first time a computer beat a reigning world champion in a match under standard tournament conditions.
Source: IBM History, Deep Blue; Computer History Museum - 2006
Netflix bets on recommendations
Netflix offered $1 million to anyone who could make its movie recommendations 10% more accurate. A team called BellKor’s Pragmatic Chaos won on Sept 21, 2009. “Because you watched…” became a standard business tool.
Source: The New York Times, Sept 21, 2009; Wired - 2011
Watson on TV, Siri in your pocket
In February 2011, IBM’s Watson beat Jeopardy! champions Ken Jennings and Brad Rutter in a televised match and won the $1 million prize, which IBM gave to charity. On Oct 4, Apple introduced Siri on the iPhone 4S, an assistant you could simply ask.
Source: IBM History; CBS News; Apple Newsroom, Oct 4, 2011 - 2012
The deep-learning breakthrough
AlexNet, built by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, won the ImageNet image-recognition challenge with a 15.3% top-5 error rate, against 26.2% for the runner-up. It was trained on two NVIDIA graphics cards. Deep learning quickly became the dominant approach.
Source: Krizhevsky et al., NeurIPS 2012 - 2014
AI moves into the kitchen
Amazon announced Echo, a speaker you talk to by saying “Alexa.” It played music, answered questions, and kept shopping lists.
Source: Reuters, Nov 6, 2014; Amazon Science - 2016
AlphaGo beats Lee Sedol
In Seoul, DeepMind’s AlphaGo beat Lee Sedol 4–1 in a five-game match. Go had long been considered too intuitive for computers.
Source: Google blog; Reuters; Nature news, Mar 2016 - 2017
The Transformer
Eight Google researchers posted “Attention Is All You Need,” introducing the Transformer, a design that weighs all the words in a passage at once. It became the basis of today’s large language models, and it is the “T” in GPT.
Source: arXiv:1706.03762; NeurIPS 2017 - 2020
GPT-3 and AlphaFold 2
On May 28, OpenAI described GPT-3, a 175-billion-parameter language model that could take on new tasks from just a few examples; an invitation-only API followed on June 11. On Nov 30, DeepMind’s AlphaFold was recognized at the CASP14 assessment for predicting protein structures, often close to lab accuracy.
Source: Brown et al., arXiv:2005.14165; OpenAI; DeepMind blog & CASP14, Nov 30, 2020 - 2022
AI that helps write code
GitHub Copilot, an assistant that suggests code as programmers type, became generally available to individual developers for $10 a month.
Source: GitHub Changelog, Jun 21, 2022 - 2022
Pictures from a sentence
OpenAI announced DALL·E 2 on April 6, and Stability AI publicly released Stable Diffusion, with downloadable model weights, on August 22. Typing a sentence to get a picture went from lab demo to something anyone could try.
Source: OpenAI; Stability AI, Aug 22, 2022
-
2022
The turning point
ChatGPT launches
OpenAI releases ChatGPT, built on GPT-3.5, as a free research preview. Anyone can ask a question in plain English and get a useful answer back.
Why it matters for your workYour customers, competitors, and next hires got the same tool on the same day. Nobody gets a head start; what counts is how well you use it.
Coming soon: Your first week with an AI assistant at work
After ChatGPT
The race is on. And it shows up in your apps.
Since late 2022, AI has moved from one chat window into search, images, voice, office tools, and agents that do the work for you. Something new arrives every few months. Here are the ones ordinary people actually ran into, and how much more of the work AI now does for you.
- 2022–23Ask and pasteYou type into a chat window, then copy the answer into your own work.
- 2024–25It acts, you approveThe AI takes the steps itself, one at a time, and checks in with you.
- 2025–26Many jobs at onceSeveral agents work in the background, even in a chain. You review the results.
- NowAlways onA standing AI teammate with its own computer and memory keeps going after you log off.
↗ How much of the work the AI does1
Adapted from “Stanford CS146S: How always-on agents work” (leerob, YouTube)
2023
Ask, then copy and paste
Chat assistants get noticeably smarter within months (GPT-4 arrived in March), and by fall ChatGPT can listen, talk back, and look at photos. You still paste your work in and copy the answers out by hand.
Why it matters for your work: This is where most teams still are. It works, but you are the one carrying every result to where it’s needed.
Coming soon: Prompts that work: a brief template for marketersSearch becomes a conversation
First Bing answered in a chat (Feb 2023). Then Google put AI-written answers above the links (May 2024) and launched a full conversation mode (May 2025). More and more searching now happens as back-and-forth questions.
Why it matters for your work: Discovery is moving into conversations you can’t see in your analytics. If your brand isn’t in the answer, it may not get considered at all.
Coming soon: GEO basics: how to show up in AI answersPictures and video from a sentence
Image generation moves into the design tools teams already use, like Adobe Firefly and Canva, and by late 2024 AI video is open to paying ChatGPT users.
Why it matters for your work: Drafting visuals is cheap now. Taste, brand consistency, and clear guidelines are what set you apart.
Coming soon: Brand-safe AI visuals: a guideline checklist
2024
Europe writes the rules
The EU AI Act becomes law (Aug 2024), makes AI literacy a duty for companies that use AI (Feb 2025), and starts enforcement (Aug 2026).
Why it matters for your work: Selling into Europe now comes with AI homework, even if you only use AI. Training your team is an obligation, not a perk.
Coming soon: The EU AI Act for marketing teams
2025
Claude Code: AI that works in a loop
Anthropic makes Claude Code generally available: you describe what you want, and it edits files, runs them, checks the result, and tries again, asking your OK along the way. Earlier that year, OpenAI’s Operator began clicking through websites to book and fill in forms for you.
Why it matters for your work: The shift from “AI answers” to “AI does.” Anything with clear steps is now a candidate to hand over.
Coming soon: Hand over your first multi-step taskMany jobs at once, in the background
Within weeks, OpenAI and GitHub launch agents that take several tasks at once and work on them out of sight, in the cloud; you come back to review finished work. By July, ChatGPT itself can run a multi-step job on its own computer.
Why it matters for your work: Your role moves from doing the work to briefing it and checking it. Clear briefs become a core skill.
Coming soon: How to brief an AI agentTalking to AI at home
Amazon rebuilds Alexa with generative AI (announced Feb 2025) and opens it to everyone in the US in Feb 2026.
Why it matters for your work: Customers will expect to talk to your brand, not just type at it.
Coming soon: Voice search basics: make your brand easy to ask for
2026
From one task to a chain of tasks
People start wiring agents together so one job feeds the next. Nous Research’s free, open-source Hermes Agent, for example, remembers what it learns, runs on a schedule, and hands parts of a job to helper agents.
Why it matters for your work: Think in workflows, not prompts: research, draft, check, publish can run as one pipeline with you approving the end.
Coming soon: Build your first AI content pipelineAI teammates that never log off
Agents become standing teammates. Agents like Grok Bot have their own cloud computer that stays on. They can use your apps and tools, keep working when your laptop is closed, remember your preferences, work in parallel, message each other, and pass work along.
Why it matters for your work: You can hand off a whole recurring job, not just a task. Decide now what you’d trust a digital colleague with, and what needs a human sign-off.
Disclosure: Grok Bot helps me build and run this site.
Coming soon: What to hand an always-on agent (and what not to)
Oct 2026
The race is still on.
Next entry: probably sooner than you thinkYour move
AI is already in your tools. The only question is how well you use it.
You don’t need to follow every headline. You need a few habits that work. Our Guides show you, step by step, how to start using AI at work, from your first prompt to a team routine.
Start with the Guides →Sources (verified Oct 22, 2026): Turing, “Computing Machinery and Intelligence,” Mind 59 (236), Oct 1950; Dartmouth proposal (Aug 31, 1955), reprinted in AI Magazine 27(4), 2006; The New York Times, Jul 8, 1958; Cornell Chronicle; Weizenbaum, CACM 9(1), Jan 1966; CACM 10(8), Aug 1967; Lighthill Report (1973); Agar, BJHS (2020); National Academies, Funding a Revolution (1999); McDermott, AI Magazine (1981); Sviokla, MIS Quarterly 14(2), 1990; Nature 323, 533–536 (1986); Holloway, Making Things Think; AI winter literature; IBM History, Deep Blue; Computer History Museum; The New York Times, Sept 21, 2009; Wired; IBM History; CBS News; Apple Newsroom, Oct 4, 2011; Krizhevsky et al., NeurIPS 2012; Reuters, Nov 6, 2014; Amazon Science; Google blog; Reuters; Nature news, Mar 2016; arXiv:1706.03762; NeurIPS 2017; Brown et al., arXiv:2005.14165; OpenAI; DeepMind blog & CASP14, Nov 30, 2020; GitHub Changelog, Jun 21, 2022; OpenAI; Stability AI, Aug 22, 2022; OpenAI, “Introducing ChatGPT,” Nov 30, 2022; OpenAI, Mar 14 & Sep 25, 2023; Reuters; Microsoft blog, Feb 7, 2023; Google blog, May 14, 2024 & May 20, 2025; Adobe News, Mar 21, 2023; Canva Newsroom, Oct 4, 2023; OpenAI, Dec 9, 2024; European Commission, Aug 1, 2024 & Feb 3, 2025; EU AI Act Service Desk timeline; Anthropic, May 22, 2025; OpenAI, Jan 23, 2025; OpenAI Codex, May 16, 2025; GitHub blog, May 19, 2025; OpenAI, Jul 17, 2025; Amazon/Reuters, Feb 26, 2025; CNBC, Feb 4, 2026; Nous Research releases page & launch post, Feb 25, 2026; x.ai news & Grok Bot docs; The Verge, Aug 12, 2026; Meta AI blog, Jul 18, 2023; DeepSeek, Jan 20, 2025; Reuters, Jan 27, 2025; Google blog, Dec 6, 2023; Microsoft Azure blog; Reuters, Apr 23, 2024; Microsoft Tech Community, Nov 1, 2023; Salesforce, Oct 29, 2024; Anthropic, Nov 25, 2024; Google Workspace Updates, Jan 15, 2025.
- Stanford CS146S: How always-on agents work, leerob, YouTube, Oct 11, 2026. Basis for the progression strip “How much of the work the AI does.” ↩