The book in context / 1950–2045
One day, you can ask a computer a question. Then it can help you write a song, build a website, or work through a research problem. The change is easier to see when you put the moments next to one another.
This timeline follows the tools, ideas, and predictions that run through the book. Follow a date back to an essay. Compare a prediction with what was demonstrated. Ask what changed—and what remains an open question.
Selected history, not a complete inventory of AI. Historical entries record events and statements; target years record forecasts. Reviewed October 22, 2026. Why stop at 2045? ↗
One thread through the timeline / 2022–2026
In November 2022, you could open ChatGPT and begin a conversation. Less than four years later, that starting point can lead to research, code, visual designs, presentations, and working websites. The change is in how much of a project you can ask it to help carry out.
A free research preview makes it possible to ask follow-up questions, develop a draft, or get help with code.
The research preview ↗Run code on uploaded data. Talk aloud, discuss an image, and create one. The conversation begins producing things you can use.
Code interpreter ↗GPT-4o brings modes of interaction together. o1 spends more time on hard problems. Canvas adds a place to revise writing and code.
GPT-4o and what followed ↗Deep research investigates a question. Codex works on a codebase. ChatGPT agent connects research, computer tools, and editable outputs.
ChatGPT agent ↗ChatGPT Work carries projects across apps and files. Image tools add planning and research. Dots can keep working toward goals over time.
Work, visual creation, and Dots ↗You still need a good question. You also need to define the task, choose what the system can access, examine its work, and decide what is ready to use. That is where AI literacy becomes a practice of judgment.
Selected milestones, not a complete release history. Each link leads to a dated entry and its original sources. Capabilities arrived gradually and vary by tools, plan, and permissions. “General-purpose agent” describes the range of tasks; it does not settle the question of AGI.
1950–2021
The internet, digital creation, and the long-running question of machine intelligence.
Alan Turing’s “Computing Machinery and Intelligence” introduces the imitation game. The question of how to recognize machine intelligence becomes a recurring reference point in later debates about AGI.
Ray Kurzweil
In The Age of Spiritual Machines, Kurzweil forecast that AI would pass a valid Turing test by 2029. His 2024 book explicitly traces the prediction to 1999.
Listen.com launches Rhapsody, a subscription service for listening to music online. The shift toward access rather than ownership helps frame the book’s later questions about AI-generated music.
Apple announces GarageBand as part of iLife ’04, with availability beginning January 16. Making music on a computer becomes another step in the longer story of who gets to create.
Ray Kurzweil
Kurzweil develops his forecast of a transformative merger of human and machine intelligence around 2045.
Apple announces the iPhone, combining a mobile phone, music player, and internet device. It goes on sale that June. The book uses earlier shifts in access to think about the scale of AI’s arrival.
Spotify launches in six European markets. Streaming’s growth becomes part of the book’s account of how platforms change music discovery, distribution, and the choices available to listeners.
The FCC’s National Broadband Plan recommends a Digital Literacy Corps to help people develop digital skills. The proposal provides a historical starting point for the book’s argument for an AI Corps.
This was a recommendation in a national plan, not the announcement of a fully operating nationwide corps.
Spotify launches in the United States. Separating the European and US launches makes the sequence of the streaming era easier to follow.
Founded in 2015, OpenAI publicly introduces itself as a nonprofit AI research organization on December 11. ChatGPT is still nearly seven years away. The organization’s original ambition is to advance digital intelligence for broad human benefit.
December 11 is the public announcement date. This entry describes OpenAI’s founding structure and mission in 2015, rather than its later corporate structure.
AlphaGo wins its five-game match against Lee Sedol 4–1. Its unexpected Move 37 in the second game becomes an example of an AI system finding a successful move outside familiar human play.
This was a breakthrough in Go, a specific domain. It did not establish general intelligence.
OpenAI publishes the work now associated with the first GPT: a model learns from large amounts of text, then is adapted to different language tasks. The research lays groundwork for a family of models that will eventually power ChatGPT.
This is a research milestone, not the release of a public conversational assistant. GPT models and the ChatGPT product are different parts of the story.
2022–2024
Conversation, images, music, and code become places to experiment.
OpenAI begins previewing DALL·E 2 with a limited group of users. The system creates and edits images from natural-language descriptions.
The preview began with limited access. The May update dates its start to April; it was not yet a general public release.
OpenAI releases ChatGPT as a free research preview. People can ask follow-up questions, request explanations, and work through writing or coding tasks in a conversation.
OpenAI introduces GPT-4 and makes it available through ChatGPT Plus. The company reports improved reasoning and instruction following while acknowledging persistent hallucinations and other limitations.
Microsoft announces Microsoft 365 Copilot, bringing natural-language assistance to applications including Word, PowerPoint, Excel, Outlook, and Teams. Initial testing involves a small group of customers.
Announcement and initial testing, rather than general availability. This is the context for the essay’s anticipation of Copilot reaching Word.
Code interpreter begins rolling out in beta to ChatGPT Plus users. People can upload files and ask ChatGPT to run code, analyze data, create charts, and edit files. A conversation can now produce a computed result or a downloadable artifact.
This dates the wider Plus beta rollout. Earlier limited tests preceded it. Executing code in a controlled environment is different from independently operating a person’s computer.
OpenAI begins adding voice conversations and image input to ChatGPT. People can talk aloud or show it a photograph, diagram, or screenshot, making the material they bring into a conversation more varied than typed text.
The announcement describes a gradual rollout to Plus and Enterprise users, with voice initially on mobile. Image understanding is distinct from generating an image.
OpenAI makes DALL·E 3 available in ChatGPT Plus and Enterprise. Users can describe an image and request revisions within the same conversation.
Suno releases v3 to all users, offering full two-minute songs generated in seconds. The company highlights improvements in audio quality, musical styles, and following prompts.
OpenAI introduces GPT-4o, trained across text, vision, and audio. Its demonstrations point toward a more immediate spoken interaction with AI. Text and image understanding begin rolling out in ChatGPT, including to free users.
Capabilities were released in stages. The new voice experience was announced for a later alpha; the launch demonstrations should not be read as universal day-one access to every mode.
Ray Kurzweil
The sequel revisits the earlier forecasts: human-level AI by 2029 and the singularity around 2045.
OpenAI releases o1-preview, a model designed to spend more computation reasoning before responding, with a focus on challenging science, mathematics, and coding problems.
OpenAI begins rolling out Canvas to ChatGPT Plus and Team users. The separate editing interface lets people revise text or code directly and ask for focused feedback on selected passages.
The initial release was an early beta for Plus and Team users. Access expanded afterward.
Lovable launches with the aim of making software building accessible through natural-language conversation. Its approach becomes central to the book's experiments in websites, applications, and design literacy.
The date comes from Lovable’s official first-anniversary account. Features added later should not be assumed to have existed at launch.
Google begins rolling out Deep Research to Gemini Advanced subscribers and gives users access to the experimental Gemini 2.0 Flash model. The research feature explores a topic and assembles a report.
ARC Prize reports a large improvement by an unreleased OpenAI o3 system on ARC-AGI-1. Results vary with the computation budget; the test concerns a defined benchmark, not a determination that AGI has arrived.
ARC Prize tested an unreleased preview, distinct from the o3 model shipped in April 2025. Benchmark performance depends on the test and computation budget.
2025–2026
Systems take on more steps. People disagree about what those capabilities mean.
Sam Altman
Altman says OpenAI is confident it knows how to build AGI, anticipates agents entering the workforce during 2025, and sets out an ambition beyond AGI. This is a statement of confidence and direction, not an announcement that AGI has been achieved.
OpenAI releases Operator as a research preview for Pro users in the United States. The agent can read a webpage and interact by typing, clicking, and scrolling to carry out tasks.
Early access was limited to US Pro subscribers, with acknowledged failures and human handoffs. The later ChatGPT agent integration followed in July.
OpenAI launches a capability that conducts multi-step web research and creates reports with citations. Initially offered to Pro users, it reaches Plus users on February 25.
Google Research introduces a Gemini-based multi-agent collaborator that proposes and refines scientific hypotheses. In one test, it independently suggests a bacterial gene-transfer mechanism that researchers had discovered but not yet published.
Human researchers guided the work and performed laboratory validation. The announcement described a research system and trusted-tester program.
Anthropic announces Claude 3.7 Sonnet and introduces Claude Code in a limited research preview. The coding agent can read and edit files, run tests, and use command-line tools while a developer directs the work.
METR introduces a way to compare agents by the human time needed for tasks they can complete at a specified success rate. Its initial study finds an approximately seven-month doubling in the 50%-success time horizon over the preceding six years.
The measure uses the time humans need to complete tasks, not how long an AI runs. The study mainly tested software and reasoning work; extending its trend into the future is an extrapolation.
GPT-4o gains native image generation in ChatGPT, with improvements in following detailed instructions, rendering text, and refining an image across turns. The creative task can become a sequence of choices about a poster, diagram, character, or other visual.
This builds on earlier DALL·E integration. It is a new image-generation approach, not the first time ChatGPT could make an image. Outputs still require review for errors and suitability.
Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean
The AI Futures Project publishes a detailed scenario of rapid AI progress with race and slowdown endings. The authors invite debate; the scenario is one possible future, not an announced schedule.
This preserves the original April 2025 scenario. Its authors subsequently clarified the uncertainty and revised their forecasting models; 2027 is not a settled arrival date or necessarily their current estimate.
OpenAI introduces the Codex software engineering agent in research preview, accessible through ChatGPT. It can read and edit a codebase, run tests, fix bugs, and propose changes for review in a cloud environment. People can delegate a task and inspect the work it produces.
This is the 2025 agent, distinct from the earlier Codex model. Initial access covered selected paid plans; the announcement emphasized reviewing the resulting changes and test evidence.
Google introduces Veo 3, which generates video with native audio, including dialogue, sound effects, and ambient sound. Initial consumer availability is for Ultra subscribers in the United States through Gemini and Flow.
ChatGPT agent combines web research with tools for acting on websites and running code. Using its own virtual computer, it can work through a request and produce editable slides or spreadsheets. The user can steer, interrupt, and review consequential actions.
The announcement combines capabilities previously associated with Operator and deep research. It describes a staged paid-plan rollout, not unrestricted autonomy or guaranteed task completion.
Mark Zuckerberg
Zuckerberg describes Meta’s goal of personal superintelligence that helps people pursue their own goals. He characterizes the rest of the decade as consequential for its direction, without giving a delivery date.
Google introduces an updated Gemini image-editing model known as Nano Banana, focused on retaining recognizable people and pets while changing outfits, combining photos, and transferring visual styles.
Eddy Keming Chen, Mikhail Belkin, Leon Bergen, David Danks
Four UC San Diego researchers argue that present AI meets a meaningful threshold of human-level intelligence. Nature publishes their argument as a Comment. This records a position in an ongoing debate, not a journal-wide declaration or a settled finding that AGI has been achieved.
Jensen Huang
On Lex Fridman’s podcast, Huang says he thinks AGI has been achieved while discussing an AI capable of creating a billion-dollar business. He then narrows the example to a potentially short-lived viral application. The claim does not establish broad expert consensus or a demonstrated system meeting every definition of AGI.
ChatGPT Images 2.0 adds a thinking mode that brings reasoning and tool use into image creation. It can use web research, plan an output, and generate multiple images from one request. Posters, explanatory visuals, and coordinated assets become more practical creative projects.
The system card describes improved instruction following and dense text generation, while also documenting remaining risks. Thinking mode and access differ by plan; finished visuals still need an editor’s judgment.
OpenAI introduces ChatGPT Work, drawing on Codex technology to carry out longer tasks across connected apps and files. It can research, analyze, code, and create documents, spreadsheets, presentations, and Sites. The work can continue through multiple steps while a person reviews progress and changes direction.
OpenAI describes projects lasting hours and scheduled follow-up tasks. Available tools, permissions, plans, and review requirements shape what an agent can actually complete. General-purpose here describes a range of work; it does not establish AGI.
OpenAI introduces Dots: persistent agents with their own cloud computers that can work toward goals, use connected apps, and learn from feedback. Available through ChatGPT and other work surfaces, they extend delegation beyond a single conversation or task.
The announcement begins a rollout to selected plans and markets. Always-on describes the intended working arrangement, not infallibility. People still set access, review work, and approve important actions.
In “What Is Worth Building?”, a student’s curiosity about her mother’s business brings the book’s argument into focus: learning to use AI starts with something that matters to the person using it.
2026–2045
Target years and imagined settings. Read these as proposals to examine, not a schedule to expect.
Dario Amodei · Forecast made October 2024
Amodei’s Machines of Loving Grace imagines highly capable, autonomous AI systems operating at scale. He says powerful AI could arrive as early as 2026, while allowing that it could take much longer.
Amodei uses “powerful AI” for his imagined systems. This is an earliest possibility with open-ended uncertainty, not a promised 2026 launch.
AI 2027 authors · Forecast made Apr 3, 2025
The original AI 2027 scenario explores what might follow if systems become exceptionally capable at coding and accelerate AI research. Its branching outcomes depend on choices about deployment, oversight, and competition.
This preserves the original April 2025 scenario. Its authors subsequently clarified the uncertainty and revised their forecasting models; 2027 is not a settled arrival date or necessarily their current estimate.
Ray Kurzweil · Forecast made 1999; reiterated 2005 and 2024
Kurzweil’s target for human-level AI includes passing a valid Turing test. It is distinct from his 2045 singularity forecast.
Passing a Turing test is one proposed threshold; there is no single test that settles every definition of AGI.
Demis Hassabis · Forecast made Apr 20, 2025
His estimate: five to ten years from the interview.
World Economic Forum · Forecast made January 2025
The Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced by 2030. These estimates cover multiple economic and technological trends, not AI alone.
Based on an employer survey and modeled projections. They are forecasts of change across the economy, not counts of jobs already created or lost.
Kyle Bylin
The Imagination Lab opens with students whose AI companions can see and hear the world around them. The scene asks what libraries and educators might help students do with that access.
An imaginative planning exercise in the book, rather than a prediction that these events will occur on this date.
The source is the essay linked above. This entry describes its imagined setting.
Kyle Bylin
Music 2045 looks toward music shaped by our moods, surroundings, and personal data. Its deeper question is about human expression and what listeners will value when generating music becomes easier.
An imaginative planning exercise in the book, rather than a prediction that these events will occur on this date.
The source is the essay linked above. This entry describes its imagined setting.
Kyle Bylin
Libraries 2045 uses a distant future to ask what libraries could become. The point is to explore choices about access, learning, and community that people can begin making now.
An imaginative planning exercise in the book, rather than a prediction that these events will occur on this date.
The source is the essay linked above. This entry describes its imagined setting.
Ray Kurzweil · Forecast made 2005; reaffirmed 2024
Kurzweil anticipates a profound expansion of human intelligence through integration with AI. This is his proposed singularity horizon, not an established endpoint of AI progress.
2045 / The edge of the map
Kurzweil imagines a transformation so profound that our present intelligence may struggle to comprehend it. If the ways we think, discover, and create were themselves changing, familiar comparisons could stop helping us predict what comes next.
That is why this timeline ends with a horizon. We use his 2045 forecast as a place to stop assigning dates and start opening questions. It is an editorial choice, not a known deadline or a guarantee that a singularity will occur.
Choose a part of life and imagine what might change.
These are invitations to imagine. Bring one question back to a choice you can make today.
Read Kurzweil’s explanation in The Singularity Is Nearer ↗No entries match this combination. Try another term, choose all essays, or reset the filters.
Bring the timeline into a conversation
Choose one milestone and one prediction. What could people actually do at that moment? What evidence would persuade you that the prediction had come true? Then find the essay that helps you think through why it matters.
Dates follow the linked sources at the precision they support. Forecasts retain the date and viewpoint of their original publication. This companion timeline does not rewrite the essays or continuously monitor new releases.