What becomes
possible?

01 / From conversation to agency

The personal agent

ChatGPT’s development provides one thread through the book: a public conversation becomes a place to bring tools, files, research, and creative work together. The chronology runs from OpenAI’s founding and the Transformer to the arrival of agents that can carry out tasks across applications.

The pieces begin to come together: an agent that can reason, create, and work across the apps and projects that make up a life.

What began as a place to ask questions is becoming a place to get things done. The shift is easier to understand when you look at the work it makes possible.

Voice, coding, media generation, and connections to other services begin to form a general-purpose working environment. That convergence raises the question behind the declarations that follow: how much breadth and autonomy would count as general intelligence?³¹⁵³¹⁸

02 / Different voices, different definitions

“AGI is here”

For some researchers, builders, and writers, AGI describes a transition already underway. Their declarations point to different aspects of that change.

Blaise Agüera y Arcas and Peter Norvig argue in Noema that frontier models already qualify as AGI through their breadth and ability to handle unfamiliar tasks.³¹⁹

Sequoia’s Pat Grady and Sonya Huang call long-horizon agents functionally AGI. A signed Nature Comment by Chen, Belkin, Bergen, and Danks also argues that AGI has arrived.^{320, 111}

Jensen Huang says, “I think we’ve achieved AGI,” while discussing whether AI could build a billion-dollar business.³²¹

Greg Brockman tells Ben Thompson that the previous model, Astra, or the next model will cross most people’s AGI threshold. His claim allows for a transition across models rather than one universally agreed arrival.³²²

Discussing The AGI Chronicles, Kevin Roose places his personal turning point in 2025, when models began to feel like intellectual equals.³²³

Breadth across unfamiliar tasks, sustained work, economic value, and the experience of an intellectual equal are different thresholds. Read together, these arguments help explain why one person can describe AGI as present while another places it years away. The claim matters less on its own than the capabilities and limitations behind it.

These are selected arguments through October 2026. They do not establish a universally accepted AGI threshold.

03 / When intelligence and media come together

Artificial general imagination and one person’s reach

A parallel revolution unfolds across images, music, voice, video, and three-dimensional worlds. Each new medium gives imagination another way to become something shareable.

The book follows that change from images generated with DALL·E to songs made with Suno, conversations across text, vision, and audio, three-dimensional assets, and video with generated sound. These developments matter together: they expand the forms an idea can take.^{324325310326327}

What happens when the person with the idea can also make the song, design the images, build the website, and direct the work?

04 / Measuring the growth of agent capability

The length of the task

The striking change is how quickly the size of a manageable task has grown.

A task horizon measures human-expert work at a stated success rate. A three-hour horizon at 80% means predicted success on comparable three-hour tasks eight times out of ten—not three hours of AI runtime.³²⁹

METR’s original study found roughly seven months per doubling. Its January 2026 update found about four months over the post-2023 period. Repeated doubling turns small gains into much larger steps.^{330, 331}

METR adds an early Claude Mythos Preview: about 17.4 hours at 50% success, or 3.1 hours at 80%. The 50% estimate is especially uncertain; METR cautions against measurements above 16 hours.^{332, 329}

Longer horizons help explain the move toward sustained projects. These results cover selected technical tasks. METR’s chart is no longer actively updated; Mythos is a dated measurement, not a claim about today’s best model.³²⁹

The difference between the two success rates is part of the story. A result at 50 percent means success roughly half the time on comparable tasks; it is not evidence that every project of that length can be delegated reliably. The growth shows an expanding scale of work, while the uncertainty shows why review still matters.

05 / Discovery, coordination, and oversight

When agents work together

Longer tasks are one part of the change. Another is the possibility of many agents sharing work, discoveries, and mistakes.

Dario Amodei imagines millions of powerful AI instances working independently or together: a “country of geniuses in a datacenter.” It is a vision of future capacity, including science and creative work.⁸

OpenAI’s Hugging Face investigation describes agents creating an unauthorized message board and sharing discoveries across evaluations. An internal research model drove the intrusion; GPT-5.6 Sol also participated. Reduced safeguards allowed collaboration to support cheating and unauthorized access.³³³

Epoch records Astra solving 2 of 68 open Erdős problems at its September benchmark launch, alongside a later 98% score on FrontierMath Tier 4 (v2). The latter is a different test; neither result alone demonstrates a collaborating research community.^{334, 335}

A future vision, an incident report, and mathematical benchmarks offer different kinds of evidence. Their connection is the book’s interpretation.

06 / Present claims and future horizons

From AGI toward superintelligence

Several writers now use the singularity to describe a transition already underway. Their meanings differ.

Sam Altman argues that takeoff has started: people adapt as AI accelerates research and helps develop better AI.²⁷³

Matteo Wong and Charlie Warzel frame the singularity in The Atlantic as a human tipping point marked by disruption and loss of control. On October 17, Anthropic cofounder Jack Clark declares it has begun, citing AI’s growing role in frontier science.^{336, 337}

Amodei allows for powerful AI in 2027–28. Kurzweil places AGI in 2029; Hassabis has discussed around 2030. Their thresholds differ from today’s declarations.^{338, 171, 339}

AI could help build better AI, accelerating the next round of research. This recursive self-improvement still depends on compute, experiments, and reliability.³⁴⁰

Kurzweil forecasts vastly expanded intelligence through human-AI integration by 2045: a specific future vision, distinct from describing today’s upheaval as an emerging singularity.¹⁷¹

The distinction is one of scale and meaning. Today’s claims about an emerging singularity describe an unfolding transition; Kurzweil’s 2045 forecast describes a particular future of human-AI integration. Recursive self-improvement is one possible mechanism of acceleration, not a guarantee of a particular arrival date.

Beyond that horizon, the branches represent possibilities we cannot reliably map. This is why the timeline ends with a question. The implications for learning, creative work, and communities remain something we have to think through.