AGI timelines
When will AI be able to do almost everything people can? Forecasts range from a year or two to many decades. This path collects who has said what, and when, then the scenarios and arguments behind the numbers.
Two things to keep in mind:
- People mean different things by AGI. Some mean beating the best human at everything, some mean doing most paid work, some mean automating AI research itself. Check the definition before comparing dates.
- When a forecast was made matters as much as what it says. Most forecasts moved sharply earlier after ChatGPT and GPT-4 in 2022–23, and some have moved back and forth since.
Who predicted what, and when
For each person, the newest statement comes first. Links go to the original or to a report of it.
People building it
| Who | When said | Prediction |
|---|---|---|
| Dario Amodei (Anthropic) | Jan 2026 | Powerful AI could be one to two years away, or considerably further |
| Oct 2024 | Powerful AI “could come as early as 2026” (Machines of Loving Grace) | |
| Sam Altman (OpenAI) | Jan 2025 | ”We are now confident we know how to build AGI” (Reflections) |
| Sep 2024 | Superintelligence possibly “in a few thousand days” (The Intelligence Age) | |
| Demis Hassabis (Google DeepMind) | Mar 2025 | AGI will start to emerge in the next five to ten years |
| Shane Legg (Google DeepMind) | Oct 2023 | Still about 50% by 2028, on the Dwarkesh Podcast |
| 2011 | About 50% by 2028, a forecast he has held since (interview) | |
| Elon Musk (xAI) | Apr 2024 | Smarter than the smartest human “probably next year, within two years” |
| Mar 2024 | Smarter than any single human next year; smarter than all humans combined by 2029 |
Researchers and forecasters
| Who | When said | Prediction |
|---|---|---|
| Daniel Kokotajlo (AI Futures Project) | Aug 2026 | Median for an AI that can replace top coders late 2027; for superintelligence early 2029 (update) |
| Apr 2026 | Median for an AI that can replace top coders moved earlier, from late 2029 to mid-2028 (update) | |
| Jan 2026 | Median for AI that beats top experts at almost every task about the end of 2030 | |
| Apr 2025 | AI 2027: superintelligence by the end of 2027 as the single most likely story; his median was already 2028 | |
| Paul Christiano | Oct 2023 | About 15% by 2030 and 40% by 2040 for AI far past human level, able to build a Dyson sphere |
| 2021 | About 8% that AI wins gold at the Math Olympiad by 2025, against Yudkowsky’s 16% or more. It happened in July 2025 | |
| Leopold Aschenbrenner | Jun 2024 | AGI by 2027 “strikingly plausible” (Situational Awareness) |
| Geoffrey Hinton | May 2023 | ”5 to 20 years but without much confidence”, after years of expecting 30 to 50 |
| Ajeya Cotra (Open Philanthropy) | Aug 2022 | Median for transformative AI moved from 2050 to 2040 (update) |
| Sep 2020 | Median about 2050, from comparing AI to the computation done by brains and evolution | |
| Eliezer Yudkowsky | Jan 2017 | Bet Bryan Caplan that unaligned AI will have ended the world by 2030 (the bet) |
| Ray Kurzweil | 1999, restated 2024 | Human-level AI by 2029, the Singularity by 2045 |
Skeptics, and one who moved
| Who | When said | Prediction |
|---|---|---|
| François Chollet (creator of Keras and the ARC-AGI benchmark) | 2025 | AGI in about five years, down from about ten a year earlier, though he still thinks scaling language models alone won’t get there |
| Jun 2024 | Language models are “an off-ramp on the path to AGI” | |
| Yann LeCun (Meta, then AMI Labs) | Oct 2024 | Human-level AI will take years, “if not a decade”; today’s language models are not the route to it |
| Arvind Narayanan and Sayash Kapoor | Apr 2025 | AI as Normal Technology: the effects will arrive over decades, as with electricity |
Crowds and surveys
| Forecast | When | Median |
|---|---|---|
| Metaculus community, “Date of AGI” | 2020 → late 2022 → 2026 | About 2055 → about 2037 → about 2030 |
| AI researchers surveyed by AI Impacts (2023 survey) | 2016 → 2022 → 2023 | About 2061 → about 2060 → 2047, for machines that outdo people at every task |
Eddie
| Who | When said | Prediction |
|---|---|---|
| Eddie | Aug 2023 | 30% by 2033, 55% by 2043, 80% by 2053 (My timelines for AGI, after GPT4 came out) |
The scenarios
Forecasts give a date; scenarios tell a story of how it could happen, which makes them easier to argue with.
- Situational Awareness (2024) — the trendline case for AGI by 2027, followed by an intelligence explosion and a US–China race. Read first; the other two respond to the world it describes.
- AI 2027 (2025) — a month-by-month forecast to superintelligence, with a race ending and a slowdown ending. Its authors’ own forecasts have moved several times since.
- AI 2040 (2026) — the same group’s recommendation: a verified US–China deal that delays superintelligence to 2040.
Why forecasters disagree
- Trendlines or bottlenecks? Short timelines extrapolate steady growth in compute, algorithms and what models can do. Longer ones expect power, chips, data, money or real-world testing to slow the curve.
- Benchmarks or real work? Models keep beating tests faster than they replace people at jobs. Which one measures progress toward AGI is disputed.
- Is something missing? Skeptics such as LeCun think current methods lack something essential, such as learning from the physical world. Others think scale and better training will get there.
- What happens at the end? If AI can automate AI research, the last stretch could compress sharply. See Intelligence explosion and AI takeoff. The step from AGI to superintelligence could be short and A feedback loop is not a speed estimate argue opposite sides.
- Who is forecasting? Lab leaders have reasons to sound close, and to sound cautious. Surveyed researchers answer quickly and inconsistently. Forecasting communities update fast but share the same information.
Why the date matters
A risk decades away and a risk a few years away call for different responses. Short timelines leave less time for alignment and for agreements such as a US–China deal, and make pausing more urgent. For many decisions, though, the difference between 2030 and 2040 matters less than whether the risk is real at all. The case for AI risk, step by step places timelines among the other steps of the argument.