AGI 何時到來?為何每項預測的終點線都不同
When will AGI happen? Why every forecast names a different finish line
AGI 到來時間沒有定論,各項預測採用不同能力定義、機率和評估方式,因此估算日期差異很大。Metaculus 在 2026 年 11 月 5 日讀取的中位日期為 15 Sep 2030;AI Impacts 2024 調查對 HLMI 和 FAOL 的 50% 預測點分別是 2042 年及 2098 年,相差 56 年。
Nobody knows when AGI will happen, and published forecasts use different definitions. On 5 November 2026, Metaculus's general-AI question showed 15 September 2030; a December 2024 researcher survey's aggregate forecast reached 50% in 2042 for unaided machines doing every task better and more cheaply than humans. To compare answers to "when will AGI happen?", write down four things for each one: the target capability, the date, the probability attached to it, and how the result would be checked.
What does AGI mean?
AGI stands for artificial general intelligence, an AI system that can do a wide range of thinking work at least as well as people. Each group writes its own acceptance test for that requirement, and the tests differ a lot.
| Who | Term | Finish line in plain words |
|---|---|---|
| Dario Amodei, Anthropic | Powerful AI | Smarter than a Nobel Prize winner in most fields, works on its own for hours to weeks |
| Sam Altman, OpenAI | Superintelligence | Far beyond human ability; no formal definition given |
| AI Impacts survey | HLMI | Unaided machines do every task better and more cheaply than humans |
| AI Impacts survey | FAOL | Every occupation is fully automatable |
| DeepMind, Levels of AGI | A ladder | From Emerging (about an unskilled human) up to Superhuman (arXiv 2311.02462) |
When will AGI happen according to forecasts?
Metaculus is a forecasting site where people give a probability distribution over dates and the site combines them. Read live on 5 November 2026, its question about the first publicly announced general AI system showed a community median of 15 Sep 2030 (Metaculus Q5121). A median means half the forecast probability sits after that date.
Screenshot of the live forecast, read 2026-11-05, from Metaculus question 5121.
| Source, date said | Target | Forecast |
|---|---|---|
| Amodei, Oct 2024 (essay) | Powerful AI | Could come as early as 2026, could take much longer |
| Anthropic, Mar 2025 (statement) | Powerful AI | Late 2026 or early 2027 |
| Altman, Sep 2024 (essay) | Superintelligence | Possibly a few thousand days |
| Bengio, Jun 2023 (FAQ) | Superhuman intelligence | 5 to 20 years, 95% interval |
| Sutskever, Nov 2025 (podcast) | Human-like learning, then superhuman | 5 to 20 years |
| Metaculus Q5121, read 5 Nov 2026 | General AI, strict criteria | Median 15 Sep 2030 |
| AI Impacts survey, Dec 2024 | HLMI | 50% in 2042 |
| AI Impacts survey, Dec 2024 | FAOL | 50% in 2098 |
The survey asked authors from top AI venues and got 1,580 responses; Appendix E reports 985 valid fitted responses for HLMI and 456 for FAOL (ESPAI 2024). Its "mean aggregate" fits a curve to each respondent's three probability and year pairs, averages the curves, and reads where the average crosses 50%. Across rounds the HLMI date moved from 2061 (2016) to 2059 (2022), 2047 (2023) and 2042 (2024). HLMI and FAOL describe nearly the same end state, every task against every job, yet their 50% dates sit 56 years apart. The two questions went to different respondents with different wording and order, and the authors say the reasons for the gap remain unclear.
What can you measure? The METR time horizon
A forecast is a judgment. METR's time horizon is a measurement on a narrow slice of work: the length of a self-contained software task, in human expert minutes, that an AI agent finishes with 50% success. The March 2025 paper found it doubled about every 7 months since 2019 (arXiv 2503.14499). Time Horizon 1.1, in January 2026, grew the suite from 170 to 228 tasks, fitted a doubling time of about 131 days on data since 2023 and about 89 days since 2024, and listed Claude Opus 4.5 at about 320 minutes, with a confidence interval of 170 to 729 minutes (METR Time Horizon 1.1).
Credit: METR, from the time horizons dashboard.
Each task has a timed human baseline. The agent runs it several times from a fresh environment snapshot, inside a harness that calls the model, runs tools and appends results; context grows with every step because an agent re-sends its history on every call. A scorer marks each run pass or fail, and a logistic curve is fitted to all runs against task length:
p(t) = 1 / (1 + exp(k × ln(t / h)))
t human expert time for the task, in minutes
h the 50% time horizon: p(h) = 1 / (1 + exp(0)) = 0.5
k steepness (k > 0): how fast success falls as tasks get longer
Repeated attempts share task difficulty, so error bars should resample whole tasks with their runs kept together. Treating 8 runs of 12 tasks as 96 independent trials makes the interval look far tighter than the evidence supports. METR's original analysis also accounts for related task families through a hierarchical bootstrap over families, tasks and attempts.
Worked example: from 50% to 80% reliability
Take h = 320 minutes and assume k = 0.9 (k is our assumption for illustration, not a METR figure). Solving p(t) = 0.8:
ln(t / h) = ln(0.25) / 0.9 = -1.54
t = 320 × 0.214 ≈ 69 minutes
p(60 min) ≈ 82%
p(960 min) ≈ 27%
Under that assumption the 80% horizon is about a fifth of the 50% horizon. Chaining makes it worse: three independent steps that each pass half the time finish 0.5 × 0.5 × 0.5 = 12.5% of the time, which is why coding pipelines add a second model as reviewer with a round cap.
Your own pilot needs error bars too. A 95% Wilson interval for 15 passes in 30 trials runs from about 33.2% to 66.8%, so 30 trials cannot tell a 35% tool from a 65% one.
What the numbers do not tell you
METR warns that estimates above 16 hours are unreliable with its suite, that the tasks are mainly self-contained software, ML and security work, and that the results do not mean AI can automate all jobs (time horizons). METR has also revised its suite and corrected a measurement mistake in March 2026, so one model can carry different numbers in different versions. The survey had a response rate of about 10% and was collected in December 2024, before most 2025 and 2026 model releases. DeepMind's Levels of AGI paper placed some 2023 chatbots at Emerging AGI in an illustrative, approximate assessment that says nothing about 2026 systems. Each figure should travel with its version, collection date and definition.
Our take
For planning, measured capability is more useful than any AGI date. Under METR's conditions, agents finish self-contained software tasks that take an expert a few hours about half the time, and the trend has been rising. Give AI well-scoped tasks with a clear check, keep a reviewer on the output, and put lasting knowledge in files and docs the tool can read: with a fixed model, a correction lives only in the context, which is why a long chat can forget your name and why applications keep facts outside the model and retrieve the one line they need from 100 pages.
Anthropic's late 2026 to early 2027 window for powerful AI is the first lab forecast that the next few months will check. When you read an AGI claim, look for the test, who ran it and on what data.
Further reading
- Metaculus question 5121: the live forecast and its full resolution criteria.
- ESPAI 2024: survey method, HLMI and FAOL wording, response counts and the authors' limits.
- METR Time Horizon 1.1: how the horizon is measured, the doubling fits by period, and the error bars.
- Machines of Loving Grace: Amodei's definition of powerful AI, October 2024.
Forecasts here are predictions with their original dates, not claims of achieved capability. We describe each person's and company's view from their own published write-up or recording and make no claims about any company's private systems. The worked example uses our own assumed k. Figures were read on 2026-11-05.
Frequently asked questions
When will AGI happen?
Nobody knows. Metaculus showed 15 September 2030 on 5 November 2026, under that question's specific criteria.
Why do AGI forecasts differ?
They use different capability definitions, probabilities and forecasting methods.
Does a benchmark score establish AGI?
No single score establishes broad capability. Report the task scope, evaluation conditions and reliability.
Sources
- Metaculus: When will the first general AI system be devised, tested, and publicly announced?
- AI Impacts: 2024 Expert Survey on Progress in AI (ESPAI 2024, PDF)
- METR (January 2026): Time Horizon 1.1
- Kwa et al. (March 2025): Measuring AI Ability to Complete Long Tasks (arXiv 2503.14499)
- Dario Amodei (October 2024): Machines of Loving Grace
- Dwarkesh Podcast (November 2025): Ilya Sutskever
- Yoshua Bengio (June 2023): FAQ on Catastrophic AI Risks
- Morris et al.: Levels of AGI for Operationalizing Progress on the Path to AGI (arXiv 2311.02462)
來源:DEV Community · dev.to

