概述
沒有人確切知道這些趨勢將走向何方。了解實際數字、專家的預測以及它們的不確定性有助於任何人將證據與炒作區分開來,並為幾種可能的結果做好計劃。
重點摘要
- Epoch AI 估計,自 2020 年以來,用於訓練前沿語言模型的計算量每年增長約五倍。
- 史丹佛人工智慧指數 2025 報告稱,2022 年 11 月至 2024 年 10 月期間,GPT-3.5 級人工智慧的使用成本降低了 280 倍以上。
- 2024 年對 2,778 名人工智慧研究人員進行的一項調查顯示,到 2047 年,機器在每項任務上都有 50% 的機會勝過人類。
- 國際貨幣基金組織估計,全球近 40% 的就業機會接觸人工智慧,而接觸意味著任務發生變化,但不一定會導致工作消失。
- IEA 預計資料中心用電量將增加一倍以上,從 2024 年的約 415 太瓦時增至 2030 年的約 945 太瓦時。
- 關於人工智慧的預測是根據既定假設進行推斷,因此有用的習慣是檢查假設並針對幾個場景進行規劃。
深入探討
Talking about the future of AI means talking about trends and uncertainty at the same time. Some things can be measured: how much computing power is used to train models, how much it costs to use them, how well they do on tests, and how many companies adopt them. Other things cannot be measured yet: whether progress will continue at the same speed, how jobs will change, and whether the most powerful systems can be kept safe and under control. A good way to think about the future is to know the trend lines, know who is forecasting what, and keep track of the assumptions behind each forecast.
The engine behind recent progress is scale. AI models learn from data using specialized computer chips, and researchers have found that larger models trained on more data with more computing power tend to perform better. Epoch AI, a research group that tracks these figures, estimates that the computing power used to train frontier language models has grown about five times per year since 2020, doubling roughly every five months, while the cost of those training runs has grown about 3.5 times per year. The Stanford AI Index 2025 reports a similar doubling time of about five months for training compute. At the same time, using AI has become much cheaper: the AI Index finds that the cost of querying a model as capable as GPT-3.5 fell more than 280-fold between November 2022 and October 2024.
Capabilities are also being tracked in more practical terms. A 2025 study titled Measuring AI Ability to Complete Long Software Tasks introduced a metric called the time horizon: the length of task, measured by how long it takes a skilled human, that an AI system can complete half the time. For frontier models in early 2025 that horizon was about 50 minutes, and it had been doubling roughly every seven months since 2019. The authors are careful to say that if the trend continued and generalized to real work, AI systems within five years could automate many software tasks that now take people a month. That is a conditional projection, not a promise, and it is the right way to read most forecasts.
Adoption is a second set of indicators, separate from raw capability. The Stanford AI Index 2025 reports that 78% of surveyed organizations used AI in 2024, up from 55% a year earlier, and that private investment in generative AI reached $33.9 billion in 2024, up 18.7% from 2023. AI is also leaving the lab in regulated, physical settings: the same report counts 223 AI-enabled medical devices approved by the US Food and Drug Administration in 2023, up from six in 2015, and Waymo providing more than 150,000 driverless rides each week. These numbers measure use and spending rather than benefit, and adoption can stall when systems prove unreliable or expensive in practice. Still, they show that questions about the future of AI are no longer only about research labs; they now involve hospitals, roads, offices and the people who work in them.
Experts disagree widely about long-range outcomes. One of the largest surveys, Thousands of AI Authors on the Future of AI, asked 2,778 researchers who had published at top AI conferences. Their combined forecast gave a 50% chance that unaided machines could outperform humans at every task by 2047, thirteen years earlier than the same survey found in 2022, but a 50% chance of all human occupations being fully automated only by 2116. The gap between those two dates, and the large spread of individual answers, shows how uncertain the field is about its own future. A fast trend in the past is not proof of where it ends.
The effects on work are expected to be uneven. The International Monetary Fund estimated in January 2024 that almost 40% of global employment is exposed to AI, meaning AI could perform or assist with a meaningful share of those jobs' tasks. Exposure is about 60% in advanced economies, about 40% in emerging markets and about 26% in low-income countries. The IMF expects roughly half of the exposed jobs in advanced economies to benefit from higher productivity, while the rest could see lower demand for labor. Exposure is not the same as job loss; it describes which tasks change, and outcomes depend on how employers, workers and governments respond.
Energy and physical infrastructure are becoming a constraint. The International Energy Agency's 2025 report Energy and AI estimates that data centers used about 415 terawatt-hours (TWh) of electricity in 2024, about 1.5% of world demand, and projects that this will more than double to around 945 TWh by 2030, with AI as the most important driver. A typical AI-focused data center uses as much electricity as about 100,000 households. Where that power comes from, and how fast grids can expand, will shape where and how quickly AI grows.
Rules are arriving in stages. The European Union's AI Act, a broad law covering AI systems sold or used in the EU, entered into force on August 1, 2024. Bans on certain practices applied from February 2, 2025, obligations for general-purpose AI models from August 2, 2025, and most remaining provisions from August 2, 2026, while several high-risk obligations now apply from December 2027 or August 2028. On safety, the International AI Safety Report, published in January 2025 by 100 experts with an advisory panel nominated by 30 countries plus the UN, OECD and EU, gathered the evidence on capabilities and risks of advanced AI and stressed how much remains uncertain.
For individuals and organizations, preparing for an uncertain future is mostly practical. Learn what current tools can and cannot do, because capabilities and prices change within months. Watch measurable indicators rather than announcements: independent benchmark results, cost per task, real adoption and incident reports. Plan for more than one scenario, faster progress, slower progress and uneven progress across different kinds of tasks, and favor decisions that make sense in all of them, such as building skills in judging AI output, protecting data, and keeping a human responsible for important decisions.
04工作範例
透過六個步驟對 AI 預測進行壓力測試
從新聞中選取一個說法,例如:到 2030 年,人工智慧將處理大部分客戶服務。
要準確地寫下什麼是真實的:哪些國家、接觸的比例是多少、由誰、在什麼日期來衡量。
尋找背後的趨勢數據,例如史丹佛人工智慧指數中的採用數據或 2025 年時間範圍研究中的任務長度趨勢。
列出索賠所需的假設,例如成本持續下降、客戶接受度、法規允許以及複雜案例的足夠可靠性。
寫出一個更快、更慢和不均勻的場景,並為每個場景記下一個指標,以顯示正在發生的情況。
設定一個日期來檢查指標並記錄原始索賠是否步入正軌、提前還是落後。
它顯示了什麼
學習者最終會得到一個可測試的預測版本、其關鍵假設和一個需要觀察的指標的簡短清單。檢查的是中立讀者是否可以在審查日期查看相同的指標並就聲明是否成立達成一致。
現實世界的實施
工作與薪資:國際貨幣基金組織在 2024 年 1 月估計,全球近 40% 的就業機會受到人工智慧的影響,在已開發經濟體中這一比例上升到約 60%,其中大約一半的就業機會可能會受益。
電力:國際能源總署的 2025 年能源和人工智慧報告估計,到 2024 年,資料中心的用電量約為 415 太瓦時,約佔世界電力的 1.5%,並預計到 2030 年約為 945 太瓦時。
法規:歐盟《人工智慧法案》於2024年8月1日生效;對某些做法的禁令從 2025 年 2 月 2 日起實施,大多數剩餘規則從 2026 年 8 月 2 日起實施。
軟體任務:測量人工智慧代理的研究人員發現,自 2019 年以來,他們以 50% 的成功率完成的軟體任務的長度大約每七個月就增加一倍。
專家期望:2024 年 1 月發布的一項針對 2,778 名已發表的人工智慧研究人員的調查顯示,到 2047 年,機器在每項任務上都有 50% 的機會勝過人類,比 2022 年的調查早了 13 年。
人工智慧的未來:趨勢、預測和懸而未決的問題
幾個懸而未決的問題將決定哪些預測成立。首先,擴展是否能持續帶來回報。 Epoch AI 的趨勢數據顯示,自 2020 年以來,前沿模型的訓練運算量每年增長約五倍,但成本、晶片供應和電力也隨之增長,國際能源總署預計到 2030 年資料中心需求將達到約 945 TWh。 其次,測試的收益是否轉化為可靠的現實工作; 2025 年時間範圍研究的作者指出,他們的軟體任務比混亂的工作更乾淨,因此 7 個月的翻倍可能不會持續下去。第三,勞動市場如何調整;國際貨幣基金組織 2024 年的分析將結果視為取決於培訓和社會安全網等政策選擇。第四,更強大、更自主的系統的安全和控制,2025 年 1 月的國際人工智慧安全報告將其描述為一個深度科學不確定性的領域。第五,治理:歐盟人工智慧法案的分階段期限為2026年至2028年,這將考驗規則是否能跟上產品的步伐,其他國家正在編寫自己的方法。讀者可以透過一些經常出現的來源來追蹤這些問題,例如年度史丹佛人工智慧指數、Epoch AI 的趨勢儀表板、國際能源總署的能源分析和官方監管時間表,並且應該將任何單一預測視為多種情景中的一種。
付諸實踐
為什麼這很重要
- 風險與安全. 災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
- 更明確的決策. 民眾和專業素養決定強而有力的安全政策在政治上是否可行。
- 突破炒作. 清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
留意
- 將存在風險視為科幻小說,同時能力複合。
- 混淆了表面產品安全與高度自治下的對準。
- 只給非英語和非專業觀眾留下低品質的資源。
後續步驟
- 單獨的產品危害、誤用和失控/失調風險。
- 詢問哪些證據會改變您對時間表和嚴重性的看法。
- 比起行銷主張,更喜歡主要來源和具體評估。
- 確定一條行動路徑:職業、政策、資金或技能——而不僅僅是意識。
資料來源與延伸閱讀
- Stanford Institute for Human-Centered AI2025年人工智慧指數報告
- Epoch AI機器學習趨勢
- arXiv (Kwa et al., 2025)衡量人工智慧完成長軟體任務的能力
- arXiv (Grace et al., 2024)數以千計的人工智慧作者探討人工智慧的未來
- International Monetary Fund人工智慧將改變全球經濟。讓我們確保它造福人類
- International Energy Agency能源與人工智慧:執行摘要
- Future of Life Institute (EU AI Act explorer)歐盟人工智慧法案實施時間表
- arXiv (Bengio et al., 2025)國際人工智慧安全報告
不斷探索
常見問題
人工智慧的未來是什麼:趨勢、預測和懸而未決的問題?
人工智慧的未來是一系列可衡量的趨勢,例如快速增長的運算能力、下降的使用成本和處理更長任務的人工智慧系統,以及有關就業、能源、安全和法律的真正開放的問題。沒有人確切知道這些趨勢將走向何方。了解實際數字、專家的預測以及它們的不確定性有助於任何人將證據與炒作區分開來,並為幾種可能的結果做好計劃。
有用的人工智慧預測應該說明什麼?
該指南稱,預測應說明假設、範圍和證據。
受控基準得到改善。這對產品的廣泛採用有何意義?
該指南將受控結果與產品可用性和採用分開。
為什麼要使用更快、更慢、進度不均勻的場景?
該指南推薦了不確定性較大且任務進度可能不均勻的場景。
哪個指標比發佈公告更好地更新能力評估?
該指南將公告與獨立確認區分開來,並建議可重複的效能。
為什麼要記錄預測以及什麼算是未預測?
該指南建議保留預測和標準,以避免移動球門柱。