尊敬的用戶您好，這是來自FT中文網的溫馨提示：如您對更多FT中文網的內容感興趣，請在蘋果應用商店或谷歌應用市場搜尋「FT中文網」，下載FT中文網的官方應用。
There was nothing new in AI in 2024 that matched the sheer “wow” factor of using ChatGPT for the first time, but rapid improvements in the underlying technology still kept the field humming. For 2025, this is how I see things panning out.
2024年，人工智慧領域沒有出現能與首次使用ChatGPT時的「驚豔」效果相媲美的新事物，但底層技術的快速進步仍然使該領域保持活力。對於2025年，我是這樣預測的。
Will AI development hit a wall?
人工智慧的發展會遇到瓶頸嗎？
In 2025, that momentum will fade. Even some of the tech industry’s biggest optimists have conceded in recent weeks that simply throwing more data and computing power into training ever-larger AI models — a reliable source of improvement in the past — is starting to yield diminishing returns. In the longer term, this robs AI of a dependable source of improvement. At least in the next 12 months, though, other advances should more than take up the slack.
到2025年，這種勢頭將會減弱。即使是科技行業中最樂觀的一些人士，最近幾周也承認，僅僅透過投入更多的數據和計算能力來訓練更大的AI模型——這一過去可靠的改進來源——開始出現收益遞減的現象。從長遠來看，這剝奪了AI一個可靠的改進來源。不過，至少在接下來的12個月裏，其他進展應該能夠彌補這一不足。
The most promising developments look like coming from models that carry out a series of steps before returning an answer, allowing them to query and refine their first responses to deliver more “reasoned” results. It is debatable whether this is really comparable to human reasoning, but systems like OpenAI’s o3 still look like the most interesting advance since the emergence of AI chatbots.
最有前景的發展似乎來自於那些在給出答案之前執行一系列步驟的模型，這些模型可以查詢並完善其初始響應，以提供更「有理有據」的結果。雖然這是否真正可與人類推理相媲美尚有爭議，但像OpenAI的o3這樣的系統仍然看起來是自AI聊天機器人出現以來最有趣的進步。
Google, which regained its AI mojo late in the year after spending two years struggling to catch up with OpenAI, also showed how the new agent-like capabilities in AI could make life easier, such as tracking what you do in your browser and then offering to complete tasks for you. All these demos and prototypes still need to be turned into useful products, but they at least show that there is more than enough in the labs to keep the AI hype going.
谷歌(Google)在花了兩年時間努力追趕OpenAI之後，於年底重新找回了其AI的活力，還展示了AI中新的類代理功能如何讓生活更輕鬆，比如跟蹤你在瀏覽器中的操作，然後爲你完成任務。所有這些演示和原型仍需轉化爲有用的產品，但至少表明實驗室中有足夠的內容來維持AI的熱潮。
Will AI’s ‘killer app’ emerge?
人工智慧的「殺手級應用」會出現嗎？
For most people, the rise of generative AI has meant constantly seeing prompts offering to complete your writing for you or edit your photos in ways you hadn’t thought of — unsought, occasionally useful tools that fall well short of transforming your life.
對於大多數人來說，生成式人工智慧的興起意味著不斷看到提示，提供爲你完成寫作或以你未曾想到的方式編輯照片的選項——這些是未曾尋求但偶爾有用的工具，遠未達到改變生活的程度。
Next year is likely to bring the first demonstrations of apps that can intervene more directly: Absorbing all your digital information and learning from your actions so that they can act as virtual memory banks or take over entire aspects of your life. But, concerned about the unreliability of the technology, tech companies will be wary about rushing these out for mass use — and most users will be equally wary about trusting them.
明年可能會首次展示能夠更直接干預的程式：這些程式可以吸收你所有的數字資訊，並從你的行爲中學習，從而充當虛擬記憶庫或接管你生活的各個方面。然而，由於擔心技術的不可靠性，科技公司會謹慎地將這些程式推向大衆使用，而大多數用戶也會對信任這些程式持謹慎態度。
Instead of true killer apps for AI, this means we’ll be left in the “AI in everything” world that technology users have already become accustomed to: Sometimes intrusive, sometimes helpful, and still not quite providing the really new experiences that would prove the AI era has truly arrived.
這意味著我們將處於技術用戶已經習慣的「AI無處不在」的世界，而不是擁有真正的AI殺手級應用：有時令人反感，有時有幫助，但仍未能提供真正新穎的體驗，以證明AI時代的真正到來。
Will Nvidia’s GPUs still rule the tech world?
輝達的GPU還能繼續主導科技界嗎？
The chipmaker’s huge profits have made it the target of the most powerful tech companies, most of which are now designing their own AI chips. But Nvidia has been moving too fast for rivals, and while a quarter or two could be bumpy as it goes through a major product transition, its Blackwell product cycles should carry it through the year comfortably ahead.
這家晶片製造商的鉅額利潤使其成爲最強大科技公司瞄準的目標，其中大多數公司現在都在設計自己的AI晶片。然而，輝達(Nvidia)的發展速度太快，競爭對手難以追趕。儘管在經歷重大產品過渡時，可能會有一兩個季度的波動，但其Blackwell產品週期應能使其在全年中輕鬆保持領先地位。
That doesn’t mean others won’t make inroads. According to chipmaker Broadcom, three of the biggest tech companies are to use their in-house chip designs for supercomputing “clusters” with 1mn chips each in 2027. That is 10 times the size of Elon Musk’s Colossus system, thought to be the largest cluster of AI chips currently in use.
這並不意味著其他公司不會取得進展。根據晶片製造商博通的說法，到2027年，三家最大的科技公司將使用其內部設計的晶片來構建超級計算「集羣」，每個集羣將包含100萬顆晶片。這是伊隆•馬斯克(Elon Musk)的Colossus系統規模的10倍，該系統被認爲是目前使用中的最大AI晶片集羣。
Even as its market share starts to erode, though, Nvidia’s software still represents a considerable moat for its business, and by the end of the year it should be on the verge of another important new product cycle.
儘管市場份額開始侵蝕，輝達的軟體仍然是其業務的重要護城河。到今年年底，它應該會處於另一個重要的新產品週期的邊緣。
Will the stock market’s AI boom continue?
股市的人工智慧熱潮會繼續嗎？
With Big Tech in the midst of an AI race that its leaders believe will determine the future shape of their industry, one of the main forces behind the AI capital spending boom will remain in place. Also, as some companies start to claim big — if unproven — results from applying the technology in their own businesses, many others will feel they have to keep spending, even if they haven’t worked out yet how to use AI productively.
在大型科技公司(Big Tech)處於一場被其領導者認爲將決定行業未來形態的人工智慧競賽中時，推動人工智慧資本支出激增的主要力量之一將繼續存在。此外，隨著一些公司開始聲稱在其自身業務中應用該技術取得了巨大但未經證實的成果，許多其他公司即使尚未弄清如何有效利用人工智慧，也會感到必須繼續投入。
Whether this is enough for investors to keep throwing their money at AI is another matter. That will depend on other factors, such as the stock market’s confidence in the deregulatory and tax-cutting intentions of the new Trump administration and the readiness of the Federal Reserve to continue with monetary policy easing.
投資者是否會繼續將資金投入人工智慧是另一回事。這將取決於其他因素，例如股市對川普政府放松管制和減稅意圖的信心，以及美聯準是否準備繼續實施貨幣政策寬鬆。
It all points to a highly volatile year, with some big corrections along the way. But with enough liquidity, Wall Street could succumb to AI hype for some time yet.
這一切都表明，今年將是一個高度動盪的年份，期間可能會出現一些大的調整。但如果流動性充足，華爾街可能會在一段時間內屈從於人工智慧的炒作。