Here are the six AI trends that will matter most in 2026. This list is based on daily research and reports from institutions like McKenzie, OpenAI, Stanford, and from analysts who are much more

這裡是2026年最重要的六個AI趨勢。這份清單基於麥肯錫、OpenAI、史丹佛等機構的日常研究和報告，以及比我更有知識的分析師。

knowledgeable than myself. In other words, don't blame me if they get it wrong. For each trend, I'll first start with a big picture, then move on to the

換句話說，如果他們錯了，別怪我。對於每個趨勢，我會先從大局開始，然後轉到

actionable takeaways so that by the end, you have a clear sense of where AI is heading and what to do about it.

可操作的要點，這樣到最後，你對AI的走向和應該怎麼做有一個清晰的認識。

Let's get started. Kicking things off with trend number one. Models don't matter much anymore. For the past few years, every new model released sparked debate about the best AI, and for good

讓我們開始吧。首先是趨勢一：模型不再那麼重要了。過去幾年，每個新模型發布都會引發關於最佳AI的辯論，這是有道理的。

reason. The difference in quality between models was significant. In 2026, though, that choice is going to matter a lot less. Taking a look at the data, this graph from artificial analysis

模型之間的質量差異很大。但在2026年，這個選擇將變得不那麼重要。看一下數據，這張來自Artificial Analysis的圖

shows how the major AI models have improved over time. Notice the clustering in the top right corner. The models are still getting smarter in absolute terms, but the gap between them

顯示主要AI模型隨時間的進步。注意右上角的聚集。模型在絕對值上仍在變得更聰明，但它們之間的差距

keeps shrinking, meaning no single model has a clear lead anymore. A Stanford study confirms this from another angle by comparing closed models like Gemini and Chachi BT against openw weight

不斷縮小，意味著沒有任何單一模型有明顯的領先優勢了。史丹佛的一項研究從另一個角度證實了這一點，通過比較像Gemini和ChatGPT這樣的閉源模型與像DeepSeek和

alternatives like Deep Seek and Llama. The trend is pretty clear. Models that are free to run are now approaching frontier performance and performance is only half the story. The cost matters as

Llama這樣的開放權重替代品。趨勢很清楚：可以免費運行的模型現在正在接近前沿性能。而且性能只是故事的一半。成本同樣

well. Data from Epoch AI shows that using powerful models has become drastically cheaper and one of the reasons is because hardware is getting more efficient. For perspective, Nvidia's latest chips uses 105,000

重要。Epoch AI的數據顯示，使用強大模型已經變得大幅便宜，原因之一是硬件變得更高效。作為參考，Nvidia最新的芯片使用的

times less energy per token than they did 10 years ago. So, what does this mean for us? In plain English, when things get cheaper and more similar, they become more like commodities. You

每個token的能量比10年前少了105,000倍。那這對我們意味著什麼？簡單說，當東西變得更便宜、更相似時，它們就變得更像大宗商品。你

don't ask who provides the best electricity, right? You ask what can I use the electricity for? And because of this, the competition is shifting from the AI model itself to the way we

不會問誰提供最好的電力，對吧？你問的是我能用電力做什麼。正因如此，競爭正從AI模型本身轉移到我們

actually use it, aka the app layer. Just think about cars. Once the engine becomes standardized, the focus shifts to the features and the design. This creates an interesting dynamic for each

實際使用它的方式，也就是應用層。想想汽車。一旦引擎變得標準化，焦點就轉向功能和設計。這為每個

of the frontier AI labs. For example, OpenAI has a mind share advantage because ChachiBT is synonymous with AI and has the largest market share. Google has a distribution advantage because

前沿AI實驗室創造了有趣的動態。例如，OpenAI有心智份額優勢，因為ChatGPT是AI的同義詞，並且有最大的市場份額。Google有分發優勢，因為

Gemini is already embedded across its existing products like search, Gmail, and Android. Anthropic has a specialization advantage given its loyal customer base in developers and enterprise customers. Notice what's

Gemini已經嵌入到其現有產品中，如搜索、Gmail和Android。Anthropic有專業化優勢，因為它在開發者和企業客戶中有忠實的客戶群。注意

missing from that list. None of them are winning because they have the best AI.

這個清單中缺少什麼。他們都不是因為擁有最好的AI而獲勝。

The competition has moved beyond raw power to reach, integration, and trust.

競爭已經超越了原始性能，轉向了觸及範圍、整合和信任。

The practical takeaway here is to stop obsessing over technical scores and instead focus on how they fit into your actual work. For example, if you live in Google Workspace, Gemini's deep

這裡的實際要點是停止糾結於技術分數，而是專注於它們如何適合你的實際工作。例如，如果你生活在Google Workspace中，Gemini與所有

integration with all of Google's apps gives it an edge that has nothing to do with raw intelligence. By the way, I'll link all the sources I mentioned today

Google應用的深度整合給它帶來了與原始智能無關的優勢。順便說一下，我會在下方連結今天提到的所有資料來源，

down below so you can check them out for yourself. Trend number two, 2026 is the year of AI workflows, not AI agents. If you spend any time on Twitter or

這樣你可以自己查看。趨勢二：2026年是AI工作流的一年，而不是AI代理的一年。如果你在Twitter或

LinkedIn, you've probably noticed the industry jump from chat bots straight to autonomous agents and completely skip the middle step where the actual value is being unlocked, AI workflows. And the

LinkedIn上花任何時間，你可能已經注意到行業從聊天機器人直接跳到自主代理，完全跳過了中間步驟——真正價值正在被釋放的地方，AI工作流。而

numbers prove this. According to McKenzie, no more than 10% of organizations in any given business function report scaling true agents.

數據證明了這一點。根據麥肯錫，任何特定業務職能中報告規模化真正代理的組織不超過10%。

Meanwhile, we see from OpenAI's enterprise report that 20% of enterprise AI use is already happening through workflow specific tools like custom GBTs and projects.

同時，我們從OpenAI的企業報告中看到，20%的企業AI使用已經通過工作流特定工具（如自定義GPT和項目）發生。

This gap tells you the market has voted for workflows, not autonomy. And we're seeing this play out across industries.

這個差距告訴你市場投票給了工作流，而不是自主性。我們在各個行業都看到了這一點。

A pharma company redesigned their clinical study workflow by using AI to analyze raw clinical data while humans focus on validation leading to 60% less prep time and 50% fewer errors. A

一家製藥公司重新設計了他們的臨床研究工作流程，使用AI分析原始臨床數據，而人類專注於驗證，導致準備時間減少60%，錯誤減少50%。一家

utility company redesigned their call center workflow where AI handles authentication and routine inquiries cutting cost per call by 50% while increasing satisfaction scores by 6%. A bank redesigned their code migration

公用事業公司重新設計了他們的呼叫中心工作流程，AI處理身份驗證和常規諮詢，每通電話的成本降低了50%，同時滿意度分數提高了6%。一家銀行重新設計了他們的代碼遷移

workflow where AI scans legacy code and generates updated versions for developers to verify, cutting the required human hours by 50%. Andre Kaparthi sums it up perfectly, calling everything an agent creates unrealistic

工作流程，AI掃描舊代碼並為開發人員生成更新版本進行驗證，將所需人工小時減少了50%。Andrej Karpathy完美地總結了這一點——把所有東西都叫做代理會製造不切實際的

expectations and confusion. Fully autonomous AI still faces massive hurdles like data security. So, we're looking at the decade of agents, not the year.

期望和混亂。完全自主的AI仍然面臨著巨大的障礙，如數據安全。所以，我們看的是代理的十年，而不是代理的一年。

>> I was triggered by that because I feel like there's some overpredictions going on in the industry. And uh in my mind this is really a lot more accurately

我被那個觸動了，因為我覺得行業裡有一些過度預測。在我看來，這更準確地

described as the decade of agents. >> Meanwhile, by integrating something like custom GBTS into an existing workflow, we've essentially created an agent light which is much more reliable at producing

描述為代理的十年。同時，通過將像自定義GPT這樣的東西整合到現有工作流程中，我們本質上創建了一個輕量級代理，它在產生

consistent results. To really ram this point home, McKenzie predicts that redesigning workflows will unlock nearly $3 trillion in economic value by 2030.

一致結果方面更可靠。要真正強調這一點，麥肯錫預測重新設計工作流程到2030年將釋放近3萬億美元的經濟價值。

And more importantly, these organizations will have the muscle memory to adopt true AI agents faster when they finally arrive. So here's your practical takeaway. Your goal for 2026

更重要的是，這些組織將擁有在真正的AI代理最終到來時更快採用它們的肌肉記憶。所以這是你2026年的實際要點：

is to turn your successful prompts into repeatable workflows. And this is something I've talked about in other videos. Pick one recurring deliverable you produce, like a weekly report. Break

把你成功的提示變成可重複的工作流程。這是我在其他視頻中談過的。選擇一個你經常產出的可交付物，比如每週報告。把

it into steps and let AI handle the predictable parts. Keep yourself in the loop for the final judgment calls. That structure is what creates true reliability. Side note, I'm actually

它分解成步驟，讓AI處理可預測的部分。讓自己參與最終的判斷。那種結構才是創造真正可靠性的東西。順便說一下，我實際上正在

developing an entire course around evergreen AI skills to give you a future proof framework that never becomes obsolete. If you're interested in learning a practical and timeless AI

開發一個關於永恆AI技能的完整課程，給你一個永不過時的面向未來的框架。如果你有興趣學習一個實用且永恆的AI

system, click the link below to join the wait list. Trend number three, the end of the technical divide. When I was at Google, non-technical teams like sales and marketing had to rely on specialist

系統，點擊下方連結加入等待名單。趨勢三：技術鴻溝的終結。當我在Google時，像銷售和行銷這樣的非技術團隊不得不依賴專家

teams to help them build stuff like dashboards. And I'm not someone who holds grudges, but a lot of my requests were depprioritized because they were too low impact and my clients weren't

團隊來幫助他們建立像儀表板這樣的東西。我不是一個記仇的人，但我的很多請求被降低優先級，因為它們影響太小，而且我的客戶不是

key accounts, but no, I'm over it. Anyways, in 2026, that's going to happen a lot less. The numbers backing this are honestly kind of insane. According to Open Eyes latest report, 75% of

重點客戶，但不，我已經釋懷了。總之，在2026年，這種情況會大大減少。支持這一點的數據老實說有點瘋狂。根據OpenAI的最新報告，75%的

enterprise users reported using AI to complete tasks they literally could not do before. Not just doing old tasks faster, they're doing entirely new things. For example, coding related messages from non-technical employees

企業用戶報告使用AI完成他們之前根本做不到的任務。不僅僅是更快地做舊任務，他們在做全新的事情。例如，來自非技術員工的編碼相關消息

grew 36% in just 6 months. These are salespeople, marketers, and operations managers writing scripts, automating spreadsheets, and building internal tools. A study from MIT confirms this.

在短短6個月內增長了36%。這些是銷售人員、行銷人員和運營經理在編寫腳本、自動化電子表格和構建內部工具。MIT的一項研究證實了這一點。

AI acts as an equalizer, disproportionately helping workers with lower technical skills close the performance gap with experts.

AI充當均衡器，不成比例地幫助技術技能較低的工人縮小與專家的績效差距。

Here's what all this means for your career. If your value is purely technical, aka you're the dashboard person, then your competitive advantage is shrinking because the marketing

這對你的職業生涯意味著什麼。如果你的價值純粹是技術性的——也就是說你是那個做儀表板的人——那麼你的競爭優勢正在縮小，因為以前

manager who used to wait in your queue can now do it themselves. But if you are that marketing manager or the salesperson who deeply understands their clients, then this is the biggest

在你隊列中等待的行銷經理現在可以自己做了。但如果你是那個深深了解客戶的行銷經理或銷售人員，那麼這是你職業生涯中最大的

opportunity of your career because the technical barrier that stood between your expertise and your execution is now gone. Here's your practical takeaway.

機會，因為站在你的專業知識和執行之間的技術障礙現在消失了。這是你的實際要點：

Attempt one impossible task this month. Identify a technical project you usually outsource like building a dashboard, cleaning a messy data set, or automating a report and try doing it yourself using

這個月嘗試一個不可能的任務。找出一個你通常外包的技術項目——比如建立儀表板、清理混亂的數據集或自動化報告——然後嘗試自己使用

Gemini Cloud or Cashibbt. You'll be surprised by what you can now pull off alone. Moving on to trend number four, from prompting to context. One of my most popular videos is this one teaching

Gemini、Claude或ChatGPT來做。你會驚訝於你現在獨自能完成什麼。接下來是趨勢四：從提示到上下文。我最受歡迎的視頻之一是這個教

you how to prompt because as we all know, if we don't phrase our request well, we get a bad result from AI.

你如何提示的視頻，因為我們都知道，如果我們沒有很好地表達我們的請求，我們會從AI得到糟糕的結果。

Unfortunately for me, that video is going to matter a lot less in 2026 because new models have gotten so much better at understanding vague instructions. However, they still have

不幸的是對我來說，那個視頻在2026年將變得不那麼重要，因為新模型在理解模糊指令方面已經變得好得多。然而，它們仍然有

one massive weakness I call the fact gap. While models know almost everything on the public internet from Shakespeare to Python code, they know nothing about your Q3 goals, your brand guidelines, or

一個我稱之為事實差距的巨大弱點。雖然模型幾乎知道公共互聯網上的一切，從莎士比亞到Python代碼，但它們對你的Q3目標、你的品牌指南或

that email your boss sent yesterday. It's like having a brilliant employee who technically knows how to complete tasks, but isn't allowed to look at any company files. they're still going to

你老闆昨天發的那封郵件一無所知。這就像有一個聰明的員工，技術上知道如何完成任務，但不被允許查看任何公司文件。他們還是會

fail, right? Because they lack context. At least that's what I told my boss during my first internship. It's the exact same thing with AI. The focus has

失敗，對吧？因為他們缺乏上下文。至少這是我在第一次實習時告訴老闆的。AI也是完全一樣的道理。焦點已經

shifted from how we ask the wording to what we give it, the context. And this explains the platform wars we're seeing right now. Google, Microsoft, and others are racing to embed AI into their

從我們怎麼問（措辭）轉移到我們給它什麼（上下文）。這解釋了我們現在看到的平臺戰爭。Google、微軟和其他公司正在競相將AI嵌入到他們的

productivity suites because whoever holds your context, your emails, your docs, your calendar, holds your attention. This is also how they'll trap you with platform lockin. The more context you build up in one ecosystem,

生產力套件中，因為誰掌握了你的上下文——你的郵件、文檔、日曆——誰就掌握了你的注意力。這也是他們將如何用平臺鎖定來困住你的方式。你在一個生態系統中積累的上下文越多，

the smarter the AI is for you and the harder it becomes to switch. There are two practical takeaways here and the non-productivity people are going to hate this. First, file management is no

AI對你就越聰明，切換就變得越難。這裡有兩個實際要點，不喜歡生產力的人會討厭這個。第一，文件管理不再

longer optional. To get value from AI, you need some sort of system to keep your files organized and clearly named.

是可選的。要從AI獲得價值，你需要某種系統來保持文件有組織且命名清晰。

If your work is scattered in random, unnamed folders, you can't point the AI to it. Second, audit where your information lives. If it's spread across three or four different platforms, you

如果你的工作散布在隨機的、未命名的資料夾中，你就無法指向它讓AI處理。第二，審核你的資訊存放在哪裡。如果它分散在三四個不同的平臺上，你

need to consolidate. If your resume lives in Google Drive, but the job description and interview notes are stored in Notion, neither Gemini nor Notion AI can help with interview prep,

需要整合。如果你的履歷在Google Drive裡，但工作描述和面試筆記存在Notion裡，Gemini和Notion AI都無法幫助準備面試，

you end up doing the synthesis manually, which leads to more friction and defeats the whole purpose. So, as a rule of thumb, prompting still matters, but it's more important to ask yourself, does the

你最終會手動進行綜合，這會導致更多摩擦並違背了整個目的。所以，作為經驗法則，提示仍然重要，但更重要的是問自己：

AI have the files it needs to know what I'm talking about? Trend number five, advertising is coming to chat bots, and it's not all bad. First of all, please

AI有它需要的文件來理解我在說什麼嗎？趨勢五：廣告即將進入聊天機器人，而這並不全是壞事。首先，

don't shoot the messenger on this one. Hear me out. At this point, it's basically been confirmed that ads are coming to CHACHBT in 2026. So, instead of debating if it will happen, let's

請不要射殺信使。聽我說完。到這個時候，基本上已經確認廣告將在2026年進入ChatGPT。所以，與其爭論它是否會發生，讓我們

talk about the implications. Imagine a world where advertising never comes to chatbots. In that reality, the best AI models stay locked behind expensive subscriptions, creating a wealth gap,

談談影響。想像一個廣告永遠不會進入聊天機器人的世界。在那個現實中，最好的AI模型被鎖在昂貴的訂閱後面，創造一個財富差距，

where only those who can pay have access to the best tools, while everyone else is left with an inferior version. Over time, this creates a compounding advantage. The wealthy use powerful AI

只有那些能付費的人才能訪問最好的工具，而其他人都只能用較差的版本。隨著時間推移，這會創造複合優勢。富人使用強大的AI

to get wealthier while everyone else falls further behind. Kind of reminds me of something I just can't put my finger on. It think of it like YouTube. Imagine

變得更富，而其他人則進一步落後。有點讓我想起了什麼，但我說不上來。想想YouTube。想像一下

if you couldn't watch videos from the top creators unless you pay for YouTube Premium. That is where AI is headed without an ad supported tier. Now that

如果你不付費訂閱YouTube Premium就看不到頂級創作者的視頻。沒有廣告支持的等級，AI就會走向那個方向。現在

we know ads are inevitable and that I'm not to blame for this, uh the thing to watch is what format those ads will take because it's going to look very

我們知道廣告是不可避免的，而且這不是我的錯，要注意的是那些廣告會採用什麼形式，因為它們看起來會非常

different from the search ads we're currently used to. For example, industry expert Eric Sufer predicts chatbot ads will not be tied to our specific questions because if an AI recommended a

不同於我們目前習慣的搜索廣告。例如，行業專家Eric Sufer預測聊天機器人廣告不會與我們的具體問題掛鉤，因為如果AI在回答中直接

product directly in its answer, we wouldn't trust it. Instead, the ad will probably look like standard display banners that stay separate from your actual conversation. Sort of like the

推薦產品，我們不會信任它。相反，廣告可能會看起來像標準的展示橫幅，與你的實際對話分開。有點像

banner ads we see on websites today. So, here's the bottom line. I don't like ads. You don't like ads. Nobody likes ads. But it's the ad revenue that makes

我們今天在網站上看到的橫幅廣告。所以，這是底線。我不喜歡廣告。你不喜歡廣告。沒有人喜歡廣告。但正是廣告收入讓

it possible for companies to offer their best models to students in developing countries, nonprofits, and casual users who can't afford another monthly bill.

公司能夠向發展中國家的學生、非營利組織和無法承受另一筆月費的普通用戶提供他們最好的模型。

Trend number six, from chatbots to robots. Everything we've covered so far has focused on AI as software. But in 2026, that software is going to appear even more in the physical world as

趨勢六：從聊天機器人到機器人。我們到目前為止涵蓋的一切都專注於AI作為軟體。但在2026年，那個軟體將更多地出現在物理世界中，作為

physical agents who can move on their own. The numbers show this is already happening. Exhibit A, Whimo. Their autonomous taxi service has now logged over 100 million fully autonomous miles

能夠自主移動的物理代理。數據顯示這已經在發生。證據A：Waymo。他們的自動駕駛出租車服務現在已經記錄了超過1億英里的完全自動駕駛里程，

and are involved in 96% fewer crashes than human drivers. Exhibit B, Amazon.

事故率比人類司機低96%。證據B：亞馬遜。

Their AI enabled warehouse robots have cut the time from order to shipping by 78%. Exhibit C, China. As early as 2023, China had deployed more industrial

他們的AI倉庫機器人將從下單到發貨的時間縮短了78%。證據C：中國。早在2023年，中國部署的工業

robots than the US and the rest of the world combined. Now, there is one caveat to all this. Humanoid robots are still mostly hype.

機器人就比美國和世界其他地方的總和還多。現在，這一切有一個警告。人形機器人仍然主要是炒作。

MIT robotics professor Rodney Brooks estimates that we are at least 15 years away from seeing functional humanoid robots in our daily lives. The real shift is what analyst Mary Miker calls

MIT機器人教授Rodney Brooks估計，我們至少還有15年才能在日常生活中看到功能性人形機器人。真正的轉變是分析師Mary Meeker所說的

AI turning capital assets into software endpoints. And here's what that means in plain English. A car, a tractor, or warehouse robot used to be a depreciating asset, which means it loses

AI將資本資產轉化為軟體端點。這用白話說是什麼意思：一輛車、一臺拖拉機或倉庫機器人過去是一種折舊資產，意味著它會隨時間

value as time goes on. Right now, these machines are becoming platforms that improve over time through software updates, exactly like our phones. A Whimo car today is actually safer and

貶值。現在，這些機器正在成為通過軟體更新隨時間改進的平臺，就像我們的手機一樣。今天的Waymo汽車實際上比

smarter than it was 2 years ago, even if the physical vehicle hasn't changed. So, what does all this mean for us? In a nutshell, while the headlines are focusing on white collar disruption for

兩年前更安全、更聰明，即使物理車輛沒有改變。那這一切對我們意味著什麼？簡而言之，雖然頭條新聞目前專注於白領工作的顛覆，

now, this trend suggests that blue collar work will also be disrupted, but over a much longer time horizon. On a more positive note, I want to leave you

但這一趨勢表明藍領工作也會被顛覆，但會在一個更長的時間範圍內。在更積極的一面，我想用

with something Ethan Mollik said. He's a professor at Wharton, and this is something I really believe in. His research on what he calls the jagged frontier of AI shows that right now we

Ethan Mollick說過的話結束。他是沃頓商學院的教授，這是我真正相信的。他關於AI「參差不齊前沿」的研究表明，現在我們

are in a unique window where expertise is being reset thanks to AI. And precisely because things are messy and undefined right now, there are no experts who know everything already. You

正處於一個獨特的窗口，專業知識正在因為AI而被重置。正是因為現在事情是混亂和未定義的，沒有專家已經什麼都知道。你

just need to be willing to learn faster than the person next to you. That is how you win in 2026.

只需要願意比你旁邊的人學得更快。這就是你在2026年獲勝的方式。

Stop worrying about developing a perfect plan to learn AI and instead just get started. I'd love to hear your thoughts on these trends, so drop them down below. Check out this practical guide on

停止擔心制定一個完美的學習AI計劃，而是直接開始。我很想聽聽你對這些趨勢的看法，所以在下方留言。接下來看看這個關於

Google Gemini next. See you on the next video. In the meantime, have a great one.

Google Gemini的實用指南。下個視頻見。在此期間，祝你一切順利。