Google 坐擁全球最頂尖的 AI 研發資源與先發優勢，卻因害怕顛覆其搜尋廣告的鉅額現金流，陷入典型的「創新者困境」，導致核心領袖出走、內部產品遭到自我審查，從 AI 領跑者淪為綁手綁腳的追趕者。
- 從霸主到困境：一年半前 Google 挾帶最強 AI 模型之姿領先全球，但如今高層地震、人才流失，外界開始質疑其未來。
- 歷史的諷刺：Google 早在 2017 年就發表了當代 AI 的基石論文「Attention is All You Need」，卻因商業模式考量將其束之高閣。
- 被內部扼殺的 GPT 前輩：前員工證實 Google 在 ChatGPT 問世前一年就擁有類似的對話產品，但因 DeepMind 被禁止推出可能干擾搜尋業務的產品而告終。
- 創新者困境的現形：理性的管理決策（保護高利潤的搜尋廣告）導致了擁抱破壞性技術的失敗，舊商業模式正被 AI 原生問答逐步侵蝕。
- 高層出走與戰略分歧：Demis Hassabis 轉向長遠科學研究，傳奇人物 Jeff Dean 則徹底離職創立 AI 新創，顯示內部創新文化已難以容納頂尖野心。
- 硬體與數據的底牌：雖然軟體層面暫時落後，Google 仍手握全球獨有的搜尋數據與自主研發的 TPU 晶片（Tensor Processing Unit）這兩大優勢。
- 開源策略的賭注：與其在封閉模型上追趕，分析師建議 Google 應全力押注開源模型生態，以此帶動其 TPU 雲端服務的普及，從底層基礎設施獲利。
-
數據壟斷與創新的悖論：Google 表面上的劣勢在於其僵化的營收結構，但真正的諷刺在於，它雖然擁有全球最龐大的使用者行為數據，這些數據卻反而成為了枷鎖。當公司習慣於利用數據優化點擊率與廣告轉化時，任何無法直接嵌入此循環的技術都會被視為高風險的「干擾源」。他們手握變革的火種，卻因為擔心燒掉眼前的糧倉而選擇將其踩熄，這證明了過度的數據依賴在破壞性創新面前，可能從資產轉變為負債。
-
組織慣性與中層管理的角色：許多分析將問題歸咎於高層，但影片點出了更深層的結構性問題：中層管理。在一個利潤極高的成熟業務中，中層管理者的績效考核往往與業務的穩定性掛鉤，而非破壞性成長。當挑戰搜尋的新技術出現時，這些管理者基於個人職位安全與部門利益的考量，會不自覺地形成強大的內部阻力，以「維護公司核心利益」為名，行抑制創新之實，這是大型企業「成功的詛咒」中最難察覺的癌細胞。
-
從「搜尋引擎」到「答案引擎」的轉向陣痛：Google 長久以來的使用者體驗是「整理資訊」並引導至外部連結，而 LLM（大型語言模型）則是直接「生成答案」。這不僅是技術路線之爭，更是存在主義的商業模式危機。當 AI 直接在搜尋頁面頂部解決了使用者問題，右側的廣告欄位與下方的 10 條藍色連結便失去了曝光價值。Google 內部的猶豫不決，反映出這間公司尚未在哲學層面解決「我們究竟是資訊的守門員，還是答案的供應者」這個根本問題。
-
Jeff Dean 出走的象徵意義：作為 Google 的第 30 號員工與核心架構的奠基者，Jeff Dean 的離職絕非單純的人事異動。這就是技術人才的最高級別投票，證明了即便是無上限的預算、最頂尖的計算資源和最寬鬆的管理承諾，都無法彌補大企業內部那種無法逃脫的引力場。當最熟悉體系的人都認為在外部從零開始比在內部改革更快時，意味著 Google 的內部創新機制可能已經發生了系統性的失靈。
-
TPU（Tensor Processing Unit）的真正戰略價值：外界常將 Google 的晶片視為 Nvidia 的替代品，但這低估了 TPU 的戰略深度。擁有自研晶片意味著 Google 是唯一一家可以垂直整合從 AI 模型、軟體框架（如 JAX）、硬體晶片到最終應用的企業。即便它在生產頂尖通用模型上短暫落後，只要開源生態迫使全世界使用其軟體棧，那麼無論模型是誰訓練的，Google 都能透過提供最具成本效益的 AI 基礎設施來收取「算力稅」，這是 Nvidia 式的硬體霸權路徑。
| 概念 |
說明 |
| Transformer 架構 |
2017 年由 Google 團隊在《Attention is All You Need》論文中提出，是現代所有大型語言模型（如 GPT、Claude、Gemini）的技術基石。 |
| 創新者困境 （Innovator's Dilemma） |
由 Clayton Christensen 提出，描述管理良好的企業因過度關注現有客戶與利潤，反而在面對破壞性技術時失去市場領導地位的現象。 |
| LM Chat （內部代號） |
Google 在 ChatGPT 問世前約一年開發的內部對話 AI 雛形，因擔心顛覆搜尋業務而未能公開推出。 |
| 非確定性 （Non-deterministic） |
AI 模型生成結果具備隨機性，無法完全預測或控制輸出內容，這與 Google 過去對搜尋產品精準可控的要求產生衝突。 |
| TPU （Tensor Processing Unit） |
Google 自主研發的 AI 加速晶片，專為機器學習任務設計，目前已發展至第八代，是其擺脫 Nvidia 依賴並建立雲端算力護城河的關鍵。 |
Google 的困境為所有大型科技公司敲響了警鐘：技術領先並不等同於市場勝利，當企業的現金流過於集中在單一模式時，內部最強的防禦機制將自動轉向攻擊自己的未來。這種「自我免疫攻擊」式的管理邏輯證明了，在 AI 時代，速度與開放性比控制權更重要。那些為了保護舊時代利潤而設下的規則與壁壘，最終只會逼走最優秀的建造者，他們會選擇在新創的貧瘠土地上重新種下森林，因為那裡沒有需要保護的老舊建築。
此外，從 Jeff Dean 到 Demis Hassabis 的轉向可以看出，未來最頂尖的 AI 人才不再將「在巨頭任職」視為終點。這群人追求的是擺脫財務季報與中層政治的自由度，他們要解決的是科學本質問題、是 AGI（通用人工智慧）的極限，甚至是疾病治癒。這預示著 AI 產業的權力重心正在從成熟企業的實驗室悄悄流向那些極具使命感的獨立小型敏捷團隊。
最後，Google 的出路暗示了 AI 商業模式的潛在終局：應用層的模型可能迅速商品化，真正的價值將沉澱在底層的基礎設施與獨家數據中。如果 Google 能放下「必須擁有最強通用模型」的執念，轉而利用 TPU 的硬體成本優勢成為「開源模型的超級工廠與託管者」，它有機會在推理與算力需求的爆炸性增長中重新定義自己的護城河。
- "I sometimes think about that Jeff Dean interview where he said they had an internal bot before ChatGPT but didn't think it was better than just googling."
- "I was part of that team basically ChatGPT one year before it came out... Google was too nervous to release it and DeepMind was blocked from shipping products that could disrupt Google."
- "Well-managed companies... can still lose market dominance... the very decisions deemed logical for near-term success are often the reasons leading to failure when confronted with disrupted technological change."
點擊展開完整字幕
0:00 A year and a half ago, Google was on top
0:02 of the world. They released the most
0:04 powerful AI model on the planet, and it
0:06 seemed like everything was going their
0:09 way. But since then, they have fallen
0:12 behind. The stumbles that Google has
0:15 seen over the past year and a half have
0:18 been significant, and we are potentially
0:20 watching the biggest bag fumble of all
0:24 time. Google just had a major shakeup in
0:26 AI leadership. Jeff Dean who is one of
0:29 the most famous computer scientists in
0:32 the world responsible for building out
0:34 much of Google's core architecture and
0:37 infrastructure today their 30th employee
0:40 left to start his own AI company. Then
0:42 we have Deis Hassabis one of the
0:45 greatest AI minds of all time co-founded
0:48 and CEO of Google DeepMind just stepped
0:51 down as CEO. Why are all these things
0:54 happening right now? And how did we get
0:55 from Google being on top of the world to
0:58 many people questioning their future?
1:00 Let me give you a little bit of
1:01 background on Google and artificial
1:03 intelligence because they have been on
1:05 the leading edge for nearly two decades.
1:07 That's what makes this story all the
1:09 more surprising. So Google has had AI
1:12 powering some of their products for a
1:14 very long time. They've been on the
1:15 cutting edge of AI. They even wrote the
1:18 seinal paper attention is all you need
1:20 in 2017. opened it up, opened the
1:23 research and let everybody see it. And
1:25 by the way, if you're not familiar with
1:26 this paper, it is the foundation of all
1:30 modern AI systems. Everything that
1:32 OpenAI built, everything that Anthropic
1:34 has built, all things like Chad GPT are
1:38 based on this paper. That's how
1:40 important this paper is. And they just
1:42 did nothing with it. And at the same
1:44 time, Google DeepMind had developed AI
1:48 that beat the best players in the world
1:50 at Go, beat the best Starcraft players.
1:53 They really had the best AI systems, yet
1:56 they just couldn't productize it. And
1:58 the reason for that, I will tell you in
1:59 a moment. It is something every company
2:02 inevitably faces. And so that brings us
2:05 to this tweet, which really highlights
2:07 the reason Google fumbled so hard. I
2:10 sometimes think about that Jeff Dean
2:11 interview where he said they had an
2:13 internal bot before Chad GPT but didn't
2:16 think it was better than just googling.
2:18 That tells a lot about what Google was
2:21 thinking in that moment. But that's
2:22 actually not the worst part. A reply by
2:25 Tibo and Tibo is one of the leads on the
2:28 codeex team at OpenAI. a very
2:31 influential person in artificial
2:33 intelligence who also just happened to
2:35 be on Jeff Dean's team at Google
2:38 DeepMind. I was part of that team
2:40 basically chat GPT one year before it
2:43 came out called LM Chat and then another
2:45 code name. He's saying they had a Chat
2:46 GPT product one year before Chat GPT
2:50 came out. And if you don't remember what
2:51 it was like when Chad GPT came out, it
2:54 almost instantly changed the world. It
2:56 was so big and so sudden. So they had
3:00 that. Don't forget Google had it. And so
3:02 he goes on, Google was too nervous to
3:05 release it and DeepMind was blocked from
3:09 shipping products that could disrupt
3:11 Google. I think about this a lot. Okay,
3:14 so much to unpack here. Google was too
3:17 nervous to release it. So they were
3:19 nervous probably for a few reasons. He
3:21 details one of them. DeepMind was
3:23 blocked from shipping products that
3:24 could disrupt Google. Google is probably
3:27 the best business ever created. You go,
3:30 you search, it serves you ads and they
3:33 generate so much revenue from that and
3:36 the profit margins are so massive. It is
3:39 this enormous cash cow that they have
3:41 been milking for almost two decades now.
3:44 And so anything that could threaten that
3:47 [clears throat] cash cow is a threat.
3:50 And even if they are developing it
3:53 themselves internally, it's a threat and
3:56 it got shut down. That is what Tibo's
3:58 saying. And so I think another reason
4:00 why they were nervous is simply because
4:02 AI by nature is non-deterministic,
4:05 meaning they can't fully control it.
4:08 They can't prevent it from saying things
4:10 that would reflect badly on Google. And
4:13 they were very nervous about that. But
4:15 the main thing to keep in mind here is
4:17 all of this is called the innovators
4:20 dilemma. And if you want to stay
4:22 uptodate on the latest in AI, check out
4:24 our newsletter forwardfuture.com
4:27 link down below. So let me read kind of
4:29 what this innovator's dilemma is. So
4:32 well-managed companies like Google,
4:35 those that listen astutely to customers,
4:37 invest aggressively in new technologies,
4:39 and allocate resources systematically
4:42 can still lose market dominance. The
4:45 paradox or innovator's dilemma is that
4:48 the very decisions deemed logical and
4:51 competent for near-term success are
4:53 often the reasons leading to failure
4:55 when confronted with disrupted
4:57 technological change. So basically to
5:00 kind of put it in the context of Google,
5:02 they have this cash cow. It's going
5:04 really really well and then they have
5:06 this new technology large language
5:08 models that is coming and it just seems
5:10 like an alternative but it doesn't seem
5:12 maybe as good and why would they disrupt
5:16 their cash cow? Their customers want
5:19 search results. Great. So give them more
5:21 search results. But it turns out that
5:23 very logical decision was actually going
5:26 to hurt them in the long run. So we have
5:29 Google's business here and then all of a
5:31 sudden we have large language models and
5:34 the large language models start to grow.
5:37 Google fails to identify and adopt AI
5:41 effectively and that's when we can start
5:45 to see the downfall the collapse of the
5:48 old business model which Google is
5:50 definitely going through right now.
5:52 their search results. Although yes, it's
5:55 still strong, everybody I know is going
5:58 directly to either Chai GPT or Claude
6:01 and yes, even Gemini, but it's all AI.
6:05 They're traditional 10 blue links and
6:08 advertising around that is being
6:10 disrupted right now. And so now let's
6:13 bring it back to Demis and Jeff Dean.
6:16 Let's talk about why these different
6:19 factors may have caused them to want to
6:21 either leave or step down and focus on
6:24 other things. And if we read Deis'
6:26 tweet, we can kind of get a glimpse of
6:28 it. Remember, the innovator's dilemma is
6:31 because of short-term thinking, not
6:34 because of bad decision-making, but just
6:36 because they're thinking about the wrong
6:37 thing. Listen to this. I'm stepping into
6:39 a new role as chair of Google DeepMind
6:42 and chief scientist of Alphabet. Now,
6:44 here's the important part. This will
6:46 allow me to focus on long-term strategy,
6:49 accelerating scientific breakthroughs,
6:52 including leaning into my own work at
6:54 Isomorphic to help cure disease. He
6:57 wants to think long term. He doesn't
6:59 want to build the Google DeepMind
7:03 company anymore because you are
7:05 basically beholdened to quarterly
7:07 earnings and shareholders. He wants to
7:10 be able to think long term because again
7:13 he identified the innovator's dilemma.
7:15 He identified that ultimately Google's
7:18 cash cow is under siege right now and he
7:21 doesn't want to try to manage that
7:23 transition. He wants to go to the
7:25 frontier. He wants to build the future.
7:29 And so that's I think a big part of why
7:31 he decided to step down. Then that
7:33 brings us to Jeff Dean. Jeff Dean didn't
7:35 just step down. He's actually leaving
7:37 Google, which is a big big deal. And I
7:42 kind of think how this played out was
7:44 Jeff was just seeing other companies
7:47 building the future, Anthropic, Open AI,
7:51 and he just wasn't feeling like he could
7:53 accomplish that within the walls of
7:55 Google, which, you know, if you think
7:58 about how important Jeff Dean is to
8:01 Google, they probably gave him anything
8:03 he wanted. They probably said, "You can
8:06 have unlimited budget. You can have
8:07 unlimited compute. You can hire or take
8:10 anybody you want to go build this. We
8:12 won't even tell you what to do. There
8:14 will be no management." And even with
8:17 all of that, he did not think it
8:20 possible to build his vision to get
8:23 excited and build the thing he wanted to
8:24 build within the walls of Google. And
8:26 that is extremely telling to the current
8:29 culture at Google. And so he went off
8:31 and now he's building his new company,
8:33 Discovery Loop. Automating discovery to
8:36 accelerate science and engineering for
8:38 the world. And by the way, if I didn't
8:41 make it clear how important Jeff Dean is
8:43 and three other people from Google who
8:46 went with him, look at the list of
8:48 incredible accomplishments that they
8:50 achieved while at Google. They built
8:52 Google search, that little business,
8:55 Google ads, content ads, Gmail, news,
8:59 translate, Gemini, cloud TPUs, a piece
9:02 of infrastructure called map reduce,
9:04 which is one of the fundamental pieces
9:07 of architecture inside Google that
9:10 allows it to do distributed systems,
9:12 meaning it can serve billions of people
9:14 throughout the world and serve all of
9:17 them in milliseconds. That's how it
9:19 works is this map produced system. and
9:21 everything else you're seeing here,
9:22 including Alpha Star, Alpha Fold, which
9:25 was the protein folding AI that they
9:27 built. So, just an incredible team,
9:29 incredible accomplishments, and they
9:31 needed to leave Google to actually do
9:33 it. But I actually thought it was pretty
9:35 funny that this is from their pitch
9:37 deck. And to think that Jeff Dean and
9:40 team actually needed a pitch deck, which
9:42 they definitely didn't, but they went
9:44 through the exercise of creating it
9:45 anyways, I thought it was hilarious. And
9:48 uh it kind of just shows the humble
9:50 nature of Jeff Dean. Now I think a lot
9:52 of the turmoil within Google and the
9:54 trouble that they're having now started
9:56 from a few different things. One again
9:58 it is very much the innovator's dilemma.
10:00 They have this massive business that
10:02 they don't want disrupted. Tibo said it
10:05 like the Google DeepMind team was
10:06 blocked from releasing their language
10:09 model product simply because they
10:11 thought it might disrupt Google search.
10:13 I mean that is the literal definition of
10:16 the innovator's dilemma. And also, I
10:18 think a lot of it is middle management.
10:20 A lot of people in there who just don't
10:22 want their jobs threatened, don't want
10:25 change. They like the way things are,
10:27 and you know, they're they're not
10:29 necessarily on the frontier of
10:31 scientific research, and that's what's
10:33 most excited to them. What's most
10:35 exciting to them is bringing home the
10:37 paycheck. And so, I think a lot of these
10:38 middle managers may have caused some of
10:40 these cultural issues at Google as well.
10:42 Now, this is all speculation by me, just
10:44 to be clear. But here's the thing. I'm
10:46 actually still quite optimistic about
10:49 Google. They lost a few of their best,
10:51 but what has happened historically when
10:53 people have left companies that I've
10:55 either worked for or that I led myself
10:57 is the team gets stronger. They get
11:00 closer. You can always hire more great
11:02 people. And obviously, Google has so
11:05 many incredibly talented people,
11:07 including friend of the show, Logan
11:09 Kilpatrick. So, uh, I'm not really
11:11 worried about Google. And there's a
11:13 bunch of other reasons. I still think
11:14 they're well positioned. So number one,
11:17 they have all the data in the world. All
11:19 of that training data will be used to
11:21 train insanely good models and much of
11:24 it is proprietary training data. No
11:27 other company in the world has it. So
11:29 they have it. So that's number one. They
11:31 also have their own chips in the TPU,
11:33 the tensor processing unit, and they're
11:35 on like the eighth generation of it. So
11:39 they're well experienced with building
11:42 their own silicon. They're not
11:43 necessarily battling every other company
11:46 to get the latest Nvidia GPU, although
11:48 they do serve with Nvidia, but they also
11:51 have their own TPUs, and that is a
11:53 strong competitive advantage. They also
11:55 have incredible hardware with Android.
11:59 They power much of the mobile phones in
12:01 the world, and they also have a cash cow
12:04 business. They have a ton of free cash,
12:06 and all of that, they can just continue
12:08 to make massive bets. So, I'm really not
12:12 all that worried. Now, if I were Google,
12:14 I would not continue pretending to be
12:17 competitive at the absolute frontier
12:19 with Fable and Soul. You just have to
12:22 admit it, okay? They're not there, and
12:24 that's okay. And if I were them, I would
12:27 actually go so hard on open source. I
12:31 would try to build the absolute best
12:33 open-source model on the planet. Get
12:36 people building on top of your
12:37 architecture. get people building on top
12:39 of your AI and then iterate from there.
12:42 And I think this is a really strong
12:43 strategy. We're already seeing it with
12:46 AI companies out of China. And there are
12:48 really only two US companies that can
12:52 afford to have a very strong open-source
12:54 strategy. Nvidia, which already has it.
12:57 They're investing over $20 billion into
12:59 open source and their Neatron family of
13:01 models. And Google, they will benefit
13:04 when open source proliferates. and they
13:07 already have the Gemma family of models,
13:09 although those are more optimized to be
13:11 small models for your mobile phone. But
13:13 if they went allin on open source and
13:17 everybody started serving open source
13:18 models and then they continue to sell
13:21 their TPUs, then they win no matter
13:24 what. Everybody will want to power the
13:26 open source models with the hardware
13:27 that powers them the best, the most
13:30 efficiently, and they already have a
13:32 strong competitive advantage with the
13:34 TPU. So, if they just released all of
13:36 these incredible models for free and
13:38 then served them and sold the TPUs as
13:41 well, I think that would be an
13:42 incredibly strong strategy. So, I still
13:45 am very hopeful for Google. I am
13:46 certainly far from counting them out and
13:48 I don't think you should either. Open
13:50 source absolutely can be a winning
13:52 strategy for Google. I actually detailed
13:55 how important open source is in this
13:57 video right
