來源：YouTube | 00:37:02 | 2026-08-11
祖克柏主張超級智慧應廣泛分配以實現安全與繁榮，但其論點忽略運算資源注定由資本壟斷，終將形成永久性不平等，矛盾地摧毀了他所倡導的賦權願景。
- 祖克柏發表長文《未來屬於每個人》，抨擊 Anthropic 等封閉實驗室的集中化安全觀，主張開源與個人賦權，認為發明而非自動化才是超智慧的核心貢獻。
- 影片作者認同大部分觀點，但指出核心矛盾：即使模型層級開源，計算與能源的稀缺性將使擁有更多資本者獲得不成比例的智慧優勢。
- 以法律訴訟、商業競爭與網路安全等思想實驗為例，說明「人人有超智慧」不等於公平，實際產出取決於能投入多少計算量。
- 討論 AI 對就業的影響，祖克柏預測公司規模將縮小但數量激增、新工種不斷湧現；作者同意此樂觀前景，但仍憂心算力不平等的階級固化。
- 祖克柏回應數據中心爭議，舉例教師獎金與水資源正效益承諾，並呼籲加速能源建設以與中國競爭；影片對封閉式水冷系統持審慎信任。
- 提出蒸餾應合法化、維持矽晶片出口管制，以及多實驗室同時達成遞迴自我改善以互相制衡的構想，作者批評該制衡機制在邏輯上不成立。
- 總結祖克柏的開放願景雖然動人，但整個論證在「算力決定一切」的經濟現實面前自我瓦解，最終仍是資本強者全拿。
-
開源神話的破口：算力決定一切
祖克柏將「超級智慧分配」簡化為模型層級的開放，但現實是智慧品質取決於推論時投入的計算量。即使開源模型人人可用，能支付更多 GPU 時數的個人或企業仍能執行更深層的推理、處理更龐大的脈絡，形成實質的「智慧不平等」。這並非技術烏托邦的意外副作用，而是資本主義在算力市場的必然延伸——運算即權力，而權力從未公平分配。祖克柏自己也承認「計算總量有限」，卻迴避了誰來決定分配這一根本政治問題。 -
「永久底層階級」的邏輯必然性
影片數次點出的「permanent underclass」並非聳動之詞。當 AGI 時代到來，個人的社經地位將被鎖定，因為所有競爭都歸結為算力競賽。祖克柏稱「人人都有超級律師就能實現司法公正」，卻忽略了雙方投入的計算資源不可能對等，正義仍將由資本定義。這與網路早期去中心化理想破滅的歷史如出一轍：開放協定最終被少數平台壟斷，而這次被壟斷的將是智慧本身——誰掌握算力，誰就掌握定義現實的權力。 -
Anthropic 的「彌賽亞情結」與 Zuckerberg 的「人人持槍」安全觀
祖克柏猛烈批判 Anthropic 那種「由我們來決定如何安全」的家長式作風，極具感染力。但深入檢視，他的替代方案——讓人人擁有超智慧以相互制衡——類似於「人人持槍，社會更安全」的邏輯。在算力不對等的前提下，這只是將權力鬥爭從少數菁英擴大到全民軍備競賽，可能加劇惡意使用與社會混亂。集中化的危險顯而易見，但極端分散化在缺乏公平資源基底的狀態下，亦難以導向真正的共同安全。 -
「發明而非自動化」的論述陷阱
祖克柏主張超智慧的主要貢獻在發明新事物，並以農業人口轉移佐證。此類比看似合理，卻忽略一個關鍵差異：AI 不僅提升特定領域槓桿，而是對認知勞動本身進行全面替代，歷史上從未有技術革命同時威脅幾乎所有腦力工作。新工作創造的速度能否跟上？即使出現新工種，它們是否也需要大量算力才能勝任？這暗示一個階級分化：能指揮 AI 的「發明階級」與被 AI 取代的「無用階級」並存，而前者正是算力的擁有者。 -
蒸餾與出口管制：國家競爭中的雙重標準
祖克柏呼籲美國國內蒸餾合法化以促進開源生態，同時支持對中國的晶片出口管制。這暴露了「開放」的雙標：對內要求知識自由流動以削弱封閉巨頭，對外則仰賴硬體封鎖維持國家優勢。他批評黃仁勳反對管制的觀點，但若中國最終建立自主晶片能力，出口管制反可能催生完全脫鉤的敵對生態系，長遠來看對美國開源主導權未必有利。這種內外有別的策略，恰恰說明了連最激進的開放派都無法逃脫地緣政治的算力民族主義。
| 概念 | 說明 |
|---|---|
| 超級智慧 (Super Intelligence) | 遠超人類認知能力的 AI；祖克柏視為應普及的工具，作者視為算力競賽的獎品。 |
| 運算 (Compute) / 能源 | 決定 AI 效能的關鍵稀缺資源；影片核心論點：這才是真正的權力槓桿，而非模型本身。 |
| 永久底層階級 (Permanent Underclass) | AGI 出現後，個人的社經地位因資本與算力差異而永久固化，相對地位無法翻轉。 |
| 遞迴自我改善 (Recursive Self-Improvement, RSI) | AI 可自主改進自身，導致智慧爆炸；祖克柏認為多實驗室同時達成可制衡，影片反駁先發者優勢將無限擴大。 |
| 蒸餾 (Distillation) | 從大模型萃取知識訓練小模型；祖克柏力主合法以利開源，但將徹底破壞封閉實驗室的商業模式。 |
祖克柏的文章與影片分析揭示了 AI 產業最深層的分裂：並非開源 vs. 閉源的技術路線之爭，而是對「權力應如何分配」的根本對立。對企業領袖而言，若接受「算力即護城河」的邏輯，那麼投資重心將從模型研發轉向能源與基礎設施的垂直整合。我們可能見證科技巨頭轉型為「智慧公用事業」，按量販售認知能力，而其他多數公司淪為算力租用者，喪失核心競爭力——這正是作者所指「矛盾」的終點。
創業者則需警覺，一人獨角獸的願景雖誘人，但在算力軍備競賽下，任何高 ROI 的 AI 應用一旦被巨頭察覺，其算力邊際成本幾乎為零，可瞬間碾壓新創。真正的防禦或許不在模型或演算法，而在於獨佔性資料、法規壁壘或緊密的使用者社群，這些是純算力難以複製的資產。同時，國家層級的能源與管線政策，將直接定義各國在智慧生產時代的階級位置。
- "I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future."
「我不懂，為何那些相信 AI 會消滅多數工作與人類存在意義的人，會急著去建設那樣的未來。」 - "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic."
「認為 AI 如此危險，以至於唯一安全之道就是極端集中權力，這個想法本身就有根本問題。」 - "Knowledge, truth, it generally finds a way out. But controlling physical components ... is much easier."
「知識與真相幾乎總能找到出路。但控制物理元件……容易得多。」
點擊展開完整字幕
0:00 Mark Zuckerberg just published an essay 0:02 about the future of AI. Abundance, 0:05 bioweapons, super intelligence for 0:08 everyone. It sounds awesome, but as I 0:10 read it, I saw one critical flaw in his 0:13 argument and I couldn't get it out of my 0:16 mind. So, I sat down to make this video 0:18 to get all of my thoughts out there. So, 0:20 here's the essay. It is called the 0:22 future is for everyone. Who will have 0:25 access to super intelligence and what 0:26 will we direct it towards? Will it be 0:28 centralized and restricted to a few 0:30 institutions or will it be a tool that 0:32 empowers everyone? So you can think of 0:35 on one side you have the anthropics of 0:37 the world and I hate to just bash them 0:39 immediately in this video but they truly 0:41 believe the way to make artificial 0:42 intelligence safe is by them deciding 0:46 how it works. And I could not disagree 0:49 more wholeheartedly. Then somewhere in 0:51 the middle you have open AI and then on 0:53 the very other side you have Meta. and 0:56 Meta is going hard on open source and 0:58 you're going to see that throughout this 1:00 entire essay and it's a long essay. We 1:02 propose a philosophy based on individual 1:05 empowerment as the source of prosperity, 1:08 invention as the primary purpose of 1:11 super intelligence and balance of power 1:14 as the foundation of safety. I could not 1:17 agree more nor have said it better 1:18 myself. Individual empowerment. 1:21 Everybody should be able to have AI 1:24 working for them. Invention as the 1:26 primary purpose of super intelligence. 1:29 That means AI to invent new things. But 1:33 what they are specifically going to 1:35 point out throughout this essay is that 1:37 it is not job automation. And then last 1:40 balance of power as the foundation of 1:43 safety. Safety in artificial 1:45 intelligence comes from everybody having 1:47 it. And that's where the one flaw that I 1:49 mentioned comes in, but I'm going to get 1:51 to that in a moment. Now, Mark 1:53 immediately comes from the top ropes 1:55 here. I do not understand why anyone who 1:59 believes that AI will eliminate most 2:01 jobs and much of humanity's relevance 2:03 would rush to build that future. I don't 2:06 understand it either. We have Frontier 2:09 Labs, again, anthropic, who Daario is 2:12 quoted saying, "We're going to have a 2:13 white collar blood bath." And then he 2:15 just continues to plow forward building 2:17 the thing that would have caused it. And 2:20 so yeah, to me it also makes no sense. 2:22 Why are you so pessimistic about the 2:25 future and yet you are building the 2:26 future? And that's like the messianic 2:29 kind of vibe coming out of anthropic. 2:32 Like, hey, we know it's going to destroy 2:34 the world, but we'd rather be the ones 2:37 destroying the world and trying to save 2:39 us. It's I don't know. It's so 2:40 backwards. The notion that AI is so 2:42 dangerous that the only safe path is an 2:45 extreme concentration of power seems 2:47 inherently problematic. Again, could not 2:50 agree more. Historically, hoping that an 2:52 absolute power will benevolently provide 2:55 for humanity if sufficiently enlightened 2:58 has not led to safe or positive 3:01 outcomes. And there's two quotes that 3:03 come to mind. Power corrupts and 3:05 absolute power corrupts absolutely. This 3:09 is a saying for a reason. And then one 3:12 of my favorites. I can't even say it 3:14 without a with a straight face. With 3:16 great power comes great responsibility. 3:19 Spider-Man. No, it was his Uncle Ben, 3:22 right? Uncle Ben said it. And if we look 3:23 at technology historically, every single 3:26 time there has been this huge 3:28 technological shift, we have figured out 3:30 what to do. There's been more 3:31 prosperity. There has been more 3:33 abundance. And a rising tide lifts all 3:36 ships. So each time there is fear that 3:38 people will be left behind. But each 3:40 time humanity has come out with more 3:42 people sharing greater prosperity, 3:44 health and freedom. And so that is why I 3:47 believe the future is bright. It 3:49 continues and I love this invention not 3:53 automation will be the greatest 3:55 contribution of super intelligence. And 3:57 that's really a key and I'm already 4:00 seeing it inside my company. We've 4:02 automated a lot of our work, but we are 4:05 busier than ever. We are doing new 4:08 things that we wouldn't have otherwise 4:10 without AI. Without being able to 4:12 automate the parts that we've been doing 4:13 manually for 2 years, we simply would 4:16 have just continued doing those things. 4:18 And now that we can automate that, we 4:20 can now invent new things to do. We can 4:23 create more better, higher quality 4:26 content for you and so much more. Now 4:29 again, Mark keeps taking shots at other 4:32 AI labs, and we know who he's talking 4:35 about here. Some argue that super 4:38 intelligence itself or a small set of 4:41 experts who control it should decide 4:44 what is best for humanity. We disagree 4:47 and I agree with that. But the argument 4:51 does break down and I'm going to get to 4:53 that in one moment. rather than 4:54 centralizing super intelligence, we 4:56 should distribute it widely and give 4:58 every person the ability to direct it. 5:00 So that I agree with that makes a lot of 5:02 sense to me. Everybody should have the 5:04 ability to have super intelligence on 5:08 tap and as much as they want, but it 5:11 does not work that way. And if you're 5:13 into entrepreneurship, if you're into 5:15 building things, I'm excited to tell you 5:18 about the sponsor of today's video, 5:19 Zapier. If you want to give your agent 5:22 all of the tools it needs to get 5:24 realworld work done, you need to connect 5:27 it to Zapier's MCP server. I've 5:29 literally been using Zapier for over 10 5:32 years. I could not recommend them more 5:34 highly, and it is now dead simple to 5:37 give your agents thousands of tools 5:40 instantly. I'm talking about Gmail, 5:43 Google Calendar, Asauna, Notion, Slack. 5:46 Any tool you use probably is available 5:49 in Zapier. Zapier has been doing 5:51 automation for a very long time. So all 5:54 the brittleleness you've experienced 5:56 elsewhere, you will not have using 5:58 Zapier. You just go in, configure the 6:01 tools that you want to use, they give 6:03 you a URL, and you simply give it to 6:05 your agent. That's it. And now your 6:07 agent has access to any tool that you 6:10 need. And it works with all of your 6:12 agents, whether it's Claude Code or 6:14 Codeex, OpenClaw, Hermes, it doesn't 6:18 matter. It will just work. And you can 6:21 set it all up without writing a single 6:22 line of code. They have a free plan to 6:24 get started and you just scale up as you 6:26 need. So check out Zapier. They've been 6:28 such a great partner. I'll drop links 6:30 down below. Thanks again to Zapier. Now 6:32 back to the video. So, here's a few 6:33 concrete examples of how Mark Zuckerberg 6:36 and Meta see super intelligence 6:38 benefiting all of society in the future. 6:41 Your agent will work 24/7 on your behalf 6:43 to improve your relationships, health, 6:45 career, finances, home management, 6:47 hobbies, and more. It will free up time 6:50 for the things you enjoy and help you 6:51 accomplish more than you could 6:53 otherwise. And I think this is actually 6:54 a really good line right here. It will 6:56 free up time. And what you decide to do 6:59 with that time is going to be up to you. 7:02 You can work less certainly, but there 7:04 definitely will be a subset of people 7:06 who just think, "Okay, I have a bunch of 7:09 free time. I'm just going to work even 7:10 harder now. I'm going to get more done." 7:12 And you can kind of see the conflict 7:14 between people who want to take more 7:16 time to themselves and people who are 7:18 overachievers and want to work even 7:20 harder. So, this is from Inc. magazine. 7:22 Meta chief technology officer Andrew 7:24 Bosworth all but rolled his eyes when an 7:27 employee asked if AI productivity gains 7:29 should be parlayed into a revival of 7:31 meta days which are like extra days off 7:35 that you earn at Meta and he said stop 7:38 asking about meta days it's very dumb 7:41 and I hope that what we do with our 7:44 extra time is do even more and cooler 7:47 stuff for the users who use our products 7:49 every day we got billions of people 7:50 using our products every day I get an 7:52 extra hour, you know what I'm going to 7:54 do with it, I put it into that. And so, 7:56 there's definitely going to be this 7:58 divide. And totally fine. You know, if 8:01 you want to go lay on a beach because 8:03 you've got all this stuff done and you 8:04 have a bunch of extra time, more power 8:06 to you. And if you want to go work even 8:08 harder and get more done, totally fine 8:10 with that as well. Now, next, Mark 8:11 Zuckerberg said, "It will have strong 8:13 privacy and security options, so you can 8:15 trust it to handle all of your personal 8:16 content knowing that no one else can 8:18 access your information." I have to 8:20 admit there's definitely some irony with 8:23 Meta saying that Meta does not have a 8:27 strong track record of privacy and then 8:30 also he mentions your glasses and I know 8:34 a lot of people like the glasses form 8:35 factor but I am not one of them. Now 8:37 this next part is extremely important. I 8:39 actually think the entire economic 8:41 landscape is going to be changing. 8:43 Entrepreneurship is going to be 8:44 changing. So, everyone will have 8:47 powerful tools to create new businesses 8:48 and the economy will become more 8:51 entrepreneurial. As an entrepreneur 8:53 myself, I love this. People are starting 8:57 to be able to manifest ideas themselves 8:59 without having to raise money or build 9:00 large teams. There's been this 9:03 prediction in Silicon Valley for 9:05 probably over a year that there's going 9:08 to be a oneperson billion dollar company 9:10 soon. Now, we haven't had it yet, but 9:13 we're really close. A company with under 9:15 10 people can reach a billion dollar 9:17 valuation today. Now, here's the key. 9:20 Many ideas that would have been too hard 9:22 or expensive to try before will now be 9:25 possible. I made a video about this 9:27 something like a year ago where I I 9:29 argued that the total universe of 9:32 problems to solve is effectively 9:33 infinite. And with artificial 9:35 intelligence, the long tale of problems 9:37 in which the ROI to solve those problems 9:40 was not there all of a sudden becomes 9:44 very achievable. It becomes very real. 9:47 And now you can use AI to solve those 9:49 longtail of problems. And all of a 9:51 sudden there's so many more problems 9:52 that are solvable. And that also leads 9:55 to many more entrepreneurs which I see 9:57 as a very good thing. And so he 9:59 concludes with more employment over time 10:02 rather than less. Next he talks about 10:05 having a personalized tutor and coach 10:06 with a PhD in every subject. Yeah, 10:08 that's awesome. I don't think anybody's 10:10 arguing against that or that that will 10:12 happen or kind of already has happened. 10:14 Everyone will benefit from scientific 10:16 advances and be able to contribute to 10:18 scientific progress. Again, it's that 10:20 long tail problem where, you know, if 10:23 you have a hundred people with some rare 10:25 disease, the pharmaceutical companies, 10:27 the bioreearch companies, it just did 10:30 not make sense to invest and find cures 10:33 for a disease that only affected 100 10:35 people because the economics were not 10:37 there. But with AI, it very much is. 10:40 Now, we're going to get to the important 10:42 part where I actually think the argument 10:43 breaks down completely. And I don't even 10:45 know if I have an answer to this yet. 10:47 So, let's see. Everyone will have free 10:50 or affordable access to these tools. 10:53 Now, that will definitely happen. We 10:55 will offer free versions that will be 10:57 accessible to billions of people. For 10:59 those who want to pay to use more 11:01 compute, there will be a dynamic auction 11:03 mechanism that will guarantee that 11:05 everyone gets the lowest price possible 11:08 for the intelligence and compute they're 11:10 using while also ensuring the capacity 11:13 is used for whatever people collectively 11:15 find most valuable. This might be the 11:19 most important sentence in this entire 11:22 essay. Now, basically what he's saying 11:23 is that there's going to be some kind of 11:25 baseline of compute allocated to 11:28 everybody on the free version to do with 11:31 what they want with their super 11:32 intelligence. But on the other hand, if 11:35 you want to pay more, you can. You can 11:37 pay more to buy more compute and you 11:40 will get more intelligence. But those 11:42 two things contradict each other because 11:45 ultimately whoever can afford to buy 11:48 more compute will have more 11:50 intelligence, will be able to accomplish 11:51 more, will be able to generate more 11:54 revenue for themselves. Ultimately, 11:56 compute is going to go towards whatever 11:58 the most valuable use case is. Whatever 12:00 the highest ROI of that compute 12:03 investment is, that's where compute's 12:05 going to go. And then also this is kind 12:08 of the entire concept of the permanent 12:10 underclass. Once AGI is here or once 12:14 super intelligence is here, whatever 12:16 capital you have in that moment, 12:17 wherever you are from a socioeconomic 12:19 perspective, that is where you will be 12:21 forever because everything is relative. 12:24 Everybody is competing for the finite 12:26 set of compute resources. So if you have 12:29 more capital and you can buy more 12:30 compute, you will be able to out compete 12:33 everybody else. All right, so let's keep 12:34 going because this essay contradicts 12:36 itself in multiple parts. Next, he 12:39 starts talking about the risks of AI and 12:42 he talks about job displacement and 12:45 ensuring local communities benefit from 12:47 data center builds, which is obviously a 12:49 huge concern for a lot of people and the 12:51 general sentiment around data centers 12:54 right now in the United States is so 12:56 negative. AI misuse related to cyber 12:58 security and biorisk to avoiding 13:00 government tyranny and surveillance 13:02 ensuring the US and democratic countries 13:05 lead and ultimately making sure humanity 13:07 maintains control over super 13:09 intelligence so it serves rather than 13:11 endangers us. So those are all of the 13:13 major risks that Mark Zuckerberg points 13:15 out in this article and he goes over 13:17 every single one and Meta's vision for 13:20 how to solve each of them. So the 13:22 conventional view is that these concerns 13:24 are about technology and that if we take 13:26 enough time then we can perfect or align 13:29 the technology to produce a single 13:31 benevolent super intelligence. I think 13:33 this view of alignment is fundamentally 13:35 flawed. And I agree this is very much 13:39 the anthropic way of thinking. We are 13:41 going to produce super intelligence. 13:43 We're going to try our darnest to align 13:45 it but we are the best and most capable 13:49 to do so. So, we're going to be the ones 13:52 to do it. Nobody else should be able to 13:53 do it. But Mark points out correctly 13:55 that there is no technological solution 13:58 that can align with everyone's opposing 14:00 interests and values at once. Any 14:03 singular super intelligence would have 14:05 to prioritize some values over others 14:08 and in the process would be incapable of 14:10 being benevolent to everyone. And again, 14:12 this is the flaw in anthropics argument. 14:17 This is the flaw in which they're saying 14:19 we need to keep AI closed and only in 14:23 the hands of a few to decide what 14:25 happens with it. This is just such a bad 14:27 argument and so obviously dangerous. The 14:30 concentration of power with AI is my 14:33 biggest fear. The counter is the best 14:35 and most realistic path to building a 14:37 positive AI future is by delivering 14:39 super intelligence to everybody. I 14:41 agree. But again, how do you do that if 14:44 the model layer, if the actual 14:46 intelligence is commoditized and then 14:49 it's just a function of how much compute 14:52 you can throw at it, how much 14:54 intelligence you can actually buy for 14:56 your dollar, who has the most dollars. 14:59 That's really what it's going to come 15:00 down to. And here they give a very 15:03 concrete thought experiment. And I 15:06 really think the argument breaks down 15:08 completely at this point. As a thought 15:10 experiment, imagine only one person had 15:12 a super intelligent lawyer. They would 15:14 have an unfair advantage in court. Yes, 15:17 true. If you had a super intelligent 15:19 lawyer who could do research and 15:21 understood every single line of every 15:24 single case, all of the case law out 15:27 there, then yes, you will have a 15:29 distinct and unfair advantage in court. 15:32 And that would lead to a worse society. 15:34 But now imagine everyone has a super 15:37 intelligent lawyer. In this case, 15:39 justice would be carried out much more 15:41 fairly and efficiently than it is today 15:44 when there is often an imbalance in 15:46 skills and resources in litigation. And 15:49 so, let's use this example. Let's say 15:51 there are two opposing sides in a court 15:53 case. Both of them have super 15:56 intelligence. Both of them have, let's 15:58 say, the exact same quality of model. 16:01 Whichever team in that court case has 16:02 more compute to throw at the model to 16:04 allow the model to think more deeply to 16:07 do more research to parallelize and get 16:10 more done in the same amount of time 16:13 will have a distinct advantage. So just 16:16 because both sides have super 16:19 intelligence does not mean that super 16:21 intelligence is equal. And that's really 16:24 important because who's going to have 16:26 the most compute? Whoever has the most 16:28 capital. And that again is the entire 16:31 premise of the permanent underclass and 16:34 it applies so well in the legal system 16:37 example. But again this argument that he 16:39 makes that once everybody has super 16:42 intelligence the justice system will be 16:44 fair because of that everybody was going 16:46 to have equal resources and equal 16:48 capabilities. It just doesn't make any 16:50 sense to me. He goes on to talk about 16:52 cyber security. If everyone has access 16:54 to cyber security super intelligence, 16:56 then all of our technical systems would 16:58 become more secure than today since the 17:00 widely deployed super intelligence would 17:02 help harden and update every system. 17:04 Once again, it's all about how much 17:06 compute you have. Now, in this example, 17:09 I actually think this is a good thing 17:12 because I suspect the good guys, so the 17:16 companies that are trying to harden our 17:18 systems that are the defensive actors, 17:21 the major companies all across the world 17:24 are going to have much more capital to 17:26 invest into infrastructure into compute 17:28 to power their super intelligence to 17:31 harden their systems than the bad actors 17:33 will. So in this case, the difference 17:37 between compute, the difference between 17:39 resources with the good guys and the bad 17:42 guys actually helps and makes our 17:45 systems hardened. But in a court case, 17:48 again, both sides hopefully would just 17:51 have equal resources and win on the 17:53 merits of their argument and the facts 17:56 rather than whoever can argue the best 17:59 or do the most research. Then we go back 18:02 to another bad example where the 18:05 argument breaks down. If only one 18:07 business had super intelligence, that 18:08 business would outco compete all others 18:10 and lead to a less dynamic and broadly 18:12 prosperous market than we have today. 18:14 But if everyone has access to super 18:16 intelligence, then everyone will have 18:18 the tools to create new things beyond 18:20 what is possible today. So I agree in 18:22 theory that makes sense. But once again, 18:25 it's really going to be about who has 18:27 the most compute. If everybody has 18:30 access to the same tap of intelligence, 18:34 whoever gets the most intelligence will 18:36 be able to outco compete everybody else. 18:38 And so in this case, once again, can you 18:41 imagine Meta has massive data centers 18:45 and they all of a sudden have 18:47 competition from a startup. They're 18:49 going to be able to buy more compute. 18:51 They own the compute and they're going 18:53 to be able to outco compete that 18:54 startup. So he keeps stating the 18:56 solution is to ensure that super 18:59 intelligence is broadly distributed to 19:00 empower people. But he keeps referring 19:03 to super intelligence in the form of the 19:06 model layer, not the compute as much 19:09 because the compute is really the 19:11 important part and honestly energy is 19:13 just as important. Compute and energy. 19:16 Once open-source proliferates the 19:18 intelligence layer is likely or at least 19:21 a lot of it is going to become 19:23 commoditized. And so it's whoever can 19:25 spend more on compute at that moment 19:28 will be able to out compete in any realm 19:30 that they're competing in. Another dig 19:32 at the competing labs. Meta is the 19:34 company primarily focused on building 19:36 personal super intelligence for everyone 19:38 which I appreciate that most of their 19:40 labs are focused on building AI for 19:42 companies, governments and other 19:44 institutions. So if those labs lead, 19:46 then the balance of power will favor 19:48 large institutions over individuals. But 19:50 again, large institutions are going to 19:53 be able to outspend individuals on 19:56 compute. Now, he goes on to talk about 19:58 job growth in the economy. This is where 20:01 I come back to agreeing with him again. 20:03 So, people have infinite demand for new 20:06 experiences and have always found new 20:09 problems to tackle. I made this argument 20:12 in a video I made about a year ago. 20:13 There are infinite problems to solve. 20:15 It's just about the ROI to solve them. 20:18 Does it make sense? And if it does make 20:19 sense, companies, entrepreneurs will go 20:21 after those problems. And AI reduces the 20:26 cost to solve these long tale of 20:29 problems, which is good because more 20:33 people will have their problem solved no 20:34 matter how rare that problem is. And you 20:37 can think of a problem again as a 20:40 disease. If somebody has a rare disease 20:42 that only a 100 people in the world 20:44 have, AI makes the economics work to go 20:47 solve that disease for the 100 people. 20:49 Okay? And I highlighted this next part 20:52 in red because it literally contradicts 20:55 the previous point or really the whole 20:57 point of this essay of super 20:59 intelligence for everyone. No matter how 21:02 intelligent AI becomes, there will be 21:05 always a finite amount of compute and 21:08 therefore an opportunity cost for how we 21:10 use it. Right there, a finite amount of 21:12 compute. That means whoever has the most 21:15 capital gets to dictate where that 21:18 compute goes. Now, the capital doesn't 21:21 just get to dictate it. It has to 21:24 actually have a return on the capital. 21:25 So hopefully the argument is whatever 21:28 the most valuable allocation of that 21:30 capital is is kind of how the economy 21:32 would play out anyways. So that's where 21:35 the compute is going to go. He goes on 21:37 to say if people can use AI to invent 21:39 incredibly valuable new things then it 21:41 will make more sense to allocate it 21:43 towards that rather than automating 21:46 existing jobs. This I agree with. But 21:48 again, let's say somebody, a startup 21:51 comes up with an incredibly valuable way 21:54 to use artificial intelligence, 21:56 something that produces a high rate of 21:59 return, then all of a sudden the company 22:01 that already has the most compute, 22:03 already has the most capital, will 22:05 immediately see it. The barrier to entry 22:08 for that company, for that massive 22:10 company to come and compete directly 22:11 with the startup is effectively zero. 22:14 And thus it is once again whoever has 22:17 the most compute will be able to out 22:19 compete everybody else. So that is still 22:22 my biggest fear and I really think this 22:26 contradicts the entire notion of 22:28 everybody's going to have super 22:29 intelligence because everybody having 22:31 super intelligence basically means the 22:34 playing field is level on the 22:36 intelligence layer and then it's how 22:38 much compute can you throw at it. Now, 22:39 he goes on to talk about how jobs are 22:41 going to change. New jobs are going to 22:43 be created. It's not going to be AI just 22:46 automates everybody's jobs and 22:47 everybody's out of work. I don't agree 22:49 with that. I never have. And I like how 22:51 he explains it. A generation ago, there 22:54 were no app developers, social media 22:57 creators, electric vehicle technicians, 22:59 and data center operators. So, he's 23:03 saying with every technological 23:04 revolution, there have been new jobs 23:07 created. And with AI, it is no 23:10 different. In the near future, there 23:11 will be new jobs that aren't common 23:13 today, like oneperson product studios 23:15 designing custom toys, furniture, or 23:17 clothes, worldbuilders, and experienced 23:19 designers creating games, stories, and 23:21 adventures. Personal biologists using 23:23 super intelligence to formulate 23:25 personalized treatments, and much more 23:27 that we can't yet conceive. I agree with 23:30 everything he said here. I am very 23:32 optimistic about job prospects for 23:34 society. And he points to not too long 23:38 ago 90% of people were farmers growing 23:41 food to survive. Now it's singledigit 23:44 percentages provide food for the entire 23:47 world. And that is the point. We have 23:50 created technologies that have allowed 23:52 fewer people to gain much more leverage 23:54 over this one domain and everybody else 23:59 benefited and then so many other jobs 24:01 were created. What did not happen is 24:04 everybody who would have been a farmer 24:06 is just now out of a job completely. We 24:09 have new and different jobs. It's not 24:11 going to be perfect and he points this 24:13 out. Of course, some aspects of the way 24:15 we work will change just as it did with 24:17 computers, the internet, and any new 24:19 technology. This means people will have 24:21 to adapt and this will be challenging. 24:23 Another prediction he's making, and 24:26 another one that I completely agree 24:27 with, is that company sizes may shrink. 24:30 Just as they did in the transition from 24:32 industrial giants to tech companies, 24:35 although tech companies have massive 24:37 labor forces, this doesn't mean fewer 24:40 jobs overall, it implies a larger number 24:42 of companies with fewer people each. And 24:44 so, again, this goes back to that 24:47 longtail ROI question. There's going to 24:50 be many more companies with fewer people 24:52 in each able to achieve higher levels of 24:54 revenue, able to tackle problems where 24:56 previously the ROI was not there, but 25:00 now it is. And so I'm excited about 25:01 that. But it does take a lot of people 25:04 who want to be entrepreneurial and 25:06 hopefully becoming an entrepreneur just 25:09 becomes easier over time. So as I 25:11 mentioned earlier, the sentiment around 25:13 data centers and AI in general is 25:15 incredibly negative right now. And so he 25:17 tries to allay some of those fears and 25:19 address some of the anger that we're 25:20 seeing today. One example that he gives 25:22 is in Richland Parish, Louisiana, Meta 25:25 built a data center and because of the 25:28 additional revenue that that town made 25:31 from that data center, each teacher 25:33 received a $50,000 bonus this year 25:37 because of the increased tax revenue 25:38 from our investment. And so this is the 25:40 way if these big tech companies are 25:42 going to be putting data centers in 25:44 towns and in cities and in states and 25:48 they're facing a lot of friction in 25:49 doing so. The best way to remove that 25:52 friction is simply to increase the 25:54 quality of life there. And the 25:55 interesting thing is with the $50,000 25:57 bonus that the teachers made in this 25:59 town, the superintendent told us that 26:01 the teachers are now moving there from 26:03 across the country and they suspect they 26:05 will have one of the best schools in the 26:07 entire nation because of this. And we 26:09 could repeat this formula again and 26:12 again. Now, a lot of the arguments 26:14 against data centers I hear today are 26:16 around water and water usage. So, Mark 26:19 Zuckerberg says, not only are their 26:23 current data centers designed to be the 26:24 most water efficient in the world, they 26:26 are also committed to being water 26:28 positive, meaning that we'll restore 26:29 more water than we use in the wersheds 26:32 where we operate by 2030. And in areas 26:35 with high water stress, our goal is to 26:37 restore 200% of the water we use. Now 26:39 again, they could say this all day long. 26:41 And maybe I'm naive to believe it, but I 26:43 know that much of the new data centers 26:45 being created, being built right now, 26:47 are closed loop systems. They do not use 26:50 a ton of water. It's not like the GPUs 26:53 are thirsty sucking up all this water 26:56 and then they it just goes nowhere. 26:58 These are closed loop systems just like 27:00 your home PC. If you have a water cooled 27:03 PC, that is exactly how it works. Okay, 27:06 so now back to the contradiction, back 27:09 to where this argument really breaks 27:11 down. So here he's talking about 27:13 securing AI and cyber security, 27:15 bioteterrorism, and more. So with AI, 27:18 the defenders will also have more 27:19 compute to generate significantly more 27:21 intelligence from the same models. This 27:24 is exactly what I said earlier. When we 27:26 have cyber defenders that tend to be the 27:29 biggest companies in the world, they 27:30 will have more resources to harden their 27:32 systems much more quickly than attackers 27:35 can try to infiltrate the systems. And 27:38 that is because the resources of 27:41 defenders are just so much higher than 27:44 attackers. But again, this exact same 27:46 logic applies everywhere in society. If 27:50 everybody has access to super 27:52 intelligence, whoever has the most super 27:55 intelligence, aka the most compute, will 27:57 have a competitive advantage. He even 28:00 says, and in some cases, they'll have 28:02 more advanced models as well. They being 28:05 the defenders. And again, the 28:07 contradiction is so plainly obvious 28:10 right here. The general pattern is that 28:12 society ensures the defenders who keep 28:15 bad actors in check must always have a 28:18 greater balance of power and resources. 28:20 Certainly, if defenders can have a 28:23 greater balance of power and resources, 28:25 it's really not the defender, it's 28:28 whoever has the most capital. Here is 28:30 the part that I do agree with. He goes 28:32 back to the centralization of power with 28:34 artificial intelligence. Restricting 28:37 capabilities leads us down the path of 28:39 centralization and lack of checks and 28:42 balances. Restricting capabilities, he's 28:44 specifically calling out companies like 28:46 Enthropic who think they are the ones to 28:49 create super intelligence and nobody 28:50 else should. They are the ones who are 28:52 responsible enough to create it well to 28:55 align it and open source probably 28:57 shouldn't exist. This is likely what 29:00 Anthropic thinks. But then again, 29:02 another contradiction. The long-term 29:04 answer isn't to withhold capabilities, 29:06 but to establish a balance of power 29:08 where super intelligence is broadly 29:09 distributed. But the balance of power is 29:12 not having access to the model. It is 29:15 how much compute you have. That is 29:18 really what it comes down to. And 29:20 actually upstream from compute is 29:22 energy. He talks about reducing 29:23 biological and chemical risks because of 29:25 AI. This is one of the main arguments 29:28 that Anthropic puts forth as to why 29:30 opensource is actually really dangerous 29:33 because if everybody has access to 29:36 knowledge of how to build chemical or 29:38 biological weapons then the world is a 29:40 more dangerous place. And I just do not 29:42 agree with that at all. And here's the 29:44 reason. Even if you have the knowledge 29:46 of how to do something, you still have 29:49 to do it. And especially for something 29:51 like building a biological or chemical 29:54 weapon, you still need all of the 29:56 physical components to do it. You still 29:58 need people, researchers with the 30:01 capabilities, even if they are supported 30:03 by AI. You still need to have a certain 30:06 level of expertise. And putting all of 30:08 these physical components together is 30:11 extremely difficult and we should make 30:13 it more difficult. So, we should focus 30:16 on limiting the physical production and 30:18 distribution of harmful materials. I 30:21 expect it will be easier to regulate and 30:23 control physical components than the 30:25 spread of knowledge. What a banger 30:28 quote. Controlling the spread of 30:30 knowledge has never been easy. 30:32 Knowledge, truth, it generally finds a 30:35 way out. But controlling physical 30:38 components, controlling chemicals, 30:40 controlling the facilities and the the 30:43 tooling necessary to build these 30:45 facilities, that is much easier. And 30:48 we've seen that with nuclear weapons. 30:50 The knowledge of how to build it is out 30:52 there. But being able to actually get 30:54 your hands on nuclear components, the 30:57 tooling, the expertise, that's the 31:00 difficult part. So, in the last few 31:01 months, I've talked quite a bit about 31:03 the race between China and the US with 31:07 artificial intelligence. And here is 31:09 something really interesting. Nvidia and 31:12 Jensen do not believe in export 31:14 controls. They believe if we stop 31:17 sending our chips to China, they're 31:19 going to develop their own chip 31:21 producing capabilities, which that makes 31:24 a lot of sense. But we do have export 31:27 controls on chips today. And Mark 31:29 Zuckerberg says export controls on 31:32 silicon have been successful for slowing 31:35 the progress of foreign labs during this 31:38 critical period. So it is the right 31:40 strategic move to continue those. So 31:43 very interesting. He disagrees with 31:45 Jensen on this point. Jensen thinks we 31:47 should just ship the best chips we have 31:49 because otherwise China is going to 31:50 build their own. And if they're 31:52 dependent on US chips, that's really 31:55 good for the US. Now obviously this 31:57 really benefits Nvidia but there is an 32:00 argument that export controls have been 32:02 working. Now he also points out that 32:05 China is producing and building much 32:08 more energy than the United States and 32:11 campaigns for the US to get even more 32:13 aggressive with building out our own 32:15 energy infrastructure accelerate 32:17 building both energy and data centers to 32:20 remain competitive. And it seems like 32:22 we're having the opposite effect right 32:23 now because of the negative sentiment in 32:25 the United States of AI and data 32:27 centers. They're being banned across the 32:29 nation as we speak. This puts us at a 32:32 distinct disadvantage in the race for 32:34 super intelligence against China. He 32:36 also, this is super interesting, calls 32:40 for distillation to just be legal. And I 32:43 think a lot of you watching this video 32:45 will probably agree with what he says 32:46 here. For the US to lead in open source, 32:49 we need to rethink our policies in 32:50 several areas, including distillation 32:53 and data use and training. The ability 32:55 for models to learn from other models is 32:57 an important principle of how the 32:59 open-source ecosystem works. All AI 33:02 models are derived from human knowledge. 33:05 Some have tried to frame distillation as 33:07 harmful, some anthropic. But I think it 33:10 is important to protect the principle 33:12 that you can learn from anything you can 33:14 observe. This is how the world works. 33:16 and the US will not be able to lead if 33:19 we restrict ourselves on this front. He 33:21 is basically calling for distillation to 33:24 be legal, which is kind of wild. 33:27 Specifically calling out anthropic. I 33:29 mean, not by name, but that's who he's 33:31 referring to when he says some have 33:33 tried to frame distillation as harmful. 33:35 But he's saying, hey, distillation is a 33:36 good thing. And when you're coming from 33:39 the angle of building an open- source 33:41 ecosystem, distillation is very 33:43 beneficial for open source. for closed 33:45 source it is not and he's basically 33:47 making the argument the closed source 33:49 labs distilled on all of human knowledge 33:52 why can't we distill on them I don't 33:54 know makes sense to me I think we just 33:56 need to follow IP law whatever the law 33:58 is let's follow that and if we need to 34:01 adjust the law to allow for distillation 34:03 fine but if distillation is legal the 34:06 frontier labs are going to lose a lot of 34:08 their value because another company can 34:11 come along distill on the frontier model 34:14 that cost billions and billions of 34:16 dollars to build and build a competitive 34:18 model at a fraction of the price and in 34:21 a fraction of the amount of time. Now, 34:24 distillation does not mean Chinese AI 34:27 labs can distill. I don't think that's 34:28 what he's saying. I think he's saying 34:30 other US companies. And so, if we want 34:32 to prevent China from distilling US 34:35 models, maybe that's still a solid 34:38 argument to make. Next, he talks about 34:40 maintaining control of super 34:41 intelligence, recursive 34:43 self-improvement. Because once that 34:45 happens, once AI is able to improve 34:47 itself and just compound on that 34:49 improvement, then the chance that we're 34:51 going to be able to keep it aligned and 34:53 keep it under our control goes down 34:55 significantly. And so his solution is as 34:58 such. The simplest is that multiple labs 35:01 could achieve RSI around the same time. 35:04 If some or all of those super 35:06 intelligences were somehow benevolent, 35:08 then they could check and balance each 35:10 other. So, here's the problem with this. 35:11 Recursive self-improvement compounds 35:13 quickly. And whoever reaches recursive 35:16 self-improvement first, maybe the delta 35:19 between first and second place, right 35:21 when it happens, is very small. But once 35:24 it happens, there's no catching up. Any 35:27 gap between first place and second place 35:29 only expands from there. That is the 35:32 entire concept of recursive 35:34 self-improvement. So I actually don't 35:36 think this is true. Even if a bunch of 35:39 labs all around the same time reached 35:43 RSI, whoever reached it first would 35:46 continue their momentum, grow their 35:47 momentum, and maintain their lead. So 35:50 that's it. Overall, I love the vision. I 35:54 am excited about open source. I want 35:57 everybody to have access to incredible 36:01 intelligence, super intelligence, but 36:03 I'm just struggling to see how it 36:05 doesn't ultimately come down to whoever 36:08 has the most compute and energy wins and 36:10 any part of the economy that they want. 36:12 Now, what I'm hopeful for is that the 36:15 value of pointing the compute at some 36:18 part of the economy is going to benefit 36:21 society as a whole. Like if society 36:23 needs some problem to be solved and 36:26 there's many people who need it solved, 36:29 then it's going to make sense for 36:30 whoever holds the compute to point their 36:32 compute at solving that problem. Because 36:34 ultimately, if you are a hyperscaler, 36:38 you have all the compute, but you're not 36:40 solving problems that people actually 36:42 need, then you're not going to get a 36:44 return on that invested compute. 36:46 Anyways, let me know what you think. I'm 36:48 still thinking through this, but I just 36:50 think these arguments, at least some of 36:52 them, break down in some areas. And I 36:54 made an entire video about why I am so 36:57 excited about open source and why it is 36:59 so important. And check that out right