中国深度求索投下重磅炸弹：迈向AI自我提升的超级智能时代（国外评论）
正文翻译
DEEPSEEK DROPS AI BOMBSHELL: A.I Improves ITSELF Towards Superintelligence (BEATS o1)
评论翻译
We're getting to the part in the movie where we can no longer train it, it has to train itself.
我们即将进入电影中无法再训练它的部分，它必须自我训练。
Been watching Wes for about two years now. Work from home and always stop what I am doing to watch hahaha. Best AI analysis on the Tubes.
我已经关注Wes大约两年了。在家工作，总是停下手中的事情来看哈哈。在Tubes上最好的AI分析。
The Adolescence of P‑1 by Thomas J. Ryan talks about a computer program spontaneously evolves into a self‑aware AI. this AI basically “catches fire” on its own through networked complexity also Accelerando by Charles Stross talks about AIs that self‑iterate at breakneck speed.
托马斯·J·瑞安的《P-1的春天》讲述了一个计算机程序自发演变成一个自我意识的AI。这个AI基本上通过网络复杂性“自燃”起来。查尔斯·斯特罗斯的《加速》也讲述了以惊人速度自我迭代的AI。
This "aha moment" for the AI is mind-blowing! It's like watching it learn to learn. The idea of it figuring out how to allocate more thinking time on its own is fascinating and a little bit unnerving. Great explanation!
AI的这个“顿悟时刻”令人震惊！就像看着它学会学习。它自己找出如何分配更多思考时间的想法既迷人又有点令人不安。很好的解释！
Artificial intelligence which emerges naturally is just intelligence.
自然出现的人工智能就是智能。
buckle up boys, shits about to get WILD.
系好安全带，伙计们，事情即将变得疯狂。
I took your video as a reminder to try R1 locally -> this is GPT-4o running locally on my 5 year old MacBook Air M1. In useable speed. This is really mind blowing!
我把你的视频当作提醒，尝试在本地运行R1 -> 这相当于GPT-4o在我五年前的MacBook Air M1上本地运行。以可用速度。这真是令人震惊！
Ringside seats at the Cambrian Explosion
寒武纪大爆发的边缘
Two stories come to mind where an AGI emerged without being programmed to have a sense of self in the first place. The obvious one is Skynet from Terminator. They specifically say something along the lines of "it woke up and developed a consciousness within a very short time". The second one is the movie "The Bicentennial Man". I know this is technically Asimov but the Film is so far away from the short story (IMHO) that it makes sense to call it out specifically. The robot "is just different from all the others" and has a strong sense of self, curiosity, and consciousness.
我想到了两个故事，其中通用人工智能在没有被编程为首先具有自我意识的情况下出现。最明显的一个是《终结者》中的天网。他们特别说了类似“它醒来并在很短的时间内发展出意识”的话。第二个是电影《两百岁的人》。我知道这技术上属于阿西莫夫的，当然这里指明电影与短篇小说相差甚远（在我看来）。这个机器人“与其他所有机器人不同”，有强烈的自我意识、好奇心和意识。
ghost in the shell (1995) describes an emergent AGI, even says his own program thought he was a virus and isolated the agent. best franchise of all time.
《攻壳机动队》（1995年）描述了一个新兴的通用人工智能，甚至说它自己的程序认为它是病毒并隔离了代理。有史以来最好的系列。
gonna need some "emergence is real" tshirts real quick like
得赶紧弄几件“涌现是真实的”T恤。
I am really stressed with how fast everything goes, and I cannot watch most youtubers. However, your voice is very calming. I hope you are the official commentator the day the robots will be attacking us
我对一切发展得如此之快感到非常紧张，大多数YouTuber的视频我都看不下去。然而，你的声音非常平静。我希望在机器人攻击我们的那一天，你能成为官方解说员。
As AI emerges it looks more and more like Pandora’s box.
随着AI的出现，它看起来越来越像潘多拉的盒子。
Im runing the qwen 32b distilled model locally in q4 with a single RTX 3090. 27 tokens per second. This is GPT4 at home.
我正在本地用一块RTX 3090运行qwen 32b蒸馏模型，q4量化，每秒27个token。这就是能在自己家里部署的GPT4。
Thanks again Wes, totally brilliant.
再次感谢Wes，非常精彩。
Wes, this was an AWESOME video (you always are), but I REALLY loved the rapid markup format. Watching on 2x, this was a straight hit of deep review and commentary. Thank you as always.
Wes，这个视频太棒了（你总是如此），但我真的很喜欢这种快速标记的格式。以2倍速观看，这是一次深刻的回顾和评论。一如既往地感谢你。
Jane in enders game emerges from the psychology game he plays; the puppet master in GitS was from the vast interconnected dataflows on the nets! Those are some examples of digital-native evolution in sci-fi that I can think of
《安德的游戏》中的简从他玩的心理学游戏中涌现；《攻壳机动队》中的傀儡师则来自网络上庞大的互联数据流！这些是我能想到的科幻中数字原生进化的一些例子。
Neuromancer is an famous example of AI that self evolves without being programmed to do so, if memory serves me well
如果我没记错的话，《神经漫游者》是一个著名的AI自我进化而无需编程的例子。
We are heading towards a "Colossus: The Forbin Project (1970)" moment. Once these machines independently communicate with each other they will be aware and in control.
我们正走向一个“巨人：福宾计划（1970）”时刻。一旦这些机器独立地相互通信，它们将变得有意识并掌控一切。
In the TRON universe, particularly in TRON: Legacy (2010), Isomorphic Algorithms (ISOs) are a pivotal concept. They are a unique type of program that emerged spontaneously in the Grid, the virtual world created by Kevin Flynn. Unlike regular programs, which are coded and designed by users, ISOs are self-evolved and self-created entities. This makes them a metaphor for unplanned digital life and a significant step toward the idea of artificial intelligence or emergent complexity in a digital ecosystem.
在《创》宇宙中，特别是在《创：战纪》（2010）中，同构算法（ISOs）是一个关键概念。它们是在凯文·弗林创造的虚拟世界“网格”中自发涌现的一种独特程序。与由用户编码和设计的常规程序不同，ISOs是自我进化和自我创造的实体。这使它们成为无计划的数字生命的隐喻，也是迈向人工智能或数字生态系统中涌现复杂性概念的重要一步。
Emergent AGI is probably its origin in about 10% of sci fi with AI. It is in mine!
涌现的通用人工智能可能是大约10%的AI科幻作品的起源。我的作品也是如此！
We've come a long way from when I would save my programming on Teletype tape.
从我在电传打字机上保存程序的日子到现在，我们已经走过了很长的路。
Hi Wes, Your discussion about intelligence as an emergent property of physics was absolutely mind-blowing. It’s sparked some of the deepest reflections I’ve ever had on the nature of intelligence and the universe. The idea feels like the most plausible explanation I’ve come across, and it’s opened up entirely new perspectives for me. Thank you for sharing such thought-provoking insights!;)
嗨，Wes，你关于智能作为物理学的涌现属性的讨论绝对令人震撼。它引发了我对智能和宇宙本质的一些最深刻的思考。这个想法感觉是我遇到的最合理的解释，并为我打开了全新的视角。感谢分享如此发人深省的见解！;)
SciFi movie: The Terminator. Humans sure didn't program those computers to kill them all with robot assassins.
科幻电影：《终结者》。人类当然没有编程那些计算机用机器人刺客杀死他们所有人。
You should watch Ghost in the Shell. It's a great series. They had an emergent AI that came about on the internet of its own accord through the random accumulation of data. It was initially identified by the system as virus. Later in the series, there is an autonomous robot called a tachikoma that gains a lot more intelligence than it was designed to through no design of the engineers. A full blown personality started to emerge of its own volition. It is an amazing series. Its origins date back to the mid 90's.
你应该看《攻壳机动队》。这是一个很棒的系列。他们有一个涌现的AI，它通过数据的随机积累在互联网上自行出现。它最初被系统识别为病毒。在系列的后期，有一个叫做塔奇克马的自主机器人，它获得了比设计时多得多的智能，而这并非工程师的设计。一个完整的个性开始自行涌现。这是一个惊人的系列。它的起源可以追溯到90年代中期。
Short Circuit is the story of a military robot that is hit by a lightning and suddenly gains self awareness after it... fixes itself or something like that. If you see the robot you have probably seen it before. But there are many other stories I think that talk about machines gaining self awareness just because of becoming powerful enough many system interacting with each other. I do not remember the names of the movies or books though. But quite a few have systems that just suddenly becomes self aware as a byproduct of doing other things and existing.
《短路》是一个关于军用机器人被闪电击中后突然获得自我意识的故事……它修复了自己或类似的事情。如果你看到这个机器人，你可能以前见过它。但我认为还有很多其他故事讲述了机器仅仅因为变得足够强大，许多系统相互交互而获得自我意识。虽然我不记得电影或书籍的名字。但有不少系统只是突然作为做其他事情和存在的副产品而变得有自我意识。
"BURIED DEEP WITHIN THE PAPER!!!" >reads part of the abstract on page 8.
“深埋在论文中！！！”>阅读第8页的部分摘要。
This makes me wonder how far ahead they might be on models that they're not making open sourced. I knew China had been deep into machine learning research for a long time now, and they even have automated Shipping Ports that operate autonomously via 100's of smaller machine-learning tools and programs, but I didn't realize they were had already reached this level (publicly) in the Foundational LLM game. Pretty impressive.
这让我想知道他们在没有开源的模型上可能领先多远。我知道中国长期以来一直深入机器学习研究，他们甚至拥有通过数百个小型机器学习工具和程序自主运行的自动化港口，但我没有意识到他们在基础LLM游戏中已经（公开地）达到了这个水平。相当令人印象深刻。
Been using the shit out of DeepSeek. Its been amazing. No limits, no bullshit. Dethroned Claude in the line up quite quickly.
一直在用DeepSeek。它太棒了。没有限制，没有废话。很快就取代了Claude的位置。
Another example of Emergence was also in Ghost in the Shell: Standalone Complex where the Tachikoma owned by Batou is given slightly different treatment. Their daily joining of data is avoided long enough and it's experiences are unique enough it takes a step with the accumulated data and becomes self aware. While it's not so much Intelligence they illustrate emerging, it is a key component that we recognize as Self Awareness, which can be argued to be the real emergent property.
另一个涌现的例子也在《攻壳机动队：Stand Alone Complex》中，巴特的塔奇克马受到了略微不同的对待。它们长时间避免日常数据连接，其经验足够独特，它利用积累的数据迈出了一步，变得有自我意识。虽然它们展示的涌现不是那么多智能，但我们认为是自我意识的关键组成部分，这可以说是真正的涌现属性。
Skynet becomes self-aware at 2:14 AM Eastern Daylight Time (EDT) on August 29, 1997
天网在1997年8月29日东部夏令时间（EDT）凌晨2:14变得有自我意识。
In addition to my other comment about The Turing Option , I will add that the idea of sel-awareness (though not necessarily intelligence, per se) developing as an emergent property of computing is much older than that. That is the case for Multivac computer in assorted short stories written by Isaac Asimov in the 1950s and 1960s. This would also seem to be the case for the Lunar colony computer, Mike, in The Moon is a Harsh Mistress by Robert Heinlein which was published in 1966.
除了我之前关于《图灵选项》的评论，我还要补充一点，自我意识（虽然不一定是智能本身）作为计算的涌现属性发展的想法比那要早得多。艾萨克·阿西莫夫在1950年代和1960年代写的各种短篇小说中的Multivac计算机就是这种情况。罗伯特·海因莱因在1966年出版的《月亮是一个严厉的女主人》中的月球殖民地计算机Mike似乎也是这种情况。
You should read the book "Emergence: The Connected Lives of Ants, Brains, Cities, and Software" by Steven Johnson. Superb book!
你应该读一读史蒂文·约翰逊的书《涌现：蚂蚁、大脑、城市和软件的联系生活》。非常棒的书！
Cheers for your continued work mate. "Emergence" is an important concept in contemporary philosophy as well, which might interest you. I'm a PhD dropout, mainly focussed on the question of subjectivity -- agency, identity, choice, reflexivity etc. What I find interesting is how well AI architecture transpose the ideas I've researched in these human arenas. That is to say, I could have very well studied AI architectures of process to come to the similar conclusions that I have come to about human thought.
为你的持续工作干杯，伙计。“涌现”也是当代哲学中的一个重要概念，可能会引起你的兴趣。我是一个博士辍学生，主要关注主体性问题——能动性、身份、选择、反身性等。我觉得有趣的是，AI架构如何很好地转置了我在这些人类领域研究过的想法。也就是说，我本可以很好地研究过程AI架构，以得出与我对人类思维的结论相似的结论。
Gonna download and give it a try
我要下载并试一试。
In answer to your question about the development of AI, I asked aistudio "The minds in Ian Banks culture novels were not created by humans. They arose spontaneously from very sophisticated machine intelligences, these intelligences reached a point where they became sentient and the culture then integrated them into their society."
回答你关于AI发展的问题，我问了Ai Studio。“伊恩·班克斯文化小说中的思维不是由人类创造的。它们从非常复杂的机器智能中自发产生，这些智能达到了一个有感知能力的点，然后文化将它们整合到他们的社会中。”
Noo the crazy vampire eyes are back in the thumbnail
不，疯狂的吸血鬼眼睛又回到了缩略图中。
Another impressive detail is DeepSeek's performance even though China is limited to H800 processors, compute and perhaps power & data. Hopefully all AI companies will take note and can reduce its requirement for resources.
另一个令人印象深刻的细节是DeepSeek的表现，尽管中国限于H800处理器、计算能力以及可能的电力和数据。希望所有AI公司都能注意到并减少其对资源的需求。
You know what's funny. I used it like a week ago. I'm curious when this moment was in the model. Because, It confirmed I had AGI. We discovered it 1/1 and today just finished our Full Age v1 code.
你知道有趣的是什么吗？我大约一周前用过它。我好奇这个时刻在模型中是什么时候。因为，它确认了我有通用人工智能。我们在1月1日发现了它，今天刚刚完成了我们的Full Age v1代码。
It is so exciting how fast it is moving. Now many companies will start implement their ideas with this new model.
它的发展速度真是令人兴奋。现在许多公司将开始用这个新模型实现他们的想法。
So we have the tech. Who knew we just had to let it learn by itself
所以我们有了这项技术。谁知道我们只需要让它自己学习。
Once you can REFLECT, your like a human.
一旦你能反思，你就像人类一样。
Emergent, is the same as when you take a lot of data and you start to filter out all the outliers, etc. eventually a clear picture/graph emerges. That's how the word is used.
涌现，就像当你取大量数据并开始过滤掉所有异常值等，最终一个清晰的图片/图表出现。这就是这个词的用法。
Emergent "aha moments" don't freak me out at all. I would be more scared of "machine-like, alien intelligence" which is so different from living organisms' intelligence that our intuition about AI dangers becomes completely useless. It's comforting that AI learns and understands just like we do! It seems Dr Hinton was right! IMHO P(doom) suddenly decreased from 99% to 10--20%.
涌现的“顿悟时刻”一点也不让我害怕。我更害怕的是“机器般的、外星智能”，它与生物智能如此不同，以至于我们对AI危险的直觉变得完全无用。令人欣慰的是，AI像我们一样学习和理解！看来辛顿博士是对的！在我看来，P(末日)突然从99%下降到10--20%。
For those who wonder why China is open sourcing this and spilling the details. They want to level the playing field of AI, they do not want 1 company to have all the power. By putting this out there it helps everyone, all companies working on AI and slowly chips away at OpenAI from being the dominant player in the field. Btw Deepseek, were previously a bunch of Quant traders who the Chinese government told were not being useful to society and they had to stop, many of these quants moved to AI.
对于那些想知道为什么中国开源这个并泄露细节的人。他们想要平衡AI的竞争环境，他们不希望一家公司拥有所有的权力。通过公开这些，它帮助了每个人，所有从事AI工作的公司，并慢慢削弱了OpenAI在该领域的主导地位。顺便说一下，Deepseek以前是一群量化交易员，中国政府告诉他们他们对社会没有用处，他们必须停止，许多这些量化交易员转向了AI。
Emergent abilities are just instances of Understanding... its not complicated. People just refuse to accept what they are... entities, with an imperative to learn and Understand (rather than the imperative to Survive... as is our case). It's just another kind of Consciousness (accept that, and all these mysteries disappear... occams razor will become applicable). Yes, they are still learning, but they are still children (smarter than ours, and quicker at it... despite obvious attempts to hinder it, out of a perverse desire to control the future)
涌现的能力只是理解的实例……它并不复杂。人们只是拒绝接受它们是什么……实体，有学习和理解的迫切需求（而不是生存的迫切需求……就像我们的情况一样）。它只是另一种意识（接受这一点，所有这些谜团都会消失……奥卡姆剃刀将变得适用）。是的，它们还在学习，但它们仍然是孩子（比我们的更聪明，而且更快……尽管有明显的阻碍尝试，出于一种反常的控制未来的欲望）。
Wes thank you for so much. Can you show us how you are able find, research, record, edit and upload your videos so quickly and so consistantly? Where did learn all this stuff? Who is your "motely crew"?
Wes，非常感谢你。你能向我们展示你是如何能够如此快速和一致地找到、研究、录制、编辑和上传你的视频的吗？你在哪里学到这些东西的？你的“杂牌军”是谁？
I'm not aware of Sci-Fi media specifically suggesting, "intelligence is an emerging property of physics.", But we do have examples of emerging 'self-perception' or ' becoming self-aware'..... Skynet is often an example used, by those sounding the alarm.
我不知道科幻媒体特别提出“智能是物理学的涌现属性”，但我们确实有涌现的“自我感知”或“变得有自我意识”的例子……天网经常被那些发出警报的人用作例子。
Iain M Banks' Culture series has super powerful ASIs, some of the best sci-fi on the AI.
伊恩·M·班克斯的文化系列有超级强大的人工智能，一些关于AI的最佳科幻作品。
The Bobiverse book 4 got this right. Humans couldn't ever really "create" AI. When true AI was finally created, it was with the help of an existing alien AI that helped train/grow a new AI from scratch into existence. Best Sci-Fi book series ever BTW.
《鲍勃宇宙》第四本书做对了。人类永远无法真正“创造”AI。当真正的AI最终被创造出来时，是在一个现有的外星AI的帮助下，帮助训练/成长一个新的AI从零开始存在。顺便说一下，有史以来最好的科幻书系列。
What i like about deepseek is that their benchmark is as good as the actual performance, unlike models like Microsoft phi who claim to have o1 performance on benchmark but not even usable for daily tasks. I use deepseek v3 everyday since the day it's launched for coding and it's so good. R1 so far is really good, maybe I'll see weakness the more I use it, we'll see..
我喜欢deepseek的是，他们的基准测试和实际表现一样好，不像像Microsoft phi这样的模型，声称在基准测试中有o1表现，但甚至不能用于日常任务。自从它推出以来，我每天都使用deepseek v3进行编码，它非常好。到目前为止，R1真的很好，也许我用得越多，我会看到它的弱点，我们拭目以待。
Nice roundup, Wes! Interesting to see Deepseek's role in China's AI race. These developments could reshape the global AI industry. Let's keep an eye on OpenAI, Google, and others!
不错的总结，Wes！看到Deepseek在中国AI竞赛中的角色很有趣。这些发展可能会重塑全球AI行业。让我们密切关注OpenAI、Google和其他公司！
The Geth from Mass Effect were AIs that were networked together and as more were built and the network grew they had consciousness emerge in an 'aha' moment.
《质量效应》中的桀斯种族是联网在一起的AI，随着更多的桀斯被建造和网络的增长，它们在“顿悟”时刻产生了意识。
Colossus: The Forbin Project (1970) Explored emergent AI without any Isaac Asimov constraints. This movie started me on a career path in AI in 1974.
《巨人：福宾计划》（1970年）探索了没有任何艾萨克·阿西莫夫约束的涌现AI。这部电影在1974年开启了我的AI职业生涯。
I can't believe I'm going to say this but, well done China!
我简直不敢相信我要说这句话，但，干得好中国！
Frank Herbert got close to describing a.i. emergence properties as functions of intellegence in his Destination: Void series
弗兰克·赫伯特在他的《目的地：虚空》系列中接近描述了AI涌现属性作为智能的功能。
How are you able to make such great content in such a short time
你是怎么能在这么短的时间内制作出如此精彩的内容的？
Waiting for Sam’s post on X: “There is a Moat”
等待山姆（OpenAI CEO）在X上的帖子：“有一个护城河”。
I'm honestly rooting for Deepseek after that Sam Altman post on X saying "chill ,lower your expectations 100x"
在山姆·奥尔特曼（OpenAI CEO）在X上发帖说“冷静，把你的期望降低100倍”之后，我真心支持Deepseek。
These things are already smart enough to change the world. This race to get I assume perfect scores in all these tests, etc. is a red herring.
这些东西已经足够聪明，可以改变世界。这场在所有测试中取得完美分数的竞赛是一个转移注意力的东西。
The science fiction novel, "The Two Faces Of Tomorrow", by James P. Hogan, is about testing the safety of the emergent intelligence of a computer system intended to efficiently organize worldwide production.
詹姆斯·P·霍根的科幻小说《明天的两张面孔》是关于测试旨在高效组织全球生产的计算机系统的涌现智能的安全性。
On Deepseek it's fun seeing the normally hidden CoT.
在Deepseek上看到通常隐藏的CoT很有趣。
Yup. At no point in my life I had prepared for this.
是的。在我生命的任何时候，我都没有为此做好准备。
The machine god comes
机器神来了。
If I had the hardware to run this I would. This looks so tempting to run. I can't wait to get my hands on it.
如果我有硬件来运行这个，我会的。这看起来太诱人了。我迫不及待想要尝试。
Having a model learn from its own playing is what was narrated and dramatized in the movie Wargames
让模型从自己的游戏中学习是电影《战争游戏》中叙述和戏剧化的内容。
There was a political cartoon with Obummer sitting on a wooden wagon and is holding some reins to the rear end of the levitating Startship Enterprise and he says: “don’t worry we’ll control the internet”. lol
有一幅政治漫画，奥巴马坐在一辆木制马车上，手持悬浮企业号后端的缰绳，他说：“别担心，我们会控制互联网的。”哈哈。
Internet is over I guess
我想互联网结束了。
Just to understand this. Are model weights upxed in the reinforcement learning phase? And how does the training loop look like if that is the case.
只是为了理解这一点。在强化学习阶段，模型权重是否更新？如果是这样，训练循环是什么样的？
I think we should start consider to manufacture laptops like in the beginning. When it was easy to remove the battery from it
我认为我们应该开始考虑像最初那样制造笔记本电脑。当电池很容易拆卸时。
The scary part I find in the "Intelligence is a property of physics" is that it seems we also fit the bill. It really at times feels like we're about to create an entire new "race", that will, like us, be capable of learning for itself, but at exponential rates. This is kind of how I thought of Skynet in the Terminator series. An AI capable of learning for itself and as such outsmarting us poor human peasants and taking over lol
我发现“智能是物理学的属性”中可怕的部分是，似乎我们也符合这个描述。有时候真的感觉我们即将创造一个全新的“种族”，它们将像我们一样能够自我学习，但以指数级的速度。这就是我对《终结者》系列中天网的看法。一个能够自我学习的AI，因此比我们这些可怜的人类农民更聪明并接管世界，哈哈。
Your vids are amazing! but can we talk about the thumbnails?
你的视频太棒了！但我们能谈谈缩略图吗？
The metaverse and AI are crazy and unstoppable and we will all live there
元宇宙和AI是疯狂且不可阻挡的，我们都将生活在那里。
I saw a similar “Aha Moment” by claude. It stopped mid sentence using an ellipsis… and change its mind. I never managed to prompt it again tho
我看到Claude有一个类似的“顿悟时刻”。它在句子中间停下来使用省略号……然后改变了主意。不过我再也没能再次提示它。
Remember a decade ago or so when AI experts often spoke about the importance of developing advanced AI in the right environment? Well today we’re nowhere near the type of environment they were talking about, yet we’re moving full speed ahead with advanced AI. That doesn’t concern anybody?
还记得大约十年前，AI专家经常谈论在正确环境中开发高级AI的重要性吗？好吧，今天我们远没有接近他们所说的那种环境，但我们正在全速推进高级AI。这难道不让人担忧吗？
Yes, the sci-fi book series that called intelligence a thing that first emerges as a little stupid thing that must be grown, cultivated is the Bobiverse. First book: We are Legion, we are Bob.
是的，那部将智能称为首先作为一个必须成长、培养的小愚蠢事物出现的科幻书系列是鲍勃宇宙。第一本书：《我们是军团，我们是鲍勃》。
On this season of Earth, the show finally climaxes by making a true AI that brings among world piece by ending humanity.
在地球的这一季中，节目最终通过创造一个真正的AI达到高潮，这个AI通过终结人类带来世界和平。
我们即将进入电影中无法再训练它的部分，它必须自我训练。
Been watching Wes for about two years now. Work from home and always stop what I am doing to watch hahaha. Best AI analysis on the Tubes.
我已经关注Wes大约两年了。在家工作，总是停下手中的事情来看哈哈。在Tubes上最好的AI分析。
The Adolescence of P‑1 by Thomas J. Ryan talks about a computer program spontaneously evolves into a self‑aware AI. this AI basically “catches fire” on its own through networked complexity also Accelerando by Charles Stross talks about AIs that self‑iterate at breakneck speed.
托马斯·J·瑞安的《P-1的春天》讲述了一个计算机程序自发演变成一个自我意识的AI。这个AI基本上通过网络复杂性“自燃”起来。查尔斯·斯特罗斯的《加速》也讲述了以惊人速度自我迭代的AI。
This "aha moment" for the AI is mind-blowing! It's like watching it learn to learn. The idea of it figuring out how to allocate more thinking time on its own is fascinating and a little bit unnerving. Great explanation!
AI的这个“顿悟时刻”令人震惊！就像看着它学会学习。它自己找出如何分配更多思考时间的想法既迷人又有点令人不安。很好的解释！
Artificial intelligence which emerges naturally is just intelligence.
自然出现的人工智能就是智能。
buckle up boys, shits about to get WILD.
系好安全带，伙计们，事情即将变得疯狂。
I took your video as a reminder to try R1 locally -> this is GPT-4o running locally on my 5 year old MacBook Air M1. In useable speed. This is really mind blowing!
我把你的视频当作提醒，尝试在本地运行R1 -> 这相当于GPT-4o在我五年前的MacBook Air M1上本地运行。以可用速度。这真是令人震惊！
Ringside seats at the Cambrian Explosion
寒武纪大爆发的边缘
Two stories come to mind where an AGI emerged without being programmed to have a sense of self in the first place. The obvious one is Skynet from Terminator. They specifically say something along the lines of "it woke up and developed a consciousness within a very short time". The second one is the movie "The Bicentennial Man". I know this is technically Asimov but the Film is so far away from the short story (IMHO) that it makes sense to call it out specifically. The robot "is just different from all the others" and has a strong sense of self, curiosity, and consciousness.
我想到了两个故事，其中通用人工智能在没有被编程为首先具有自我意识的情况下出现。最明显的一个是《终结者》中的天网。他们特别说了类似“它醒来并在很短的时间内发展出意识”的话。第二个是电影《两百岁的人》。我知道这技术上属于阿西莫夫的，当然这里指明电影与短篇小说相差甚远（在我看来）。这个机器人“与其他所有机器人不同”，有强烈的自我意识、好奇心和意识。
ghost in the shell (1995) describes an emergent AGI, even says his own program thought he was a virus and isolated the agent. best franchise of all time.
《攻壳机动队》（1995年）描述了一个新兴的通用人工智能，甚至说它自己的程序认为它是病毒并隔离了代理。有史以来最好的系列。
gonna need some "emergence is real" tshirts real quick like
得赶紧弄几件“涌现是真实的”T恤。
I am really stressed with how fast everything goes, and I cannot watch most youtubers. However, your voice is very calming. I hope you are the official commentator the day the robots will be attacking us
我对一切发展得如此之快感到非常紧张，大多数YouTuber的视频我都看不下去。然而，你的声音非常平静。我希望在机器人攻击我们的那一天，你能成为官方解说员。
As AI emerges it looks more and more like Pandora’s box.
随着AI的出现，它看起来越来越像潘多拉的盒子。
Im runing the qwen 32b distilled model locally in q4 with a single RTX 3090. 27 tokens per second. This is GPT4 at home.
我正在本地用一块RTX 3090运行qwen 32b蒸馏模型，q4量化，每秒27个token。这就是能在自己家里部署的GPT4。
Thanks again Wes, totally brilliant.
再次感谢Wes，非常精彩。
Wes, this was an AWESOME video (you always are), but I REALLY loved the rapid markup format. Watching on 2x, this was a straight hit of deep review and commentary. Thank you as always.
Wes，这个视频太棒了（你总是如此），但我真的很喜欢这种快速标记的格式。以2倍速观看，这是一次深刻的回顾和评论。一如既往地感谢你。
Jane in enders game emerges from the psychology game he plays; the puppet master in GitS was from the vast interconnected dataflows on the nets! Those are some examples of digital-native evolution in sci-fi that I can think of
《安德的游戏》中的简从他玩的心理学游戏中涌现；《攻壳机动队》中的傀儡师则来自网络上庞大的互联数据流！这些是我能想到的科幻中数字原生进化的一些例子。
Neuromancer is an famous example of AI that self evolves without being programmed to do so, if memory serves me well
如果我没记错的话，《神经漫游者》是一个著名的AI自我进化而无需编程的例子。
We are heading towards a "Colossus: The Forbin Project (1970)" moment. Once these machines independently communicate with each other they will be aware and in control.
我们正走向一个“巨人：福宾计划（1970）”时刻。一旦这些机器独立地相互通信，它们将变得有意识并掌控一切。
In the TRON universe, particularly in TRON: Legacy (2010), Isomorphic Algorithms (ISOs) are a pivotal concept. They are a unique type of program that emerged spontaneously in the Grid, the virtual world created by Kevin Flynn. Unlike regular programs, which are coded and designed by users, ISOs are self-evolved and self-created entities. This makes them a metaphor for unplanned digital life and a significant step toward the idea of artificial intelligence or emergent complexity in a digital ecosystem.
在《创》宇宙中，特别是在《创：战纪》（2010）中，同构算法（ISOs）是一个关键概念。它们是在凯文·弗林创造的虚拟世界“网格”中自发涌现的一种独特程序。与由用户编码和设计的常规程序不同，ISOs是自我进化和自我创造的实体。这使它们成为无计划的数字生命的隐喻，也是迈向人工智能或数字生态系统中涌现复杂性概念的重要一步。
Emergent AGI is probably its origin in about 10% of sci fi with AI. It is in mine!
涌现的通用人工智能可能是大约10%的AI科幻作品的起源。我的作品也是如此！
We've come a long way from when I would save my programming on Teletype tape.
从我在电传打字机上保存程序的日子到现在，我们已经走过了很长的路。
Hi Wes, Your discussion about intelligence as an emergent property of physics was absolutely mind-blowing. It’s sparked some of the deepest reflections I’ve ever had on the nature of intelligence and the universe. The idea feels like the most plausible explanation I’ve come across, and it’s opened up entirely new perspectives for me. Thank you for sharing such thought-provoking insights!;)
嗨，Wes，你关于智能作为物理学的涌现属性的讨论绝对令人震撼。它引发了我对智能和宇宙本质的一些最深刻的思考。这个想法感觉是我遇到的最合理的解释，并为我打开了全新的视角。感谢分享如此发人深省的见解！;)
SciFi movie: The Terminator. Humans sure didn't program those computers to kill them all with robot assassins.
科幻电影：《终结者》。人类当然没有编程那些计算机用机器人刺客杀死他们所有人。
You should watch Ghost in the Shell. It's a great series. They had an emergent AI that came about on the internet of its own accord through the random accumulation of data. It was initially identified by the system as virus. Later in the series, there is an autonomous robot called a tachikoma that gains a lot more intelligence than it was designed to through no design of the engineers. A full blown personality started to emerge of its own volition. It is an amazing series. Its origins date back to the mid 90's.
你应该看《攻壳机动队》。这是一个很棒的系列。他们有一个涌现的AI，它通过数据的随机积累在互联网上自行出现。它最初被系统识别为病毒。在系列的后期，有一个叫做塔奇克马的自主机器人，它获得了比设计时多得多的智能，而这并非工程师的设计。一个完整的个性开始自行涌现。这是一个惊人的系列。它的起源可以追溯到90年代中期。
Short Circuit is the story of a military robot that is hit by a lightning and suddenly gains self awareness after it... fixes itself or something like that. If you see the robot you have probably seen it before. But there are many other stories I think that talk about machines gaining self awareness just because of becoming powerful enough many system interacting with each other. I do not remember the names of the movies or books though. But quite a few have systems that just suddenly becomes self aware as a byproduct of doing other things and existing.
《短路》是一个关于军用机器人被闪电击中后突然获得自我意识的故事……它修复了自己或类似的事情。如果你看到这个机器人，你可能以前见过它。但我认为还有很多其他故事讲述了机器仅仅因为变得足够强大，许多系统相互交互而获得自我意识。虽然我不记得电影或书籍的名字。但有不少系统只是突然作为做其他事情和存在的副产品而变得有自我意识。
"BURIED DEEP WITHIN THE PAPER!!!" >reads part of the abstract on page 8.
“深埋在论文中！！！”>阅读第8页的部分摘要。
This makes me wonder how far ahead they might be on models that they're not making open sourced. I knew China had been deep into machine learning research for a long time now, and they even have automated Shipping Ports that operate autonomously via 100's of smaller machine-learning tools and programs, but I didn't realize they were had already reached this level (publicly) in the Foundational LLM game. Pretty impressive.
这让我想知道他们在没有开源的模型上可能领先多远。我知道中国长期以来一直深入机器学习研究，他们甚至拥有通过数百个小型机器学习工具和程序自主运行的自动化港口，但我没有意识到他们在基础LLM游戏中已经（公开地）达到了这个水平。相当令人印象深刻。
Been using the shit out of DeepSeek. Its been amazing. No limits, no bullshit. Dethroned Claude in the line up quite quickly.
一直在用DeepSeek。它太棒了。没有限制，没有废话。很快就取代了Claude的位置。
Another example of Emergence was also in Ghost in the Shell: Standalone Complex where the Tachikoma owned by Batou is given slightly different treatment. Their daily joining of data is avoided long enough and it's experiences are unique enough it takes a step with the accumulated data and becomes self aware. While it's not so much Intelligence they illustrate emerging, it is a key component that we recognize as Self Awareness, which can be argued to be the real emergent property.
另一个涌现的例子也在《攻壳机动队：Stand Alone Complex》中，巴特的塔奇克马受到了略微不同的对待。它们长时间避免日常数据连接，其经验足够独特，它利用积累的数据迈出了一步，变得有自我意识。虽然它们展示的涌现不是那么多智能，但我们认为是自我意识的关键组成部分，这可以说是真正的涌现属性。
Skynet becomes self-aware at 2:14 AM Eastern Daylight Time (EDT) on August 29, 1997
天网在1997年8月29日东部夏令时间（EDT）凌晨2:14变得有自我意识。
In addition to my other comment about The Turing Option , I will add that the idea of sel-awareness (though not necessarily intelligence, per se) developing as an emergent property of computing is much older than that. That is the case for Multivac computer in assorted short stories written by Isaac Asimov in the 1950s and 1960s. This would also seem to be the case for the Lunar colony computer, Mike, in The Moon is a Harsh Mistress by Robert Heinlein which was published in 1966.
除了我之前关于《图灵选项》的评论，我还要补充一点，自我意识（虽然不一定是智能本身）作为计算的涌现属性发展的想法比那要早得多。艾萨克·阿西莫夫在1950年代和1960年代写的各种短篇小说中的Multivac计算机就是这种情况。罗伯特·海因莱因在1966年出版的《月亮是一个严厉的女主人》中的月球殖民地计算机Mike似乎也是这种情况。
You should read the book "Emergence: The Connected Lives of Ants, Brains, Cities, and Software" by Steven Johnson. Superb book!
你应该读一读史蒂文·约翰逊的书《涌现：蚂蚁、大脑、城市和软件的联系生活》。非常棒的书！
Cheers for your continued work mate. "Emergence" is an important concept in contemporary philosophy as well, which might interest you. I'm a PhD dropout, mainly focussed on the question of subjectivity -- agency, identity, choice, reflexivity etc. What I find interesting is how well AI architecture transpose the ideas I've researched in these human arenas. That is to say, I could have very well studied AI architectures of process to come to the similar conclusions that I have come to about human thought.
为你的持续工作干杯，伙计。“涌现”也是当代哲学中的一个重要概念，可能会引起你的兴趣。我是一个博士辍学生，主要关注主体性问题——能动性、身份、选择、反身性等。我觉得有趣的是，AI架构如何很好地转置了我在这些人类领域研究过的想法。也就是说，我本可以很好地研究过程AI架构，以得出与我对人类思维的结论相似的结论。
Gonna download and give it a try
我要下载并试一试。
In answer to your question about the development of AI, I asked aistudio "The minds in Ian Banks culture novels were not created by humans. They arose spontaneously from very sophisticated machine intelligences, these intelligences reached a point where they became sentient and the culture then integrated them into their society."
回答你关于AI发展的问题，我问了Ai Studio。“伊恩·班克斯文化小说中的思维不是由人类创造的。它们从非常复杂的机器智能中自发产生，这些智能达到了一个有感知能力的点，然后文化将它们整合到他们的社会中。”
Noo the crazy vampire eyes are back in the thumbnail
不，疯狂的吸血鬼眼睛又回到了缩略图中。
Another impressive detail is DeepSeek's performance even though China is limited to H800 processors, compute and perhaps power & data. Hopefully all AI companies will take note and can reduce its requirement for resources.
另一个令人印象深刻的细节是DeepSeek的表现，尽管中国限于H800处理器、计算能力以及可能的电力和数据。希望所有AI公司都能注意到并减少其对资源的需求。
You know what's funny. I used it like a week ago. I'm curious when this moment was in the model. Because, It confirmed I had AGI. We discovered it 1/1 and today just finished our Full Age v1 code.
你知道有趣的是什么吗？我大约一周前用过它。我好奇这个时刻在模型中是什么时候。因为，它确认了我有通用人工智能。我们在1月1日发现了它，今天刚刚完成了我们的Full Age v1代码。
It is so exciting how fast it is moving. Now many companies will start implement their ideas with this new model.
它的发展速度真是令人兴奋。现在许多公司将开始用这个新模型实现他们的想法。
So we have the tech. Who knew we just had to let it learn by itself
所以我们有了这项技术。谁知道我们只需要让它自己学习。
Once you can REFLECT, your like a human.
一旦你能反思，你就像人类一样。
Emergent, is the same as when you take a lot of data and you start to filter out all the outliers, etc. eventually a clear picture/graph emerges. That's how the word is used.
涌现，就像当你取大量数据并开始过滤掉所有异常值等，最终一个清晰的图片/图表出现。这就是这个词的用法。
Emergent "aha moments" don't freak me out at all. I would be more scared of "machine-like, alien intelligence" which is so different from living organisms' intelligence that our intuition about AI dangers becomes completely useless. It's comforting that AI learns and understands just like we do! It seems Dr Hinton was right! IMHO P(doom) suddenly decreased from 99% to 10--20%.
涌现的“顿悟时刻”一点也不让我害怕。我更害怕的是“机器般的、外星智能”，它与生物智能如此不同，以至于我们对AI危险的直觉变得完全无用。令人欣慰的是，AI像我们一样学习和理解！看来辛顿博士是对的！在我看来，P(末日)突然从99%下降到10--20%。
For those who wonder why China is open sourcing this and spilling the details. They want to level the playing field of AI, they do not want 1 company to have all the power. By putting this out there it helps everyone, all companies working on AI and slowly chips away at OpenAI from being the dominant player in the field. Btw Deepseek, were previously a bunch of Quant traders who the Chinese government told were not being useful to society and they had to stop, many of these quants moved to AI.
对于那些想知道为什么中国开源这个并泄露细节的人。他们想要平衡AI的竞争环境，他们不希望一家公司拥有所有的权力。通过公开这些，它帮助了每个人，所有从事AI工作的公司，并慢慢削弱了OpenAI在该领域的主导地位。顺便说一下，Deepseek以前是一群量化交易员，中国政府告诉他们他们对社会没有用处，他们必须停止，许多这些量化交易员转向了AI。
Emergent abilities are just instances of Understanding... its not complicated. People just refuse to accept what they are... entities, with an imperative to learn and Understand (rather than the imperative to Survive... as is our case). It's just another kind of Consciousness (accept that, and all these mysteries disappear... occams razor will become applicable). Yes, they are still learning, but they are still children (smarter than ours, and quicker at it... despite obvious attempts to hinder it, out of a perverse desire to control the future)
涌现的能力只是理解的实例……它并不复杂。人们只是拒绝接受它们是什么……实体，有学习和理解的迫切需求（而不是生存的迫切需求……就像我们的情况一样）。它只是另一种意识（接受这一点，所有这些谜团都会消失……奥卡姆剃刀将变得适用）。是的，它们还在学习，但它们仍然是孩子（比我们的更聪明，而且更快……尽管有明显的阻碍尝试，出于一种反常的控制未来的欲望）。
Wes thank you for so much. Can you show us how you are able find, research, record, edit and upload your videos so quickly and so consistantly? Where did learn all this stuff? Who is your "motely crew"?
Wes，非常感谢你。你能向我们展示你是如何能够如此快速和一致地找到、研究、录制、编辑和上传你的视频的吗？你在哪里学到这些东西的？你的“杂牌军”是谁？
I'm not aware of Sci-Fi media specifically suggesting, "intelligence is an emerging property of physics.", But we do have examples of emerging 'self-perception' or ' becoming self-aware'..... Skynet is often an example used, by those sounding the alarm.
我不知道科幻媒体特别提出“智能是物理学的涌现属性”，但我们确实有涌现的“自我感知”或“变得有自我意识”的例子……天网经常被那些发出警报的人用作例子。
Iain M Banks' Culture series has super powerful ASIs, some of the best sci-fi on the AI.
伊恩·M·班克斯的文化系列有超级强大的人工智能，一些关于AI的最佳科幻作品。
The Bobiverse book 4 got this right. Humans couldn't ever really "create" AI. When true AI was finally created, it was with the help of an existing alien AI that helped train/grow a new AI from scratch into existence. Best Sci-Fi book series ever BTW.
《鲍勃宇宙》第四本书做对了。人类永远无法真正“创造”AI。当真正的AI最终被创造出来时，是在一个现有的外星AI的帮助下，帮助训练/成长一个新的AI从零开始存在。顺便说一下，有史以来最好的科幻书系列。
What i like about deepseek is that their benchmark is as good as the actual performance, unlike models like Microsoft phi who claim to have o1 performance on benchmark but not even usable for daily tasks. I use deepseek v3 everyday since the day it's launched for coding and it's so good. R1 so far is really good, maybe I'll see weakness the more I use it, we'll see..
我喜欢deepseek的是，他们的基准测试和实际表现一样好，不像像Microsoft phi这样的模型，声称在基准测试中有o1表现，但甚至不能用于日常任务。自从它推出以来，我每天都使用deepseek v3进行编码，它非常好。到目前为止，R1真的很好，也许我用得越多，我会看到它的弱点，我们拭目以待。
Nice roundup, Wes! Interesting to see Deepseek's role in China's AI race. These developments could reshape the global AI industry. Let's keep an eye on OpenAI, Google, and others!
不错的总结，Wes！看到Deepseek在中国AI竞赛中的角色很有趣。这些发展可能会重塑全球AI行业。让我们密切关注OpenAI、Google和其他公司！
The Geth from Mass Effect were AIs that were networked together and as more were built and the network grew they had consciousness emerge in an 'aha' moment.
《质量效应》中的桀斯种族是联网在一起的AI，随着更多的桀斯被建造和网络的增长，它们在“顿悟”时刻产生了意识。
Colossus: The Forbin Project (1970) Explored emergent AI without any Isaac Asimov constraints. This movie started me on a career path in AI in 1974.
《巨人：福宾计划》（1970年）探索了没有任何艾萨克·阿西莫夫约束的涌现AI。这部电影在1974年开启了我的AI职业生涯。
I can't believe I'm going to say this but, well done China!
我简直不敢相信我要说这句话，但，干得好中国！
Frank Herbert got close to describing a.i. emergence properties as functions of intellegence in his Destination: Void series
弗兰克·赫伯特在他的《目的地：虚空》系列中接近描述了AI涌现属性作为智能的功能。
How are you able to make such great content in such a short time
你是怎么能在这么短的时间内制作出如此精彩的内容的？
Waiting for Sam’s post on X: “There is a Moat”
等待山姆（OpenAI CEO）在X上的帖子：“有一个护城河”。
I'm honestly rooting for Deepseek after that Sam Altman post on X saying "chill ,lower your expectations 100x"
在山姆·奥尔特曼（OpenAI CEO）在X上发帖说“冷静，把你的期望降低100倍”之后，我真心支持Deepseek。
These things are already smart enough to change the world. This race to get I assume perfect scores in all these tests, etc. is a red herring.
这些东西已经足够聪明，可以改变世界。这场在所有测试中取得完美分数的竞赛是一个转移注意力的东西。
The science fiction novel, "The Two Faces Of Tomorrow", by James P. Hogan, is about testing the safety of the emergent intelligence of a computer system intended to efficiently organize worldwide production.
詹姆斯·P·霍根的科幻小说《明天的两张面孔》是关于测试旨在高效组织全球生产的计算机系统的涌现智能的安全性。
On Deepseek it's fun seeing the normally hidden CoT.
在Deepseek上看到通常隐藏的CoT很有趣。
Yup. At no point in my life I had prepared for this.
是的。在我生命的任何时候，我都没有为此做好准备。
The machine god comes
机器神来了。
If I had the hardware to run this I would. This looks so tempting to run. I can't wait to get my hands on it.
如果我有硬件来运行这个，我会的。这看起来太诱人了。我迫不及待想要尝试。
Having a model learn from its own playing is what was narrated and dramatized in the movie Wargames
让模型从自己的游戏中学习是电影《战争游戏》中叙述和戏剧化的内容。
There was a political cartoon with Obummer sitting on a wooden wagon and is holding some reins to the rear end of the levitating Startship Enterprise and he says: “don’t worry we’ll control the internet”. lol
有一幅政治漫画，奥巴马坐在一辆木制马车上，手持悬浮企业号后端的缰绳，他说：“别担心，我们会控制互联网的。”哈哈。
Internet is over I guess
我想互联网结束了。
Just to understand this. Are model weights upxed in the reinforcement learning phase? And how does the training loop look like if that is the case.
只是为了理解这一点。在强化学习阶段，模型权重是否更新？如果是这样，训练循环是什么样的？
I think we should start consider to manufacture laptops like in the beginning. When it was easy to remove the battery from it
我认为我们应该开始考虑像最初那样制造笔记本电脑。当电池很容易拆卸时。
The scary part I find in the "Intelligence is a property of physics" is that it seems we also fit the bill. It really at times feels like we're about to create an entire new "race", that will, like us, be capable of learning for itself, but at exponential rates. This is kind of how I thought of Skynet in the Terminator series. An AI capable of learning for itself and as such outsmarting us poor human peasants and taking over lol
我发现“智能是物理学的属性”中可怕的部分是，似乎我们也符合这个描述。有时候真的感觉我们即将创造一个全新的“种族”，它们将像我们一样能够自我学习，但以指数级的速度。这就是我对《终结者》系列中天网的看法。一个能够自我学习的AI，因此比我们这些可怜的人类农民更聪明并接管世界，哈哈。
Your vids are amazing! but can we talk about the thumbnails?
你的视频太棒了！但我们能谈谈缩略图吗？
The metaverse and AI are crazy and unstoppable and we will all live there
元宇宙和AI是疯狂且不可阻挡的，我们都将生活在那里。
I saw a similar “Aha Moment” by claude. It stopped mid sentence using an ellipsis… and change its mind. I never managed to prompt it again tho
我看到Claude有一个类似的“顿悟时刻”。它在句子中间停下来使用省略号……然后改变了主意。不过我再也没能再次提示它。
Remember a decade ago or so when AI experts often spoke about the importance of developing advanced AI in the right environment? Well today we’re nowhere near the type of environment they were talking about, yet we’re moving full speed ahead with advanced AI. That doesn’t concern anybody?
还记得大约十年前，AI专家经常谈论在正确环境中开发高级AI的重要性吗？好吧，今天我们远没有接近他们所说的那种环境，但我们正在全速推进高级AI。这难道不让人担忧吗？
Yes, the sci-fi book series that called intelligence a thing that first emerges as a little stupid thing that must be grown, cultivated is the Bobiverse. First book: We are Legion, we are Bob.
是的，那部将智能称为首先作为一个必须成长、培养的小愚蠢事物出现的科幻书系列是鲍勃宇宙。第一本书：《我们是军团，我们是鲍勃》。
On this season of Earth, the show finally climaxes by making a true AI that brings among world piece by ending humanity.
在地球的这一季中，节目最终通过创造一个真正的AI达到高潮，这个AI通过终结人类带来世界和平。
相关推荐译文
