显示语言 中英对照 中文 English
点击段落跳转到对应时间
0:00 好，我先把一件事说清楚。Claude Opus 5.5 是有史以来最好的 AI 模型，而且现在你几乎干什么事都应该用它。我运气很好，Anthropic 提前给了我这个模型的测试资格。 All right, so I want to make one thing clear here. Claude Opus 5.5 is the best AI model ever released, and it is the one you should be using for pretty much everything right now. So I was very lucky. Anthropic gave me early access to this model. 0:12 我一直在不停地用它，前几天他们把权限收回去的时候，我是真的难过。咱们从头说起，接下来我会做演示，带你完整走一遍，看看它做出来的东西。 I have been using it nonstop, and when they took it away a couple days ago, I legitimately felt sadness. Just to start off with here, and I'll go through demos, take you through it, all that, show you what it's built. 0:22 它比 Fable 聪明，而且差距不是一点半点；它也比 ChatGPT 6 Astra 聪明。没错，它就是目前最聪明的 AI 模型。 It is smarter than Fable, it's not even really close, and it is smarter than ChatGPT 6 Astra. That's right, it is the smartest AI model out there. 0:32 但与此同时，它明显更快，也明显更便宜。Anthropic 给我早期测试资格的时候，根本没告诉我它叫什么。他们就说：「嘿，这是一个模型，你测一测，告诉我们你的想法。」 And yet at the same time, it is significantly faster, and it is significantly cheaper. So when Anthropic gave me early access to this model, they didn't tell me what the name was. They said, "Hey, here's a model. Test it out. Let us know what you think." 0:43 我当时几乎可以拿一切打赌，觉得这肯定是 Fable 5.5 或者 Fable 6 之类的。因为它比我用 Fable 得到的结果强太多了，简直是巨大的飞跃。 And I could have sworn on everything this was going to be like Fable 5.5 or Fable 6 or whatever. Like it was just such a massive leap over the results I was getting from Fable. 0:55 然后今天早上，他们给我发消息说：「对了，那个其实是 Opus 5.5。」我简直不敢相信。我第一反应就是：「不可能，这怎么可能是 Opus。」要知道，它比 Fable 还聪明。 Then this morning, they messaged me, and they go, "By the way, that was Opus 5.5." I could not believe it. It's the first thing I said, I'm like, "No, there's no way this is Opus." I mean, it was smarter than Fable. 1:05 我把 Astra 写的代码丢给 Opus，它直接说：「这代码写得太烂了，我来帮你收拾一下。」 I was taking code that Astra wrote, I was giving it to Opus, and it was going, "Oh, this code's disgusting. Let me clean this up for you." 1:12 基本上其他任何模型给我写的代码，它都能到处挑出问题。速度快得离谱，我印象里 Opus 从来没这么快过。 Like it was just finding issues left and right with basically all the code any other model has written for me. The speed is unbelievable. I mean, I can't remember Opus ever being this fast. 1:23 跟 Fable 或 Astra 比都没必要了，它比那两个快了不知道多少光年。但就算跟之前的 Opus 模型比，它感觉也更利落。 I mean, let's not even compare it to Fable or Astra because it is just light years faster than both of those. But even comparing it to previous Opus models, it just feels zippier. 1:33 我早期测试的时候就在想：「他们肯定有了什么突破，因为这个速度快得不可思议。」然后他们今天发布了。我终于拿到了 benchmark 数据，看到发布结果我就……我真的不敢相信。 Like I was testing it early, and I thought, "Oh, they must have had some sort of breakthrough here 'cause the speed is incredible." And so they release it today. I finally get my hands on the benchmarks, and they release it, and I'm just like... I just can't believe it. 1:45 他们居然还降价了。所以它不光单位任务的成本比市面上任何模型都低，连标价本身也直接降了。这种事你真的从来没见过。 They actually reduced the price. So not only is it cheaper per task than any other model out there, they also just reduced the straight-up price. You literally never see this. 1:56 他们的决定是：「嘿，这个模型我们就少赚点。」我猜他们的想法是——这个想法很棒——如果模型足够好，大家自然就会用得更多，用量上去了，钱也赚得更多。 They decided, "Hey, we're just gonna make less money on this model." I guess the idea is, and this is a great idea, if this model is so good, people are just gonna use it more, and we'll get more usage out of it and make more money off of it. 2:09 这说得通。对用户友好，对公司也友好。价格降了，单任务成本降了，智能水平却大幅提升。 And it makes sense. It's both consumer-friendly and friendly for the company. The price goes down, the cost per task goes down, and the intelligence goes way up. 2:20 你们可以看到，根本没有能跟它接近的。这边是 Astra，它比 Astra 聪明。 As you can see here, there's nothing even close. You can see Astra. It's smarter than Astra. 2:24 Opus 5.5 的高档位比 Astra 6 的最高档还强，而成本只是人家的零头。说真的，这种事太少见了。 The high setting on Opus 5.5 is better than the max setting on Astra 6, all while being like a fraction of the cost. I mean, you very rarely see this. 2:35 这等于是速度、智能、成本三项全能，全都更好，全都提升，全都是同级最强。太不可思议了。 This is like the trifecta of great speed, intelligence, and cost, all better, all improved, all like best in class. That is unbelievable. 2:46 如果你打算用 Claude Opus 5.5，你肯定想把它用到极致，而用到极致的方法就是 Claude Skills。很多人不知道怎么配置 Claude Skills。 And if you're gonna be using Claude Opus 5.5, you wanna make sure you get the absolute best out of it, and the way you get the best out of it is with Claude Skills. Many people don't know how to set up Claude Skills. 2:56 这就要靠我在 HubSpot for Startups 的朋友们了。他们刚推出一份免费资源，里面有五个 Claude Skills，帮你把创始人主导的内容做出彩。 That's where my friends at HubSpot for Startups come in. They just put out this free resource that gives you five Claude Skills that help you crush founder-led content. 3:05 装好这五个技能之后，你就不用每次从零开始给 Claude 写提示词了，而是跟一个已经懂你工作流程的系统协作。 And once you have these five skills installed, you stop prompting Claude from scratch every time and start working with a system that already knows your workflows. 3:14 里面有五个可安装的技能，能把你的原始素材变成贴合你个人风格、可以直接发渠道的内容，全程不到两小时。其中有个声音构建器，写出来真的像你本人。 Inside, you're going to get five installable skills that turn your raw work into voice-matched, channel-ready content in under two hours. You got a voice builder that actually sounds like you. 3:26 看看这个输出。还有个每周素材收集技能，能把你的笔记、Slack 消息什么的变成真正可用的内容。我最喜欢的是钩子和角度生成器，它让我做内容变得太轻松了。 Look at this output. A weekly capture skill that takes your notes, Slack messages, whatever, and turns into actual real content. My favorite is the hook and angle generator. This has made creating content so easy for me. 3:38 我随便给它一些粗略的想法，它就给我能用的钩子、角度和草稿。做内容真的变得太舒服了。 I just give it kinda rough ideas, and it gives me hooks I can use, angles, drafts. It just makes creating content so nice. 3:45 点击下方链接领取全部五个技能，也感谢 HubSpot 赞助本期视频。不过说了这么多，这些还不是这个模型最厉害的地方。 Grab all five skills from the link down below, and thanks to HubSpot for sponsoring the video. And now despite all of that, that's not even the best part of this model. 3:56 过去一周我一直在不停地跟它聊天，跟你说吧：它是有史以来聊起来最舒服的模型。 I've been talking to it nonstop for the last week, and let me tell you this. It is the best model to talk to ever. 4:03 过去几个月 AI 出现了一个很离谱的问题，真的把我烦透了。 There's been like this crazy issue going on with AI the last few months, and it's been pissing me off so much. 4:09 基本上，从 Opus 4.8 之后发布的每个模型，聊起来都让人受不了，而且这是所有公司都有的毛病。ChatGPT 是这样，Grok 也是这样。 Basically, every model to come out since like Opus 4.8 has been unbearable to talk to, and this is across all of the different companies. This is ChatGPT. This is Grok. 4:21 全都这样。呃，基本上从 Opus 4.8 之后——如果没记错的话大概是五月份——任何一家公司出的模型，聊起来都让人受不了。 This is all of them. Uh, basically every model to come out from any company since Opus 4.8, which if my memory serves me correct was like May, has been unbearable to talk to. 4:33 它们都形成了一种奇怪的腔调，根本不像人说话。全是些奇怪的术语，smoke test 这个词能给你说一千遍。 They've all developed like this weird language that they talk in that's like not even human. They use all these strange terms. They use smoke test a thousand times. 4:42 它们就是爱用这种莫名其妙、没什么意义的行话。你读这些模型输出的一大段一大段文字，读到最后会想：「我刚才到底读了什么？」这种现象到头了。 They just like use this weird jargon that doesn't really make sense. And you read like paragraphs of text from all these models, and you read it to the end, and you go, "What did I just read?" That is dead. 4:54 结束了，他们在 Opus 5.5 上把这个问题解决了。Anthropic 的模型当年最了不起的地方，也是我大约一年前死心塌地用它的原因，就是它一直是聊起来最舒服的模型。 That is over. They fixed it with Opus 5.5. What was amazing about Anthropic models and why I like ride or died with Anthropic like a year ago was it was just straight up the best models to talk to. 5:06 它们给人感觉温暖、像人，说话简洁。跟这些模型聊天就是很愉快，这方面比其他公司高出一大截。后来他们把这点丢了。 They felt warm. They felt human. They talked concisely. You... And it was just like pleasant to talk to the models, more on a level than any other company out there. They lost it. 5:17 这让我特别恼火，而现在它终于回来了。Opus 5.5 聊起来不亚于 Opus 4.5，甚至更好，也不输 Sonnet 3.5——那个我最爱聊天的元老级模型。 It pissed me off so much, and now it is finally back. Opus 5.5 is just as good if not better to talk to than Opus 4.5, than Sonnet 3.5, the OG model I love talking to. 5:31 它又说人话了，这点确实更好。接下来我就用一直在做的这个现场演示，给大家好好过一遍它到底强在哪儿。 It's better that it talks human again. So I'm gonna go through so many of the things it does better right now with this live demo I've been doing. 5:38 过去这一周，我基本都在用它做一款相当复杂的游戏，当然还做了很多别的事，但我觉得这个游戏最适合拿来演示。而且它的表达方式读起来特别顺，特别舒服。 So I have been basically been building, uh, this really complex game for the last week with it as ma- as well as many other things too, but I think this makes for the best demo. But the way it talks is so easy to read, so nice. 5:50 “完整构建和四个测试正在后台跑，大概要 30 分钟。跑完我会收到通知，目前还没有测试结果。这次构建会替换正在运行的二进制文件，所以我得先把你的游戏关掉。”你看，它说话就跟真人一样。 "The full build and four tests are running in the background, about 30 minutes. I'll be notified when they finish, and there are no test results yet. The build replaces the running binary, so I had to close your game." Like it just talks like a human being. 6:01 要是换成 Opus 5、Astra 6 或者 Grok 4.6 之类的——虽然它们也都是非常非常好的模型——但它们说话的方式是这样的：“完整冒烟测试已通过。 If this was Opus 5 or Astra 6 or Grok 4.6 or whatever, while those are great, great, great models- Like, the way they talk is like, "The full smoke test has passed. 6:13 冒烟测试已通过功能验证，二进制文件已重置边界，现在需要你自行进行冒烟测试以确认可用性。” The smoke test has passed the functionality, and the binary has reset the bi- boundaries, and it is now up to you to be smoke test to make sure it is available." 6:21 就是一堆多余的词和术语。而这个就非常直白：“嘿，我们构建好了，看起来不错。这是我们改进的所有内容。我会看一下截图、核对数据，等跑完就重新启动游戏。所有测试都通过了。 And like, it's just like all these extra words and jargon and this is just straight up, "Hey, we built it. Looks good. Here's everything we improved. I'll look at the screenshots, check the numbers, and relaunch the game when it finishes. Every test passed. 6:34 这是我们做的事情。”它还会配上漂亮的图表和表格，非常好读。感觉他们训练这个模型就是为了让它成为史上读起来最轻松的模型。要点列表也让阅读体验特别好。我说：“嘿，这个游戏有不少问题。 Here's the things we did." It has, like nice charts and tables, very easy to read. It, it's like they trained this just to be the easiest to read model ever. Bullet points make it super easy to read as well. I say, "Hey, there's a lot of issues with the game. 6:46 我们能改进什么？嘿，我测试了游戏。这是我看到的所有问题。你受到了 601 点伤害。”它就是像真人一样跟你说话。它调试代码的能力也非常惊人。 What can we improve? Hey, I tested the game. Here's all the issues I'm seeing. You took 601 damage." Like, it just talks like a human being. It's also spectacular at debugging. 6:57 就像我前面说的，我把一堆用 Astra 6.0 写的代码交给了它——再说一次，Astra 6.0 非常厉害，在这之前一直是我的日常主力模型——但它接手了所有代码，一下子就找出了超多问题。 I handed, as I said earlier, a bunch of code that I wrote with Astra 6.0, which again, Astra 6.0, incredible. It was my go-to model up until this point, but it took all the code. It just found so many issues. 7:07 我真的印象特别深。我没让它调试，只把代码给它：“能改进一下吗？”它说：“我发现了这些问题，要我帮你修吗？”太强了，又聪明又有人味儿。 Like, I, I, I really, really imp- And I didn't even ask it to debug. I just handed it and said, "Hey, can you improve these things?" Like, "Oh, by the way, I found all these things. Do you want me to fix it?" It's just so good. It's so smart. It's so intelligent and feels human. 7:18 所以我想给大家展示一个例子，看看它只用几条提示词就能给出多深的思考。这是一款完整的俯视视角 3D 撤离射击游戏，它只用几条提示词就构建出来了。 So I want to show you an example of like the depth and thinking it gives you in just a couple prompts. This is a full top-down 3D extraction shooter it built in a few prompts. 7:29 这个游戏的深度非常高，甚至比我用 Astra 做出来的还要深，而且完成得很漂亮，bug 还特别少。我觉得这一点有点被低估了——它写的代码几乎没 bug。 There's so much depth to this game, even more depth than I was getting out of Astra, and it does it so nicely with so few bugs. I think that's like kind of the underrated part here is like it writes bug-free code. 7:41 我不知道他们是怎么训练这个模型的，但它交付给我的无 bug 代码量，是我以前从没达到过的。那我们直接来看看。进游戏之后，它有一整套武器库存系统。 I don't know how they trained this model, but it's like shipping me the most code that's bug-free that I've ever done before. So let's just go ahead and just look at this. So you go in, it has this like entire weapon inventory system. 7:53 你可以直接进来，这个游戏的内容真的很有深度。第一眼可能看不出来，但你点进来看看——先看看游戏里所有的物品，再看看地图和里面的东西，有不同的地点。现在敌人出来了。 You can go right in here, and there's so much depth to this game. It might not like look at it at first, but if you go here, just first of all, look at all the items in the game, look at the map and everything in it. It has different locations. Now we got enemies. 8:05 我可以击杀敌人。还有一整套完整的战利品系统、战利品掉落。你可以把它放进你的背包——如你所见，还有完整的装备系统等等。它能给我做出这样的东西，真的太惊人了。 I can kill the enemy. There's like a full like loot system, the loot reels. You can put it in your inv- As you can see, there's like full equipment and all that. It's really amazing what it was able to build for me. 8:15 3D 建模也基本和 Astra 6.0 一样好。还有完整的撤离系统，我可以按住 E 键撤离，带着进来的所有战利品撤出去。顺便问一下，你们会玩这个游戏吗？在评论区告诉我。 The 3D modeling also basically just as good as Astra 6.0. Has a full extraction system. I can hold E to extract, and I get out with all the loot I came in with. By the way, would you play this game? Let me know down below. 8:27 还有一点是它提出新点子的能力。如果你跟我一样，会跟模型做大量头脑风暴，那 Opus 5.5 在出新点子这件事上简直强得离谱。 One other thing is its ability to come up with novel ideas. If you're anything like me, you do tons of brainstorming with your models. Opus 5.5 has been unbelievable at coming up with novel ideas. 8:39 我刚买了新款 Apple Watch Ultra 4，就去找 Opus 5.5，我说：“嘿，你能不能帮我想想怎么折腾这个东西？可以往上装什么应用，能做哪些 AI 相关的玩法。” So I just bought the new Apple Watch Ultra 4. I go to Opus 5.5. I say, "Hey, can you help me come up with ideas for how we can hack this? Apps we can put on it, different AI things we can do." 8:50 它提出了一堆点子，还真给我做出来了。最后我们做出来的效果是：我现在可以在手腕上直接跟我的智能体对话。 It comes up with all these ideas, builds it out for me, and what we end up doing is I can now talk to my agents on my wrist. 8:59 它给我的 Apple Watch 做了一个定制应用，我所有的智能体在上面都能看到——它们写代码或者干别的活儿时我都能实时看到——我还能按个按钮跟它们说话、发消息，出门在外不用带手机。 So all my agents, it built this custom app for the Apple Watch where I can see all my agents working as they're building code or whatever, and I can hit a button and talk to them and text them while I'm on the go, no phone required. 9:11 它就是能想出这些特别有意思的点子。我说：“行啊，听起来不错。”它就动手做出来了。它自己想办法把开发者应用装到了我的手机和手表上。我自己压根不知道怎么往手表上装应用。 It just like came up with all these really interesting ideas. I said, "Yeah, that sounds good," and it went ahead, built it. Figured out a way to like put developer apps on my phone and on my watch. I have no idea how to put apps like on my watch. 9:22 我完全不懂 Apple——iOS 手表是怎么运作的。它就直接说：“嘿，顺便说一下，我把应用装到你手表上了。”我说：“不可能。”结果我一低头，应用真的就在我手表上。我完全不知道怎么回事，我也没把手表插到任何东西上，它自己就搞定了。 I don't know how Apple o- iOS watches work. It just was like, "Hey, by the way, I put the app on your watch." I'm like, "No, you didn't." And I look down, and there's just the app on my watch. I have no idea. I didn't plug my watch into anything. It just figured it out. 9:34 所以它超级聪明、超级快、超级便宜，还能想出这些新点子。我说：“嘿，我这儿大概有 17 台电脑。 So it's super intelligent, super fast, super cheap. Comes up with these novel ideas. I said, "Hey, I have like 17 computers here. 9:42 我们能用本地模型做些什么很酷的事？”它就帮我搭好了这一整套本地模型系统。如果你会用 AI 来碰撞想法——你绝对应该这么做——就随口一问：“嘿，你会怎么用这个？你会在这儿做点什么？ What are some cool things we can do with local models?" It set up this entire local model system for me. If you use AI to bounce ideas, which you definitely should be doing, just coming up, "Hey, how would you use this? What would you create here? 9:54 你会在这儿搭什么？”在这方面它就是有史以来最好的模型。不过有一个缺点：我是否认为 Claude Code 的配套工具和 ChatGPT 的一样好？说实话，不是。 What would you build here?" This is like the best model for that ever. Now, the one downside is, is like do I believe the Claude code harness is as good as the ChatGPT harness, for instance? No. 10:08 ChatGPT 的配套工具有个语音模式，好用到令人震惊。 The ChatGPT harness has like voice mode, which is just like mind-blowingly good. 10:12 Claude 确实也有语音模式，但还没 ChatGPT 的语音模式那么好，不过我预计 Anthropic 很快会在这方面推出改进。 Claude does have a voice mode, but it is not quite as good as ChatGPT's voice mode, although I'd expect Anthropic to come out with improvements for that soon. 10:22 所以配套工具确实还差点意思，但这个模型实在太强了，我认为你必须用起来。 So the, the harness isn't quite there yet, but this model is so good that I do believe you need to be using it. 10:29 它基本上解决了 Anthropic 一段时间以来的所有问题和挑战：用量上效率高得多，说话方式上聊起来顺畅得多，还有就是实打实地能把事办成。 It solves basically every problem and challenge Anthropic has had for a while now, from usage, it's way more efficient, to the way it talks, it's way easier to talk to, to just being able to get things done. 10:41 他们终于回来了。Anthropic 回来了。他们差不多有五个月没跟上竞争了，这真的挺可惜的。 They finally are back. Anthropic is back. They've been out of the game for like five months now, which has been really unfortunate. 10:48 但他们回来了，而当公司回归、重新竞争的时候，最终赢的永远是我们消费者。本周和下周，我们又一次身处 AI 史上最激动人心的两周。 But they are back, and when companies come back and they compete, us, the consumer, always ends up winning. We are again in the two most exciting weeks in AI history this week and next. 11:00 记得订阅并打开通知，第一时间了解 AI 领域的所有最新动态。我会给大家带来所有演示、所有内容、所有独家消息，都会在这个频道发布。希望这期视频对你有帮助。非常感谢你收看我的内容，我们下期视频见。 Make sure to subscribe and turn on notifications to stay up-to-date on all the latest news on what is going on with AI. I'll give you all the demos, everything, all the exclusives. You'll find them here on this channel. I hope this was helpful. I'm so grateful you'd watch my content, and I'll see you in the next video.
没有匹配的文稿
69:03
11:43
14:05