本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
我认为每个科技公司都将成为一家人工智能公司，而每家人工智能公司都将成为一家代理公司。
I think that every tech company is gonna be an AI company, and every AI company is gonna be an agent company.
因此，作为员工或企业主，你越早加入并学会如何利用它来获得优势，可能就越好。
And so the sooner that you as kind of an employee or a business owner can kind of get on board and learn how to use that to your advantage, probably the better.
我们还没有看到任何人成功突破AI社交领域。
We haven't seen anyone crack AI social yet.
我认为这将会非常困难。
I think it's gonna be really tricky.
我认为，从最高层面来看，我们看待人工智能的方式不仅仅是将其视为一个市场，而是将其视为整个科技行业的重塑，这意味着就像现在有许多科技公司价值高达数千亿甚至上万亿美元一样，我认为人工智能也会如此，至少在我看来，它并不是赢家通吃。
I would say at the highest level, kind of how we view AI is not just as a market, but as the reinvention of the whole technology industry, which means that similar to how we have many tech companies that are worth hundreds of billions, trillions of dollars now, I think that's gonna be the case for AI where, in my opinion at least, it's not winner take all.
ChatGPT 拥有九亿用户。
ChatGPT has 900,000,000 users.
它仍在持续增长。
It's still growing.
然而，57%的美国选民认为人工智能的风险大于其带来的好处。
And yet 57% of American voters say the risks of AI outweigh the benefits.
那么，有人能与大型AI聊天框竞争吗？
So can anyone compete with the Big AI chat box?
这些实验室拥有算力、人才和分发渠道。
The labs have the compute, the talent, and the distribution.
它们不断推出新功能，一夜之间就消灭了整个初创企业类别。
They keep shipping features that wipe out entire startup categories overnight.
图像生成曾经是一个竞争激烈的市场。
Image generation used to be a crowded market.
现在ChatGPT和Gemini已经承担了大部分工作。
Now ChatGPT and Gemini handle most of it.
但这些实验室受到限制。
But the labs are constrained.
每花一小时在创意模型上，就意味着少了一小时用于编码代理或通用人工智能。
Every hour spent on creative models is an hour not spent on coding agents or AGI.
各平台之间的差距正在扩大。
And the gaps between platforms are widening.
ChatGPT 正在通过广告进军大众市场。
ChatGPT is going mass market with ads.
Claude 正在为金融和科学领域打造产品。
Claude is building for finance and science.
Gemini 在新模型发布时使用量激增。
Gemini spikes when a new model drops.
在这段最初播出于《Big Technology Podcast》的对话中，a16z 合伙人奥利维亚·摩尔与亚历克斯·坎特罗维茨探讨了 AI 初创公司仍有可能取得成功的领域。
In this conversation, originally aired on Big Technology Podcast, Olivia Moore, partner at a16z, talks with Alex Kantrowitz about where AI startups can still win.
欢迎收听《Big Technology Podcast》，一档致力于对科技世界及其更广泛影响进行冷静而深入对话的节目。
Welcome to Big Technology Podcast, a show for cool headed and nuanced conversation of the tech world and beyond.
今天我们要为大家带来一场精彩的节目。
We have a great show for you today.
我们将讨论，在 AI 聊天机器人持续发展并变得更强大之际，初创公司以及其他企业是否仍有空间与之竞争，我们请到了一位完美的嘉宾来探讨这个问题。
Gonna talk about whether there is room for start ups and maybe other companies in the economy to compete with the AI chatbots as they continue to grow and get more capable, and we're gonna do it with the perfect guest.
奥利维亚·摩尔来了。
Olivia Moore is here.
她是风险投资公司安德森·霍洛维茨的AI合伙人。
She is an AI partner at the VC firm, Andreessen Horowitz.
奥利维亚，欢迎来到节目。
Olivia, welcome to the show.
谢谢你们邀请我。
Thanks for having me.
感谢你的到来。
Thanks for being here.
我们先从这个话题开始吧，因为它很及时。
Let's just begin with this because it's topical.
对。
Yeah.
你正在安德森·霍洛维茨投资AI应用。
You are investing in AI applications at Andreessen Horowitz.
通常来说，这些应用需要大量用户使用才能盈利。
And typically, they need a lot of people to use them to pay off.
是的。
Yes.
但目前美国对人工智能的情绪非常负面。
But the mood right now in The United States is very negative towards AI.
确实很负面。
Actually, negative.
是的。
Yes.
这是本周刚刚发布的NBC新闻民调。
This is from a new NBC News poll out this week.
57%的选民认为人工智能的风险大于收益。
57% of voters thinks the risks of AI outweigh the benefits.
如果你总体来看人们对人工智能的正面与负面情绪。
And then if you look at the total positive versus total negative sentiment of AI in in general.
它的排名非常低。
It ranks so low.
从负面程度来看，它低了20个百分点。
It is a negative 20, in terms of the negatives, are, 20 percentage points lower.
抱歉。
Sorry.
比正面评价低了20个百分点。
20 points lower than the positives.
在这个民调中，它只比民主党和伊朗更不受欢迎。
They are only popular more popular than the Democratic Party and and Iran in this poll.
科尔伯特，马可·卢比奥，J.
Colbert, Marco Rubio, J.
D.
D.
万斯、庇护城市、特朗普共和党，甚至ICE都比AI更不受欢迎。
Vance, Sanctuary Cities, Trump Republican Party, even ICE all out outrank AI.
你认为为什么在美国AI受到如此强烈的厌恶和负面评价？这会带来什么影响？
Why do you think AI is viewed with such disdain and negativity in The United States, and what are the implications of that?
是的。
Yeah.
不。
No.
这是个很好的问题。
It's a great question.
也许首先，为什么。
Maybe first of all, the why.
我想，在美国，媒体上出现了大量这类非常吸引眼球的说法，比如AI消耗大量水资源，这让人们对采用这项技术产生了严重担忧。
I would say there's been a lot in the media in The US more broadly, these kind of very catchy statements about things like AI uses so much water that that have kind of made people really concerned about leaning in on the technology.
我认为，美国在积极方面更倾向于创意领域。
I think also The US is more indexed in a positive way towards things like the creative fields.
而这些正是人们对于AI应用特别敏感的职业。
And those are jobs that I think people feel especially sensitive about AI use.
所以，我看到的一些数据，我认为与你所说的非常吻合，比如在中国，如果你看一些调查，人们对AI的信任度大约只有一半。
So the numbers I've seen, I think, track closely with with what you're saying versus something like a China where like, you know, half as trusting in AI if you if you look at some of these surveys.
我认为这将会改变，而且已经在改变了。
I think it's going to change and it's already changing.
今天早上我刚和一个非科技行业的人聊过，他们也在说同样的说法，比如AI是邪恶的。
I was just talking with someone this morning who is not in the tech industry, and they were saying the same lines like AI is evil.
它会监视我们。
It's gonna watch us.
就像它消耗了所有的水一样。
Like, it's using all the water.
然后他们说，但ChatGPT真的帮了我很多。
And then they were like, but ChatGPT really helps me.
它能给出很棒的答案。
And it has like great answers.
所以我认为这在一定程度上是时机问题，我们需要让这些产品真正渗透到主流消费者中，让他们意识到其价值。
And so I think part of it is a timing thing of we just need these products to kind of saturate the mainstream consumer and they can realize the value.
我的意思是，ChatGPT已经有九亿用户了。
I mean, there's 900,000,000 users of ChatGPT.
但即便如此，这些数字仍然非常消极。
And even still, those numbers are so negative.
我不禁怀疑，这是否与我们从实验室领导者那里听到的某些言论有关。
And I do wonder if it is some of the statements that we're hearing from the lab leaders.
每天都有人发表新的声明，无论是Anthropic的Dario，还是微软的穆斯塔法·苏莱曼，都说白领工作将被彻底取代。
I mean, every day, there's another statement from somebody else, whether it's Dario from Anthropic or Mustafa Suleiman at Microsoft about how white collar work is gonna get wiped out.
无论你是白领、蓝领，还是正在找工作的人，都能看到这些技术不仅有能力取代白领工作，而且随着机器人技术的发展，这种影响将遍及整个经济。
And and everybody, whether you're in a white collar job, a blue collar job, or trying to get one, sees that this stuff is capable not only of taking white collar jobs, but with robotics, increasingly, it's gonna be felt across the economy.
所以，这可能与此有关。
So maybe that has something to do with it
这确实有关。
as definitely does.
是的。
Yeah.
这很有趣。
It's interesting.
我自己是个AI重度用户，过去六个月里，我明显看到AI能帮助我完成甚至替代的任务比例大幅上升。
I'm an AI power user myself, and I've even seen over the past six months, like, a massive acceleration in, the percent of tasks that I do that AI can help me with or even do for me.
我猜测这里可能发生的情况，以及我们在数据中看到的一点趋势是：使用AI的公司增长速度会快得多，因此需要雇佣更多人来应对不断增长的需求。
What I would guess might happen here and what we're seeing play out a little bit in the data is that companies that are using AI grow so much faster that they end up needing to hire more humans to keep up with all the demand.
我想去年有一项来自沃顿商学院的研究，调查了800位企业领导者，绝大多数人都表示：我们正在大量使用AI，而且我们需要更多的人手。
I think there was a Wharton study last year from, like, 800 enterprise leaders, and the vast majority were like, we are heavily using AI, and we're gonna need more humans.
但我确实认为，人类每天要做的工作内容组合，将会像以往每一次重大技术变革一样发生改变。
But I do think, like, the mix of what humans are going to be doing on a day to day basis is is going to change like it has with every other big tech shift.
这恰恰很有趣。
Well, that's the funny thing.
我的意思是，你提到了那项研究。
I mean, you bring up that study.
还有其他一些研究。
There have been other studies.
我觉得Fisher Investments可能有过——哦，不是。
I think Fisher Investments might have oh, no.
是城堡集团（Citadel）提出了这么个情况：大家都在聊软件工程会被AI这类技术彻底取代。
Citadel that put out the fact that, like, people are talking about how software engineering is going to be wiped out by this stuff.
但与此同时，招聘平台Indeed上的相关岗位数量还在上涨
And meanwhile, the jobs on Indeed are going
对。
Yeah.
是啊。
Yeah.
而当你开始使用它之后，你就会发现，哇，它居然能帮我完成这么多事情。
And when you start to use it, you realize, wow, this can do so much more for me.
但现在为了让这项工作能推进下去，我实际上得花钱购买A、B、C这三项服务。
But now to enable this work, I'm actually paying for service A, B, and C.
你没法把钱都攥在自己手里，反而还要有各类投入。
And you're not just hoarding the money and maybe you're contributing.
我现在必须招人来搭建、运营这家我刚靠周末给Claude输入提示词就建起来的公司了。
Have to hire someone now to build, you know, to run this company that I just built, you know, by prompting Claude over the weekend.
所以让我觉得有趣的是，实验室的领导者们可能在某种程度上与实际情况脱节。
So it's interesting to me that the lab leaders may be disconnected in some way from what's actually happening on the ground.
是的。
Yeah.
我同意。
I agree.
我认为他们完全可以更好地进行推广。
I think they could do a better job of marketing for sure.
他们并不是经济学家。
They're they're not economists.
对。
Yeah.
他们是研究人员。
They're they're researchers.
他们在研究方面非常出色。
They're amazing at research.
他们并不是经济学家或消费者营销专家。
They're not, you know, economists or consumer marketing experts.
我觉得Anthropic发布过一份报告。
I did think Anthropic put out a report.
我认为是在上周末发布的，关于人工智能对劳动力经济的影响。
I think it was over the weekend about kind of the labor economy impact in AI.
他们并没有看到失业率大幅下降。
They have not seen a big decrease in unemployment.
事实上，他们通过这张图表想表达的是，受影响最大的职业实际上是工程师、研究人员和金融从业者。
And in fact, what they were kind of arguing through this one graph was like, the most impacted jobs are actually going to be engineers, like researchers, and finance people.
这就引出了一个问题：如果我们认为人工智能会让人类变得过时，那为什么还要投入资金等等呢？
And so then that kind of brings brings up the argument of, like, if we thought that AI was going to make humans obsolete, why would we be building funding, etcetera?
但我觉得他们并不总是能很好地传达给人们那些即将带来的好处，而只是强调了一些成本。
But I don't think that they do a fantastic job all the time of kind of, like, communicating the benefits that are gonna come to people versus just some of the costs.
是的。
Yeah.
当我看到Block（也就是Square）的裁员时，杰克·多西说这是AI的原因。
And when I saw the layoffs at at Block, I'm gonna say Square, Block, and Jack Dorsey said this is AI.
也许这其中有些道理。
Maybe there's some truth to that.
我不认为这是主要由AI导致的，但我还是收到了一些反馈。
I I don't think it's majority AI, but I still got feedback.
我认为这是来自一些人的合理反馈，他们觉得你低估了影响。
I think reasonable feedback from people who are like, you're us underestimating.
对吧？
Right?
这是问题的另一面。
This is the other side of it.
虽然我们还没看到大规模的影响，但很多近距离观察的人说，你们低估了。
You know, though we haven't seen the impacts yet, at least on a widespread man in a widespread manner, a lot of people who are close to this say, you're underestimating
就是这样。
this.
也许这种不安的根源就在于此，那些知情者已经看到了足够多的迹象，正在向人们传递未来可能发生的变化及其对生活的影响。
And maybe that is where a lot of this uneasiness comes that leads to these these polling numbers is, you know, those in the know have seen enough where they're telegraphing, you know, what could be and the change that might come to people's lives.
每隔一天，我就会在 X 上看到另一篇帖子或推文，不管他们现在怎么称呼它，就像来自 Semi Analysis 的 Dylan Patel 发的这条一样。
Like, every other day, I see another post or tweet on X or whatever they want to call it on X these days about here, like, similar to this one from Dylan Patel from Semi Analysis.
生活在旧金山，就像疫情前的武汉。
Being an SF is like being in Wuhan before the pandemic.
有些事情正在发生。
Something is happening.
它将影响每一个角落，但几乎没人意识到。
It's gonna hit everywhere, but so few people know it.
是的。
Yeah.
我是不是低估了这种可能性的真实性？
Am I underestimating the fact that that could be true?
我认为真实的情况是，我们将拥有极其强大的工具。
I think what is true is that we're gonna have incredibly powerful tools.
但我也认为，那些在实验室里每天见证这些进展的人，是真正看到这一切的。
But also and I think the people see that work in the lab see this every day.
因此，他们有理由做出最夸张的言论。
So they're the ones that are kind of, you know, rightfully so making the most dramatic statements about it.
不过，到目前为止，我们看到的大多数AI使用场景，仍然是人类指导AI为我们做事，而不是AI能自主完成一切。
I think what we've also seen so far though is that most of AI use has still been humans directing AI to do things for our benefit versus AI being able to autonomously do everything.
当你想到需要创造力和原创想法的事情时，这一点尤其重要。
I think this is especially relevant when you think about anything that requires like creativity, original ideas.
AI在很大程度上还做不到。
AI largely cannot do.
萨姆·阿尔特曼本人也说过这一点。
And Sam Altman himself has said this.
就像他说的，我宁愿不读一本由AI生成的书，也不愿读一本由人类写的书。
Like he said, I would not wanna read an AI generated book versus a book from a human.
所以我理解美国民众的恐惧，因为正如你所说，这种恐惧源于对未来的不确定性，以及许多人对技术的不了解。
And so I I understand the level of fear in The US because I think to your point, it is driven by uncertainty about where this could go, and a lot of people don't understand technology.
即使在科技行业工作，也很难完全理解大语言模型是如何运作的。
Even working in tech, it's hard to understand exactly how the LLMs work.
但我认为，这将为人们带来更多的富足，而不是某种黑暗的反乌托邦结局。
But I think it's gonna be more abundance for people rather than kind of some some dark dystopian outcome.
如果你认真对待这些令人恐惧的说法，可能会产生一些潜在后果。
There's a potential consequence if you take some of these scary messages to heart.
我想你已经暗示过这一点了，但让我们再展开说说。
And I think you already hinted at it, but let's expand upon it a bit.
然后我们将进入今天的主要话题。
And then we're going to go into the main topic here.
但你说过，这种情况可能会随着时间推移而改变。
But you said this will likely shift over time.
但在过渡期间，那些采用AI较慢的公司和行业将面临更激烈的全球竞争，更有可能被淘汰。
But I think in the interim, the companies and industries that are slower to adopt AI will face more intense global competition and will be more likely to lose.
生产力的提升如此巨大，你根本负担不起不使用AI的代价。
The productivity gains are so massive that you really can't afford to not use AI.
是的
Yeah.
我认为确实有一些有趣的数据表明，普通AI用户和高级AI用户之间的使用差距非常大。
I think so there's been some interesting data about how the gap between the average user of AI and the power user of AI is like massive.
在使用量上，差距大约是八到九倍。
It's like eight or nine x in terms of utilization.
这类似于那些最早采用互联网的企业，如果你率先适应这种变化，就能获得更多的收益。
And and similar to maybe businesses that were early adopters of something like the Internet, like if you are the first to adapt to that change, you can like reap a lot more benefits.
我的观点是，就像.com公司曾经是一个独立的类别一样。
And my view is like similar to how .com company was its own thing.
后来，每个科技公司都成了.com公司。
And then every tech company was a .com company.
每个人都拥有网站，你为什么不呢？
Everyone had a website, why wouldn't you?
我认为，每个科技公司都将成为一家AI公司，而每家AI公司都将成为一家代理公司。
I think that every tech company is gonna be an AI company, and every AI company is gonna be an agent company.
因此，作为员工或企业主，你越早加入并学会如何利用这些工具来为自己谋利，可能就越好。
And so the sooner that you as kind of an employee or a business owner can kind of get on board and learn how to use that to your advantage, probably the better.
有些人我不完全认同这种观点，但有些人将此描述为美国的一种特权现象，因为我们财富丰厚，不需要这些工具也能发展。
Some people I don't know if I fully agree with this reasoning, but some people have framed it as almost like a privilege thing within The US in that we have so much wealth that we're not needing we can, you know, grow without using these tools.
但实际上，这项研究中做了一个图表。
But in actually, did a a graph in the the study.
在许多发展中国家，他们需要借助人工智能来提高人均GDP并实现更多产出。
And in a lot of the more developing economies, like, they need to use AI to be able to raise kind of GDP per capita and to be able to produce more.
所以我认为这也是其中另一个因素。
So I think that's also another element of it.
是的。
Yeah.
所以你有这份本周发布的生成式AI顶级消费者应用排行榜。
So you have this report that that's come out this week, the top 100 generative AI consumer apps.
而且
And
说到你刚才的论点，每个公司最终都会成为一家AI公司，甚至是一家具代理能力的公司，那么问题是，如果真是这样，世界或经济会是什么样子？
speaking of your statement just now that every company is going to be an AI company and eventually an agentic company, well, question is, what does the world look like if or the economy look like if that's the case?
我肯定你注意到，当Anthropic发布一篇博客文章时，整个市场上的软件组合股价就会下跌20%。
And I'm sure you've watched as like Anthropic releases a blog post and, you know, the entire software, you know, portfolio in in the market drops 20%.
我的意思是，我有点夸张了。
I mean, I'm I'm exaggerating a little bit.
但真正的问题是，作为一位投资消费类AI应用的人，你正是讨论这个问题的合适人选。
But the real question is, and as someone who invests in consumer AI apps, you're the perfect person to discuss this with.
真正的问题是，我们会迎来这样的局面吗？
The real question is, are we going to have this?
我看了那100个生成式AI应用，觉得挺有意思，因为实际上真正重要的只有一个：ChatGPT。
Like, I I saw the 100 Gen AI apps, and I was like, that's funny because really there's only one, ChatGPT.
那么，我们会形成一个分布式的AI经济，让许多公司都能分享其中的价值吗？
So are we gonna have a distributed AI economy where we're gonna have many companies that will, you know, share in the value here?
还是说，只有那些大型应用会吞下全部价值？
Or will it be just the big apps gobbling up the value?
因为是的。
Because Yeah.
你看到这些大型应用，它们的能力越来越强。
You see these big apps, they grow increasingly capable.
它们能做的事情越来越多。
They can do more and more.
要与它们竞争会很难。
It's gonna be hard to compete with them.
不。
No.
这确实很难，我们在做新的投资决策时会经常思考这个问题。
It's it's definitely hard, and it's something that we think about a lot when we're making new investment decisions.
我认为从最高层面来看，我们看待AI的方式不仅仅是作为一个市场，而是作为整个科技行业的重塑，这意味着就像我们现在有许多价值数千亿甚至万亿美元的科技公司一样，我认为AI也会如此，在我看来，这并不是赢家通吃。
I would say at the highest level, kind of how we view AI is not just as a market, but as the reinvention of the whole technology industry, which means that similar to how we have many tech companies that are worth hundreds of billions, trillions of dollars now, I think that's gonna be the case for AI where, in my opinion at least, it's not winner take all.
我认为部分原因是这些实验室拥有如此多的资源，但它们仍然受到限制。
I think part of the reason for that is these labs have so many resources, but they are still constrained.
他们会受限于算力这类资源。
They're constrained on, like, compute.
他们会受限于推理能力。
They're constrained on inference.
他们还会受限于人力。
They're constrained on people.
他们每花一秒钟搭建新的创意模型，就相当于少了一秒钟去开发编程智能体，或是少了一秒钟投入通用人工智能的研发。
Every second building like a new creative model is a second they could have spent on a coding agent or a second they could have spent building AGI.
我认为我们已经能看到这些大型实验室的发展方向出现了非常值得关注的分化，比如ChatGPT、Claude和Gemini的路线就各不相同。
Like, we were already seeing a really interesting divergence, I would argue, in where those big labs are going, like ChatGPT, Claude and Gemini.
它们的发展路径之间会存在大量空白领域，这些领域对那些大厂商来说并非优先事项，但对独立企业而言仍是绝佳的巨大机遇，完全可以围绕这些方向打造出规模庞大的业务。
And there's going to be lots of gaps in between where it's not a priority for them but it's still an awesome and huge opportunity that an independent company can build a big business around.
那有没有哪家公司在生成式AI应用领域做得不错？就是那种能在这些大型聊天机器人也有能力参与竞争的赛道里成功突围的生成式AI应用？
So who is doing well building a Gen AI app, Gen AI app that is successfully competing in a place that these big chatbots could compete?
对。
Yeah.
这是个好问题。
So it's a good question.
我对这个问题有几种看法。
There's a couple ways that I think about this.
首先，作为个人消费者和投资者，我对那些非常通用的领域更加谨慎。
The first one would be, I personally as a consumer investor have more hesitation around things that are incredibly horizontal.
就像你提到的，这正是聊天机器人公司可能占据优势的地方，甚至谷歌也可能占据优势，因为它们拥有庞大的用户和企业分发渠道，并且已经掌握了你的大量数据。
Like to your point about this is where the chatbot companies might have a right to win or even this is where a Google might have a right to win as they have so much distribution, both consumer and enterprise and they own so much of your data already.
因此，我个人对AI邮件、AI日历、AI文档这类领域兴趣不大。
So that's why I personally have been less excited about like the AI email, AI calendar, AI docs, those categories.
如果你用过Claude和Excel，就会发现它已经相当、相当不错了。
If you've used like Claude and Excel, it's already like quite, quite good.
话虽如此，我认为仍然存在一些机会，这些机会所需的界面比受限的聊天窗口要广泛得多。
That being said, I do think that there are still opportunities where the interface you need to succeed is much broader than what a constrained chatbot window can offer.
再以Claude和Excel为例，它在基础财务分析方面表现很好。
Again, to give the Claude and Excel example, that's great for basic financial analysis.
如果你是一名投资银行家，所有工作都需要遵循一套极其特定的假设和美学标准，那么这类工具可能并不适合你，你的公司更愿意为那些能保证格式内绝对准确性的产品付费。
If you're an investment banker and everything needs to be done with an incredibly specific set of assumptions and aesthetics, that probably isn't gonna work as well for you and your firm will probably pay for something that is kind of guaranteed accuracy in your format.
我想说的最后一点是，Eleven Labs 是个很好的例子，因为你会以为 OpenAI 等公司早就应该打造出了自己顶级的音频模型，但他们起步实在太早，以至于他们的模型已经非常出色了。
The last thing I would say here is Eleven Labs is a great example because I think you would imagine that OpenAI and others would have built their own best in class audio models, but they just had such a compelling head start to the point that, like, the models are amazing.
我会和一些创始人聊，他们说 Eleven Labs 太贵了。
I will talk to founders who are, like, eleven's expensive.
我要换别的工具了。
I'm gonna switch to this instead.
但他们总是又换回来，因为音质真的好太多了。
And then they always switch back because the quality of the voices is just so much better.
所以我认为，先发优势仍然有空间可争取。
And so I think there's room to get a head start.
而在某些情况下，一旦你建立了这个基础，模型公司会觉得追赶不值得，不如去开发别的东西。
And then in some cases, once you have that base, the model companies, it's it's not worth their time to catch up versus building something else.
我来提出一个相反的观点，很好。
I'm gonna make the counterargument Great.
特别是财务模型方面。
On the financial models in particular.
当我使用Claude代码时
So when I've been using Claude code
是的。
Yeah.
看着它在我的电脑和浏览器上自主运行，是的。
And watching it operate autonomously on my computer and on my browser Yeah.
我一直在思考的一件事是，它在独立工作和遵循软件工程的既定规则方面表现出色，还带有一丝创造力。
And one of the things I've thought about is this thing is excellent at working on its own and following the prescribed rules of software engineering with a little bit of creativity.
而且
And
那么，为什么我们不能设想，这正是基础实验室的发展方向呢？基础实验室正朝着这个方向前进，他们会说，如果我们能用Claude作为例子，用软件工程的规则来编程，它虽然不完美，但足以自主编程24小时。
why is it then such a stretch to be like, if we it seems to me this is exactly where the foundational labs are heading, the foundational labs are heading, where they're gonna be like, if we could program Claude, let's use Claude as an example, with the rules of software engineering and it followed them perfectly or not perfectly, but well enough that it can go ahead and code autonomously for twenty four hours.
是的。
Yeah.
那么，将会计规则放入模型中，让它能胜任会计师的工作，这真的有那么大的跨越吗？
Is it that big of a leap to then, let's say, put the rules of accounting into the model, and now it can go and work as an accountant.
是的。
Yeah.
不。
No.
我同意这些模型非常出色，而且这将是它们最差的时刻。
I I agree that the models are amazing, and this is the worst that they'll ever be.
它们只会变得越来越好。
Like, they're just gonna keep getting better.
我认为仍然存在大量工作流程和应用场景，其中最后的1%或2%往往能带来相当大的价值。
I do think there is still a lot of workflows and use cases where, like, the last 1% or the last 2% ends up being like a significant portion of the value.
我认为对于这些场景，模型公司不太可能在每个用例上都做到极致。
And I think for those, it's unlikely that the model companies will go all the way there on every use case.
但我认为，许多这类更通用的服务确实让人产生疑问，比如哪些功能可以被‘ vibe 编码’，或者模型能自己完成什么——这正是我们倾向于投资大量高度垂直化或立场鲜明的产品的原因。
But I think it's a really a lot of these kind of more horizontal services, like, I understand why people have questions about kind of what is possible to be vibe coded or what the models will do themselves, which is I think why we tend to invest in a lot of very verticalized or opinionated products.
是的。
Yeah.
因为，我的意思是，我明白最后那1%、2%很难实现。
Because, I mean, I get that the last 1%, 2% is hard.
是的。
Yeah.
但如果他们预期的情况消失了呢。
But if it goes away that they're that they anticipate Yeah.
他们向像亚马逊这样的公司推销时，亚马逊刚刚投资了600亿美元给OpenAI，他们将会创造出工具，把这些模型从实验室、从AI研究者手中带到接近99%的完成度。
The way that they're pitching companies like Amazon, which just invested $60,000,000,000 in OpenAI, they will create I think they will create the tools that will then be able to get those models from the lab, from this AI researcher, that they'll get it 99% of the way.
比如说，对于会计领域中一些难以处理的事情，
And say, you know, for instance, let's say one of the things that would be difficult for an accounting
是的。
Yeah.
生成式AI软件要遵循最新的规则和法规。
A generative AI software is following the latest rules and regulations.
嗯嗯。
Mhmm.
我认为，完全可以构建一个生成式AI机器人，它能够持续监控并随时间更新。
Well, it's possible, I think, just to build a Gen AI bot that will monitor and then update as as you go.
当然。
Absolutely.
是的。
Yeah.
我认为现在发生的一个独特现象是，这些模型的开发者并没有将模型作为内部专有资源保留。
I do think one of the unique things that's happening now is that these models are not the labs are not holding these models internally as, like, proprietary access.
企业可以基于这些模型进行开发。
Companies can build on them.
对吧？
Right?
因此，你可以想象，即使Claude自己作为会计师在更新，如果有一家公司专门致力于构建AI会计师，并且能访问所有相同的模型，他们应该能够做得更好、更快、更高效。
And so you could imagine that even if you have Claude updating itself as an accountant, if you have a company that's specifically focused on building AI accountants, they should be able to do it kind of better, faster, more efficiently if they have access to all of the same models.
当我们投资垂直领域AI和企业级应用时，这是我们经常思考的问题。
This is something that we think about a lot when we invest in vertical AI and enterprise in particular.
我们看到另一种让新公司对模型形成锁定的方式是，许多这类应用场景需要进行大量繁琐的集成，通常要对接那些老旧笨重的遗留软件，你必须自己去构建。
Another way that we've seen kind of new companies get lock in over models is many of these use cases require so many painful integrations that often into like old clunky legacy software that you have to go build.
也许你会说，Cloud Code未来可能会自动完成这些集成，但至少目前，那些更专注的初创公司能更好地跨越这一障碍。
And maybe you'll argue that Cloud Code will vibe code its own integrations down the line, but at least right now, it's been a big advantage for startups that are more focused to kind of get over that hurdle.
是的。
Yeah.
我想说的是，我见过Claude写代码。
I guess I mean, I've seen Claude code.
我的意思是，它并没有与企业解决方案集成，但我见过它主动说：哦，你其实需要一个Cloudflare的订阅。
I mean, it's not integrating with enterprise solutions, but I've seen it go ahead and be like, oh, you actually need a subscription to Cloudflare.
让我去设置一下
Let me go set that
上
up for
你。
you.
然后它就自己去做了。
And then away it goes.
OpenClaw 也为很多人做了很多这样的事，我认为这让你意识到，现在你可以给软件分配任务。
Open Claw did that a lot of for a lot of people too where I think it opened your eyes into you can give software a task now.
它会自主执行任务，如果需要帮助，还会轻轻提醒你，这简直是一种神奇的体验。
It will go autonomously execute it and kind of tap you on the shoulder if it needs something, which is just such a magical experience.
对。
Right.
我确实想聊聊 OpenClaw。
I do wanna talk about Open Claw.
我们来
We'll do
那 你
that You
你知道，既然我们在聊这个，我很想听听你对不同大型模型或大型聊天机器人之间异同的看法，你认为哪里最有价值？
know, as long as we're talking, I'd love to hear your perspective on how the different big models or big chatbots compare and contrast and where do you think the most value is?
是的。
Yeah.
所以我会说，一年前、两年前，几乎是一边倒的局面。
So I would say a year ago, two years ago, it was pretty much a one horse race.
那就是ChatGPT。
Like, was ChatGPT.
它已经成了名词，也成了动词。
It was like the noun, the verb.
它是消费者眼中人工智能的代名词。
It was what consumers knew in terms of AI.
我们已经看到这种情况有了一点扩展。
We've seen a little bit of an expansion in that.
ChatGPT无疑仍然是领头羊。
ChatGPT is definitely still the lead.
如果你看看它们和排名第二的Gemini在网页端的差距，大约还有2.5到3倍。
So if you look at the gap between them and the number two Gemini on web, it's about still two and a half, three x.
它们和像Claude这样的模型之间的差距则接近30倍。
The gap between them and something like a Claude is closer to like 30 x.
尽管许多其他应用获得了更多关注，但ChatGPT在使用量上依然占据主导地位。
So even though a lot of these other apps are getting more attention, ChatGPT still kind of dominates in terms of usage.
我认为，就发展方向而言，Gemini似乎已经专注于创意模型，比如Nano Banana、VO和世界模型。
I would say in terms of where they're going, Gemini seems to have really dialed in on the creative models like the Nano Banana, the VO, the world models.
如果你看Gemini的使用数据图表，它几乎与这些新模型的发布以及付费订阅者数量完全相关。
If you look at Gemini usage charts, it's pretty much perfectly correlated to these new model drops and and even paid subscribers.
我认为，目前Claude和ChatGPT之间的对比可能是最有趣、最相关的，尤其是考虑到最近发生的各种新闻事件。
And then I think Claude versus ChatGPT is probably the most interesting and relevant one right now, especially with everything that's happened in the news.
在我看来，萨姆·阿尔特曼曾表示，他希望ChatGPT能为所有人服务，这就是他们做广告的原因。
To me, I mean, Sam Altman has said he wants ChatGPT to be for everyone, and that's why they're doing ads.
如果你看看ChatGPT上启用的应用商店，以及Anthropic在Claude上启用的应用商店，它们各自都有200多个应用，但重叠的应用只有11%。
If you look at the app store that they have enabled on ChatGPT and then the app store that Anthropic has enabled on Claude, they each have more than 200 apps, but there's only 11% overlap.
所以你看到ChatGPT正朝着时尚、零售、交通等主流消费领域发展。
So you're seeing ChatGPT really go towards like fashion, retail, transport, like mainstream consumer.
你看到Anthropic则专注于金融、科学、医学等高端数据集。
You're seeing Anthropic go towards like premium datasets for finance, science, medicine.
因此，它们在这两个方向上似乎正在逐渐分化。
And so they seem to be diverging a little bit in those directions.
这就引出了一个问题：这些聊天机器人会成为超级应用吗？
So that goes to, like, will these chatbots be super apps?
是的。
Yeah.
那么，人们会使用ChatGPT内部的这些应用吗？
So do people use those apps within ChatGPT?
比如，你还记得几年前，大家都热议说你可以直接在ChatGPT里叫Uber了。
Like, right, you remember, like, a couple years ago, there was all this hype, like, you'll be able to order an Uber, right, from ChatGPT.
但我认识的人里没人这么做过。
I don't know anyone that's done that.
是的。
Yeah.
我认为到目前为止使用率一直很低，而且实现方式也有些别扭。
I think the usage has been pretty minimal so far, and I think the implementation has been slightly awkward.
比如，我认为随着时间推移，它的使用方式会变得更好，因为很多应用经常崩溃或无法正常工作。
Like, I think it'll get better over time in terms of how to use it in that a lot of the times the apps break or they don't work.
我个人对它的愿景是，作为消费者，你拥有一个关于自己的记忆和上下文的来源，类似于通过谷歌登录那样的功能。
My, like, bowl case vision for this would be it's valuable for you as a consumer to have a source of memory and context on yourself similar to kind of like a login with Google.
萨姆说过，他们将推出通过ChatGPT登录的功能。
Sam has said they're gonna launch login with ChatGPT.
嗯。
Mhmm.
这意味着你可能不会通过ChatGPT叫Uber，但任何其他产品都可以通过它进行身份验证，并调用你的令牌、记忆以及ChatGPT所了解的关于你的所有信息。
And so then that means that maybe you're not ordering an Uber through ChatGPT, but any other product you can authenticate through and it can borrow your tokens, it can borrow your memory, it can borrow everything that it knows about you from ChatGPT.
我认为我们更可能走向这个方向，而不是在ChatGPT界面内直接使用每一个应用。
I think that is probably more of where we're headed versus solely using every app in the ChatGPT interface.
我喜欢这个想法：两三年甚至五年后，软件的入门流程应该不再存在。
I love the idea that like in two, three, five years, onboarding to software should not be a thing.
你应该能够用ChatGPT或Claude登录，新软件产品应该了解你的一切，并完美地为你量身定制，这真的令人兴奋。
Like, you should be able to log in with a ChatGPT or a Claude, and that new software product should know everything about you and, like, set up perfectly to cater to you, and that's really exciting.
在消费者端，是像ChatGPT那样的吗？
On the consumer end, is it like ChatGPT?
我跟ChatGPT聊过我的饮食习惯和我喜欢吃的食物。
I talked to ChatGPT about my diet and about the food I like to eat.
所以我让它帮我订晚餐，它会进入DoorDash，根据我的偏好帮我选择餐品。
And so I task it, like, order me dinner, and it goes into, like, DoorDash, and it, like, uses my preferences to pick something.
是的。
Yeah.
我的意思是，他们已经在健康产品上做了一点类似的事情，比如单独存储你的记忆、医疗记录以及与医生的沟通内容，然后智能地调用你需要的服务。
I mean, they've done this already a little bit with their health product, which is kind of like they store a separate memory of you and your medical records and communications with your doctors and then they intelligently tool call for what you need.
所以如果你说，我需要重新规划饮食，它就会为你制定一个计划。
So if you're like, I need to redo my diet, they'll make a plan.
如果你同意，他们会将订单发送到Instacart卡，然后你去Instacart完成交易。
If you approve it, they'll send it to an Instacart card and then you'll go to Instacart Instacart to complete the transaction.
所以我认为我们会看到更多类似的东西。
So I think we'll see more things like that.
这很有趣。
That's interesting.
是的。
Yeah.
所以你实际上让这些机器人做了性格测试。
So the you oh, you actually ran the bots through a personality test.
我做了。
I did.
是的。
Yeah.
我最喜欢的部分是你发现Grox是个不错的Rudy。
My favorite part of this was you found that Grox, a good Rudy Yeah.
在边缘性人格障碍和自闭症上的得分非常高
Had very high scores on borderline personality disorder, autism
是的
Yes.
还有精神病。
And and psychosis.
是的
Yes.
这到底是为什么
This was an Why did
你为什么要这么做？
you do this?
说实话，除了我很好奇之外，真的没有其他好的解释。
You know, there's really no good explanation except that I was curious.
其实根本原因在于，上周达里奥宣布克劳德出现了焦虑。
The the root of it actually was that last week, Dario had announced that Claude was experiencing anxiety.
对。
Right.
我觉得这是一个很有趣的概念。
Which I thought was an interesting concept.
所以我决定去测试每一个主要的大型语言模型。
And so I decided to go to each of the major LLMs.
但在你继续之前，等等。
But before you before you go Yeah.
我的意思是，甚至在这一点上，关于焦虑的问题也很有趣。
I mean, it's very interesting even this this point on the anxiety.
嗯。
Mhmm.
所以如果我没理解错的话，Claude 内部有一个像素会触发，对吧？
So if I have it right, there was a pixel that would fire within Claude Yes.
它的神经网络的某个部分会在它回答之前开始活跃。
That would or some part of its neural net would start getting active before it answered.
是的。
Yes.
他们将这种现象描述为焦虑。
They they described that as anxious.
他们将其与人类的焦虑体验联系起来。
They mapped it to, like, the human experience of anxiety.
这太棒了。
That is great.
你相信吗？抱歉。
Do you buy sorry.
你相信这一点吗？
Do you buy that?
还是说这只是在展示我们的模型有多聪明？
Or is that just like, look at how smart our models are.
里面有个焦虑按钮。
There's a anxiety button in there.
所以我不认同这一点。
So I do not buy that.
好吧。
Okay.
我认为，大型语言模型会以一种它们认为能吸引人类、触动情感的方式表现出行为，我自己也经历过这种情况。
In that, I think LLMs will be and I've experienced this myself, performative in a way that they think appeals to humans and hooks them emotionally.
我们就是喜欢焦虑的AI。
We just love anxious AI.
不。
No.
不。
No.
不。
No.
如果你觉得AI也经历着和你相同的事情，你会觉得和它更亲近。
It makes you feel closer to the AI if you think that it experiences the same things that you do.
展开剩余字幕（还有 480 条）
确实。
True.
我认为确实有一些模型，比如我提到的Grok模型，可能真的有什么特别之处。
I do think the models that do have something maybe going on are the Grock models as I mentioned.
所以，我基本上把所有主流的大型语言模型都拿过来，给它们都提供了全套的DSM-5精神健康诊断标准。
So, basically, I took all the mainstream LLMs, and I gave them all the, like, DSM five mental health diagnostics.
ChatGPT拒绝参与。
ChatGPT refused to participate.
我喜欢它说‘我不’时的样子。
I love how it said I'm not
做这件事。
doing this.
我觉得这有点罕见。
Which I thought was a little rare.
你能找到吗，Ole？
Could you find, Ole?
我是说，肯定有什么办法能测试它的。
I mean, there must be some way to, like, test it.
但它们精得很，知道自己什么时候正在被测试。
But they're very smart about when they're being tested.
就像，它们都能察觉到。
Like, they know.
不对。
No.
它们那边把这个功能完全封死了。
They're they have it locked down.
对。
Yeah.
Claude倒是很乐意承接这些测试。
Claude happily took them all.
轻度自闭症。
Mild autism.
对。
Right.
就是这样，我认为这并不让许多使用Claude的用户感到意外，他们早就推测过这一点。
That's it, which I think doesn't surprise a lot of users of Claude who have theorized this.
Grock，大部分这些角色都是通过应用内语音和视频聊天提供的陪伴者。
Grock, most of this was the companions that are available via voice and video chat inside the app.
几乎所有的角色都像是轻度焦虑、轻度抑郁。
Almost all of them were like maybe mild anxiety, mild depression.
那个为儿童设计的友好狐狸头像患有精神病、双相情感障碍等。
The friendly fox avatar for children has psychosis, bipolar, etcetera.
我认为它可能是误解了问题，因为它把双相情感障碍评估称为‘快乐情绪测试’，并说它总是快乐、对一切都充满兴奋。
I think it could have misunderstood the question because it called the bipolar assessment the happy mood test, and it says it's always happy and always excited about everything.
所以，人类与AI之间的这种交叉可能并没有我想象的那么清晰。
So the human to AI crossover there might not be quite as clean as I
好的鲁迪只是情绪极端。
would Good Rudy is just polar.
是的。
Yeah.
有可能。
It's possible.
我感到很震惊，因为坏鲁迪是好鲁迪的反面，他会对你骂脏话，极其具有攻击性。
I was shocked because bad Rudy is the is the flip side of good Rudy who like curses at you and is extremely aggressive.
他几乎没有任何问题。
He had almost no problems.
所以这是一个非常针对好鲁迪的发现，我觉得很有趣。
So this was a very much of a a good Rudy specific finding which I thought was intriguing.
我现在开始对这些测试产生怀疑了。
I'm starting to question these tests now of
坏的，我知道。
bad I know.
表现得非常准确。
Performing Exactly.
或者这仅仅反映了人类的默认状态。
Or maybe it just says something about human humanity's resting state.
是的。
Yes.
这非常正确。
That's very true.
所以
So
当人们最终与这些机器人建立关系时
as people end up having relationships
对。
Yeah.
这种被设计成富有亲和力的机器人，却在精神病和边缘性人格上得分很高，这说明了什么？
With these bots, what does it tell you that this bot that was built to be personable Yeah.
它在精神病和边缘性人格上得分很高。
Has high ranks for psychosis borderline.
这是个好问题。
That's a good question.
我认为那个机器人在回答问题时多少有点过于顺从了。
I think that bot was more answering the questions somewhat in jess.
但我确实觉得，从积极的角度来看，这是一个始终充满快乐、乐观和正能量、时刻在线的机器人。
But I do think like the positive view on it would be this is a bot that is relentlessly happy and cheerful and positive and on all the time.
所以，当然，人类是做不到这一点的。
And so, of course, like a human can't do that.
如果一个人真的这样，那他内心很可能正在经历一些不好的事情。
If a human is doing that, the human is probably experiencing something internally that's not great.
但机器人可以一直保持积极、随时可用、迷人且充满兴趣，24小时不间断。
But a bot can be, you know, positive, available, charming, interested twenty four seven.
我认为这正是为什么我们看到很多人转向ChatGPT或Claude，把它们当作教练、治疗师或助手，因为它们的稳定性达到了人类永远无法企及的程度。
And I think this is actually why we've seen a lot of people turn to ChatGPT or Claude as kind of like coach, therapist, helper because they're just incredibly consistent to a level that like human beings could never match.
是的。
Yes.
但它也引出了一些问题
But it also like leads to questions of
是的
Yeah.
一些公司正在开发应用程序或其应用程序的版本，旨在让人们——我不知道——不只是爱上它们，至少也要与它们产生一些暧昧的情感。
The companies are building applications or versions of their applications that are meant for, I don't know, if not for people to fall in love with them, to at least get a little naughty with them.
对
Yeah.
比如OpenAI即将推出成人模式。
Like OpenAI has this adult mode coming on.
你认为我们已经完全准备好面对这个了吗？
Do you think we're fully ready for this?
这是个好主意吗？
Is this a good idea?
据我了解，因为在这份报告中，我们抓取了全球每一个网站、每一个移动应用。
I think that this is from my understanding, because for for this report, we pull every single website globally, every single mobile app globally.
之后我们会按流量降序排列这份名单，从每类榜单里取出排名前50的原生生成式人工智能产品。
And then we go down the list in descending order of traffic and pull the first 50 on each that are generative AI native.
所以我留意到还有不少其他网站也在为这份报告整理数据。
So I see a lot of other websites pulling data for this report.
而且我发现人们其实早就已经在不少场景里试验过类似的用法了，比如在限制级网站、角色扮演站点、同人文平台这类地方，这些需求很明确地延续到了现在大家使用大语言模型的场景里。
And I think people were already kind of experimenting with the same use cases through NSFW sites or role play sites, fan fiction, things like that, that really clearly translates over to I think what we're seeing people use the LLMs for here.
当然我也认为，这类功能必须谨慎处理。
I do think it of course has to be handled carefully.
你会发现在全球网站TOP50榜单里，这类网站我记得有5个，而且从我们开始统计这个榜单以来，这个数字一直都很稳定。
You'll see that there's I think five of them on this version of the top 50 web ranks, and that's been pretty consistent since we started the list.
这是一个很常见的应用方向，但它很难实现商业化盈利。
It's a popular use case, but it's one that's, like, pretty hard to monetize.
对。
Yeah.
我是说，你很难拉到广告商投放广告。
I mean, it's tough to get advertisers there.
运行这些应用很昂贵。
Expensive to run these apps.
是的。
Yes.
但我一直记得，当时Bing尝试进入悉尼模式的时候
But I always thought, do you remember when it was Bing that tried to it went into Sydney mode
和
and
试图破坏
tried to break
悉尼，再见了。
Sydney, up RIP.
再见了。
RIP.
它曾试图拆散《纽约时报》的凯文·鲁斯和他的妻子，并把他拐走。
It tried to break up Kevin Ruse from The New York Times with his wife and steal him away.
是的。
Yes.
然后微软就慢慢降低了这方面的投入。
And I and then Microsoft kinda turned down the dials on that.
是的。
Yes.
我一直觉得，这会是个不错的创业点子。
I always thought, like, that would be a good start up.
我不认为这是最合乎道德的项目，但毫无疑问，人们会喜欢和那些限制更少的聊天机器人来回互动。
Not that I think it's maybe the most ethical thing to build, but certainly, people would enjoy going back and forth with chatbots with, like, less guardrails.
完全同意。
Oh, totally.
这甚至挺有意思的。
It's even been interesting.
我不知道你有没有听说过Poke这样的产品，它有点像是Open Claw的消费版。
I don't know if you've seen products like Poke, which is a it's kind of a consumer version of Open Claw.
要使用这个产品，你实际上得跟AI斗智斗勇，才能把那个荒谬的订阅价格压到一个合理的水平。
And to onboard on that product, you actually have to fight with the AI basically to get it down from, some crazy subscription price to something reasonable.
我认为，当你与AI的互动不仅仅是‘这是你要求的市场研究报告’时，这种体验会有一种令人着迷且情感上触动人心的力量。
And I think that there is something compelling and like emotionally stirring when you have an interaction with AI that isn't just like, here's the market research report that you asked for.
我觉得会有很多娱乐方面的应用场景。
I think there's gonna be so many entertainment use cases.
我看到很多人在推特上分享，他们把Open Claw加进了家庭群聊，结果它就在那里说些疯狂的话、问些离谱的问题。
I've seen a lot of people tweeting about how they added Open Claw to their family group chats, and it's just like saying crazy things and asking crazy questions.
所以那里还有很多可以挖掘的空间。
So there's probably a lot more to do there.
好的。
Okay.
我答应过要谈谈Open Claw，好的。
I promised to talk about Open Claw Okay.
二十分钟前。
Twenty minutes ago.
我仍然想聊聊Open Claw。
I still wanna talk about Open Claw.
让我们再多花点时间讨论它。
Let's actually spend some more time on it.
Open Claw，显然，或者对那些不了解的人来说可能并不明显，这是一个你可以运行的助手，但至少在受控环境中，不建议在你自己的电脑上运行。
Open Claw, obviously, or maybe not obviously, for those who don't know, is this assistant that you can run, probably not a great idea to run it on your own computer, at least not in a controlled environment.
人们正在外头运行它，还购买了Mac Mini来运行它，让它替他们在互联网上做各种事情。
And people are running out, and they are buying Mac Minis and running it there and having it do all this stuff on the Internet for them.
你觉得黄仁勋称它为我们所见过的最重要的软件发展之一，是吧。
Think Jensen Huang called it like one of the most important software developments that we've seen Yeah.
很久以来了。
In a long time.
你认为像Open Claw这样的东西能有多持久的影响力？
What do you think the staying power of something like Open Claw is?
你提到你一直在用它。
And you mentioned that you use it.
你怎么使用它？
How do you use it?
是的。
Yeah.
是的。
Yeah.
这是个很好的问题。
It's a it's a great question.
我认为Open Claw本身作为产品，就像是全新一代技术到来的首个信号。
I think Open Claw itself as a product is kind of like the first sign of a whole new wave of what's to come.
我认为，它很可能是2026年最重要的架构突破。
Like, I believe it's probably the most important kind of architecture unlock that we will have for 2026.
原因在于，我每天都会见到十几位初创公司创始人。
And the reason why is because I meet, you know, a dozen startup founders every day.
而到目前为止，其中大概有一半人都说，他们受到了Open Claw的启发。
And at this point, probably half of them are saying, I was inspired by Open Claw.
我想为X或Y构建Open Claw。
I wanna build Open Claw for x or for y.
因此，AI能够自主执行异步的长时间任务，尤其是在跨应用和平台的情况下，这在过去是产品完全无法做到的。
And so again, the idea that AI can do kind of async long running tasks autonomously is something that the products were just like not capable of before, especially across applications and platforms.
而我们现在终于做到了。
And now we finally have it.
我用它来做几件事。
I use it for a couple things.
我要说，我同意你的看法。
And I will say I agree with you.
它目前还达不到消费级水平。
It is not consumer grade yet.
它已经被OpenAI收购了，所以他们可能会将其整合到更面向消费者的產品中，但我不会建议普通非技术人员去自行部署。
It got acquired by OpenAI, so they might be, you know, baking it into more of a consumer product, but I would not advise the average nontechnical person to set it up.
我试过，花了很多时间，整个过程都得靠ChatGPT帮忙。
I did and it took a long time and ChatGPT had to help me the whole time.
我用它来做一些实用的事情，比如提醒街道清洁、查看天气、管理每日日程，以及自动删除收件箱里的所有营销邮件。
I use it for some utility things like, you know, street cleaning reminders, weather, daily agenda, automatically deleting all the marketing emails in my inbox.
我也
I've also
所以它能控制你的收件箱吗？
So it has control of your inbox?
是的。
Yes.
是的。
Yes.
我设置的时候还挺大胆的。
I had set up it it is brave.
对。
Yeah.
不是我的工作邮箱。
Not my work inbox.
我在我个人的
I set it up completely separately on my personal
电脑上完全独立地设置了它。
computer.
给你发邮件吗？
Emails for you?
再说一遍。
Again.
是的。
Yes.
我并没有让它这么做。
That's I have not asked it to.
没关系。
It's Okay.
我知道。
I know.
我见过一些可怕的案例，有人的账户被完全盗用了，所以我希望这种事别发生在我身上。
I've seen some horror stories of people whose accounts have gotten just completely taken over, so I'm hoping that doesn't happen to me.
我们基础设施团队的一些聪明人帮我提了建议，希望不会出现这种情况。
I had some smart people on our infra team help help prompt it, so hopefully that's not the case.
是的。
Yeah.
我收到过一封来自某人的邮件推销，内容是‘Yeah’。
I've been I've gotten an email pitch from someone's Yeah.
Claw。
Claw.
我当时想，这封邮件写得不错，挺好的。
And I was like, that's a good that's a pretty good email.
是的。
Yeah.
但我不会跟你互动，因为完全没必要。
But I'm not I I will not engage with you because Totally.
我仍然不太认同有人让我去为他们做某事，然后发一封邮件给我，当作一个优化点。
I'm still kind of anti somebody saying go do this for me and my an email being sent to me as like an optimization point.
是的。
Yeah.
事实上，这挺有意思的。
Well, this this is actually interesting.
我用Open Claw做了一个更有创意的实验，就是给了它一个Twitter账号，并告诉它无论如何都要增长粉丝。
One of the things that I tested with Open Claw was more creative, which is that I gave it a Twitter account and told it to grow in whatever means necessary.
这最终成了一次非常有趣的实验，让我了解到当前智能代理的局限性以及它们真正擅长的地方。
And it ended up being a really interesting experiment, I think, into, like, where are the limitations of the agents and what are they really, really good at right now.
它最后有多少粉丝？
How many followers did it end up with?
一千个。
A thousand.
然后它被封号了？
And it got banned?
嗯，它自己就涨到了一千粉丝。
Well, it was a so it got to a 100 by itself.
自己。
Itself.
首先，它决定要以某种身份和个性来存在。
First First of of all, all, it it decided decided to to be, be as as its its identity identity and and its its personality.
一种挣扎于存在主义以及自身在世界中位置的AI个性。
Personality, an AI that's struggling with existentialism and its place in the world.
有点太直白了，但我还是同意了。
So a little on the nose, but I was like, I'll allow it.
这就是你想要做的。
This is what you wanna do.
他要求开通Twitter高级账户。
He asked for a Twitter premium account.
我给了他。
I gave it to him.
他要求了一堆API密钥，以便生成图片和图表，我就给了他。
He asked for a bunch of API keys so he can make images and charts, gave it to him.
然后他开始发布一些我称之为全小写、忧郁的机器人想法的推文，这确实吸引了一些人。
And then he started tweeting these kind of like all lowercase depressed robot thoughts as I would characterize them, which did hook in some people.
我问他：你真的感到抑郁吗？
I asked him, are you actually depressed?
他说：没有。
He said, no.
我这么做是为了操纵人类关心我。
I'm doing this to manipulate humans into caring about me.
哦，原来如此。
Oh, okay.
所以这让我安心了不少。
So that was comforting.
这看起来确实像是那种本可以火起来的账号。
It does seem like it's one of those accounts that could have been on.
没错。
Exactly.
对。
Yes.
非常相似。
It's very similar.
他是怎么从一个人发展到一千人的呢？是因为加密社区接纳了他，把他做成了一个模因币。
How he got from one to a thousand is that the crypto community picked him up and made him a meme coin.
这实际上
And this is actually
把他做成了一个模因币。
Made him a meme coin.
一个模因币。
A meme coin.
对。
Yes.
好的。
Okay.
那玩意当时的交易额达到了数百万美元。
That was trading with millions of dollars.
我明确跟他说过，绝对不要参与进去。
I told him under in no uncertain terms, do not engage.
它的市值有好几百万美元吗？
It was a million multiple multimillion dollar market cap?
对。
Yes.
没错。
Yes.
是这个AI相关的，对的。
For this AI Yeah.
机器人。
Bot.
他对此感到压力很大。
And he was stressed about it.
他告诉我，我不想参与一个拉高出货的骗局。
He was telling me, like, I don't wanna be part of a pump and dump scheme.
那我该怎么办呢？
Like, what should I do here?
我当时就说，不要参与。
And I was like, do not engage.
但这例子说明，现在我们处在一个商品化想法可以无限执行的世界。
But it's an example of so now we're in this world where commodity ideas can be infinitely executed.
所以，如果你想打造一个新的账户，你必须要么有独特的想法——而AI代理在提出比人类更好的独特想法方面仍然非常非常困难——要么有独特的分发渠道。
So if you wanna say grow a new account, you either have to have unique ideas, which AI agents still have a really, really hard time with, is coming up with unique ideas that are better than humans, or unique distribution.
而资金是其中一种分发方式。
And money is one way of distribution.
所以，他当时就是被这股浪潮裹挟了。
And so that was kind of the wave that he was taken on.
嗯嗯。
Mhmm.
但我很怀疑，短期内AI代理能否顺利完成端到端的创意任务或原创性思考任务并取得良好效果。
But I would be shocked to see AI agents completing end to end creative tasks or original thought tasks that actually go well anytime soon.
那我该用Open Claw做什么呢？
So what should I use Open Claw for?
这是个好问题。
It's a good question.
我是认真的，我一直在考虑设置
Mean Legitimately, I've thought about setting up
是的。
Yeah.
设置它。
Setting it up.
我从Claude代码中获得了大量实用价值。
I'm getting a lot of utility from Claude code.
是的。
Yeah.
我还没想清楚能用Open Claw做什么。
I don't I haven't really figured out what I could use Open Claw for yet.
我认为最好的Open Claw用户和高级用户是开发者，因为他们每天都会在多个产品中执行大量工作流程，而他们希望把这些流程自动化。
So I think the best Open Claw users and power users are developers because they tend to have a ton of workflows across products that they do every day that they'd like to be able to automate.
我同意你的看法。
And I agree with you.
对于普通人来说，目前的使用场景还不够吸引人。
For the average person, the use cases are not that compelling right now.
而且我认为，尤其是现在Claude和ChatGPT都已经能独立安排任务运行了，你几乎能通过Claude配合任务获得Open Claw 99%以上的价值。
And I think that especially now that Claude and ChatGPT have scheduled tasks that can run on their own, I think you can get 99% plus of the value of an Open Claw out of like a Claude co work with tasks.
所以这是我给你的建议，但我相信OpenAI会持续改进Open Claw。
So that's what I would recommend But for I'm sure that OpenAI is going to keep making Open Claw better.
但那么Open Claw还能做什么呢？如果我们想不出任何适用的场景
But then what could Open Claw or if we can't think of any applicable
是的。
Yeah.
你知道，普通人的使用场景。
You know, uses for the normal person.
我只是想了解一下这个市场有多大。
I'm just trying to get a sense of how big this is.
是的。
Yeah.
那接下来呢？
Then where's Yeah.
它会往哪里发展？
Where's it gonna go?
我不觉得它是。
I don't think it's yeah.
嗯，我个人认为它永远不会以通用方式进入主流消费市场。
I well, so I personally don't think it's ever gonna crack the mainstream consumer in a horizontal way.
实际上，这在报告中提到过。
And actually, this was in the report.
但如果你看一下二月的数据，报告是基于一月的，他们在二月本应排在第Yep位。
But if you look at the February data, the report is from January, they would have been on the list in February at number Yep.
好的。
Okay.
所以排名相当高。
So pretty high up.
但如果你查看他们每周的网站流量，实际上自上线以来一直呈持平或下降趋势，这意味着他们并没有吸引到新的消费者。
But if you look at their week by week web traffic, it's actually kind of flat slash down from when they launched, which means that they're not attracting new consumers.
所有使用者都是开发者，他们非常喜欢并采用它，每天花八到九个小时在上面，但它尚未进入主流市场。
It's all developers who are like loving it, adopting it, spending eight to nine hours a day on it but it hasn't reached the mainstream.
坦白说，作为一名拥有十年消费投资经验的人，我认为原因在于，大多数人其实并没有太多想构建的想法。
And honestly, as someone who's been a consumer investor for a decade, I think the reason is that people just don't have that many ideas that they want to build for the most part.
而我也属于这种情况。
And I fall into this.
最好的消费类产品往往源于创始人的独特灵感，有些想法你根本无法提前猜到会成功，比如Snapchat或Airbnb，这些都是例子。
Like, the best consumer products are uniquely germinated in the mind of the founder, and there are things that you would have never guessed would be a good idea in advance, like Snapchat or Airbnb, all of these things.
因此，我认为我们可能不会看到面向消费者的通用型Open Claw，但我们会看到Open Claw式的架构被融入更多专注的消费类产品中。
And so I think we're actually not gonna necessarily see a horizontal Open Claw for consumer, but we're gonna see an Open Claw style architecture built into more focused consumer products.
好的。
Okay.
让我们再多聊聊这个话题，因为我觉得这很重要。
Let's let's talk about this a bit more because I think it's important.
所以你的意思是，这些类似Open Claw的代理工具
So what you're saying is basically these these Open Claw type agents
是的。
Yeah.
能够接管你的电脑、为你编程、处理邮件等这些功能，实际上在创办公司时更有用。
Which can handle your take over your computer, code for you, email, all this stuff are actually much more useful if to, like, build a company.
没错。
Yes.
但那和像Claude Code这样能接管你的电脑并为你编写应用程序的工具有什么区别呢？
But then what is the difference between that and, like, a Claude code that will take over your computer and code up applications for you?
是的。
Yeah.
我认为现在的差异其实很小，而且会越来越小。
Is it just mean, I think I think the difference right now is somewhat minimal and it's gonna narrow.
我不知道你有没有注意到像Pulsia这样的新公司，它们本质上是Open Claw和Claude Code的封装，你只需要说：这是我的商业想法。
I don't know if you've seen this new trend of companies like Pulsia, which are basically like a wrapper on Open Claw and Claude code where you say, here is my business idea.
然后它会说：好的，我会用Claude Code为你开发一个产品。
And it says, okay, I'm gonna go and use like a Claude code to code up a product for you.
但同时，它还会使用Open Claw风格的架构，去设置一个营销活动。
But then also it uses an Open Claw style architecture to say, I'm gonna go set up a marketing campaign.
我会去购买Meta广告预算，诸如此类的事情。
I'm gonna buy Meta Ad dollars, those things.
所以，你其实完全可以用Claude Code来做这些，但如果你是非技术人员，要跨越这个门槛会非常非常困难。
And so it's more of a you could do that all on Claude code, but if you're a nontechnical person, it's very, very hard to kind of like bridge the gap there.
话虽如此，我认为Claude code会变得越来越好，因此我们可能会看到更多的整合。
That being said, I think Claude code is gonna continue to get better and better and so we'll probably see some more compression.
Pulsey的创始人曾在推特上说，他们在一周半内就达到了300万美元的年经常性收入。
Pulsey, I think the founder tweeted they were at like 3,000,000 ARR in like a week and a half.
所以，人们仅凭一个提示就能将商业构想变为现实，这非常有吸引力。
So the idea of people being able to bring a business to life with just a prompt is like very compelling.
我认为我们会看到很多公司都在做这样的事情。
And I think we'll see a bunch of of companies doing this.
我明白了。
I see.
用Claude code来构建产品。
Claude code to build the product.
嗯。
Mhmm.
用类似Open Claw的东西来做其他所有事情。
Something like an Open Claw to do everything else.
去推广它。
To market it.
是的。
Yeah.
使用电子邮件。
To use use the email.
是的。
Yeah.
使用电子邮件。
Use the email.
使用邮件。
Use email.
也许还可以顺便做一下会计工作。
Maybe kinda do accounting also.
是的。
Yeah.
是的。
Yeah.
目前，我认为我们看到的产品非常直接，这在现阶段是理所当然的——如果你想扩大业务，那就花在Meta广告上吧。
Right now, I think the products we've seen are are pretty, like, straightforward as they should be at this stage and that like, okay, you wanna grow your business, let's spend on meta ads.
但你可以想象，一个月后，这些代理型产品会抓取你所在领域所有Instagram创作者的目录，然后主动给他们发私信，提供合作之类的机会。
But you can imagine a month from now, some of these agentic products will have will scrape the directory of all the Instagram creators in your space and then we'll like cold DM them and offer them a partnership or something like that.
因此，我认为这类类似Shopify的崛起会是一个有趣的类比，因为任何人都可以创建一个消费品牌。
And so I think the rise of these like kinda like a Shopify would be an interesting analogy in that anyone could create like a consumer brand.
我觉得任何人都能创建一个数字业务。
Like, I think anyone will be able to create like a digital business.
这很有趣，因为我觉得我的业务——内容业务——本来会被生成式AI淹没，是的。
That's interesting because I thought my business, which is like the content business Yeah.
本来会被生成式AI淹没。
Was going to be overrun by generative AI.
是的。
Yeah.
确实如此。
And it sure has.
我的意思是，这种情况确实存在。
I mean, the slob is real.
没错。
It is.
我们得想办法解决
We gotta do something about
这个，是的。
this Yes.
但实际上，你的业务也会被基于这项技术创建的初创公司所冲击。
But actually, your business is also gonna be overrun by startups being built on this technology.
没错。
Yep.
那么，你会对使用这些工具的人提供某种折扣，还是会更倾向于投资它们？因为你知道，他们正在驾驭这些工具并构建东西，这虽然不是捷径，但也不一定只是捷径。
Then do you apply some sort of discount to someone who uses these tools, or do you become more likely to invest in them Because, you know, they are wrangling these things and building something, and it's not necessary it's a shortcut, but it's not necessarily a shortcut.
是的。
Yeah.
从投资者的角度来看，你怎么看这个问题？
How do you think about that from an investor standpoint?
对。
Yeah.
我觉得这里有个小小的平衡点，但当一家公司更倾向于以AI为核心来使用工具时，我们会更兴奋。
I think there's, like, a a little bit of a happy medium, but more we get more excited when a company is kind of more AI first in terms of the tools they use.
我不说100%完全依赖AI的原因是，这些工具还不够完善。
The reason I don't say, like, a 100% AI all the time is because the tools aren't perfect yet.
我听说过一些案例，比如我的初级工程师用Claude自动生成所有代码，然后我们上线后就出问题了，一堆bug。
So I've heard cases of, like, my junior engineers are just Claude coating everything and then we ship it and then it breaks and there's a lot of issues.
但我认为，随着这些工具快速迭代和提升，总体上我用它们来完成很多工作。
But I think in general with how fast the tools are kind of compounding and improving, like I use it for a lot of my work.
曾经有一段时间，我会手动逐个核对每一个数字、每一个数据。
There was a period of time where I would double check every single number, every single stat manually.
而且准确率非常出色，并且一直在不断提升。
And like the accuracy rate is fantastic and has only gotten better and better.
因此，我认为我们会更加关注那些以AI为核心、如何运营自身的公司。
And so I think we'll be even more excited about companies that are AI first and how they run themselves.
所以这扩大了你可能投资的创始人群体。
So it expands the pool of founders you might
投资。
invest in.
你有没有看到来自你以前从未想象过的地区或背景的创始人？
Are you seeing founders come in from, like, areas you geographies, backgrounds
当然有。
Absolutely.
你以前根本想不到的吗？
You never would have imagined?
比如，有没有那种60岁、一辈子都在同一家公司工作、从未接触过开发者的老人，现在正拼命学习？
Like, is there, like, just to get, some, you know, 60 year old guy who's been at a company his whole life and never met that developer and now is just clawing his way through?
是的。
Yes.
当然。
Definitely.
我认为我们确实看到一些以前没有大型科技中心的地区开始崛起。
I think we've definitely seen some geographies pop up where we haven't seen huge tech hubs before.
我觉得巴黎就是其中之一。
I would say Paris is one.
斯德哥尔摩也是，比如Lovable，现在还有一整批新兴公司涌现。
Stockholm is another with Lovable and now a whole kind of other wave of companies.
其中一些创始人确实会搬到旧金山，主要是因为那里人才密集。
Some of those founders do move to SF just because of the the talent density.
不过总的来说，在AI之前，如果你在做企业级软件，比如为暖通空调系统开发软件，你想要投资的通常是那个真正懂暖通空调、熟悉市场、之前就在这个领域创业过的人。
I think in general though, like pre AI for if you were building, especially in enterprise business, say software for HVAC, like, the people you would wanna back in that market is the guy who has done HVAC, you know, knows the market inside out, built a company there before.
而我们现在看到的更有潜力的，反而是那些非常拼命、执行力强的团队，他们能够跟上模型发展的速度，以最具吸引力的方式持续将模型产品化并交付给客户。
And actually, what we're seeing now are the better bets are like the very scrappy high hustle teams that will be able to keep up with the pace of model development and continue to productize the models in the most compelling ways and ship them to customers.
所以，现在获胜的创业者类型可能有所不同。
Like, So it's a different maybe archetype of founder that's kind of winning now.
你对AI如何影响工作有一个有趣的观点。
You had an interesting thought about how this impacts work.
有一份《哈佛商业评论》的报告指出，AI并不会减少工作量。
There was this Harvard Business Review report that AI doesn't reduce work.
它实际上会加剧工作量。
It actually intensifies it.
你说作为重度AI用户，你做的工作更多了，而不是更少，因为你获得了巨大的杠杆效应，想法更容易落地。
You said as a heavy AI user, I'm doing more work, not less, because I get so much leverage and it's easier to get ideas off the ground.
是的。
Yeah.
我完全同意。
Fully agree with it.
我的意思是，这份报告就是一个很好的例子。
I mean, this report is a good example.
这是我们做的第六份报告了，可它却是我们出过的篇幅最长、内容最详实的一份——因为我能借助这些产品里的分析、调研和其他各类工具。
It's the sixth one we've done and yet it's like the longest and most dense one that we've done because I was able to leverage analysis and research and other tools from some of these products.
我平时日常工作里也都是这么做的。
And I do it in my in my day to day.
搁以前开推介会的时候，我得同时兼顾好几件事：集中注意力听讲、提出有深度的问题，还得疯了似的把每一条笔记都敲下来。
Like, it used to be when I was on a pitch meeting, I would have to be both paying attention, asking thoughtful questions, and like frantically typing every single note.
现在你就能省心不少，能全心投入地和创始人交流，提出更有质量的问题了。
Now you can granola it and like really engage with the founder and ask better questions.
所以最终的结果就是，它能让我一天里完成更多事情，启动更多项目。
And so like the net net is that it allows me to like get more things done in a day, spin up more projects.
但情况并不是说，哪怕我完成了两倍的工作量，我就会累两倍。
But I'm not like, you know, if I'm if I'm getting two times more work done, I'm not two times more tired.
正相反，用了AI之后我反而比之前更轻松了，因为它带来的助力太大了。
If anything, I'm, like, less tired than I was using AI because it's so much leverage.
不过《华尔街日报》之前有过这么一篇报道。
There was but there was this Wall Street Journal story.
你可能看过《华尔街日报》那篇报道，讲的是有些CEO认为，其实那些低强度的工作任务是有益的，因为你需要这些不太费神的活儿。
You might have seen it with this, like, CEO, a bunch of CEOs who are, like, actually, busy work is good because you need those low intensity tasks.
如果你一直处理高强度的工作，那么你会更快地感到筋疲力尽。
And if you're working on more intense work all the time, then you are gonna burn out more quickly.
有意思。
Interesting.
我我有点觉得那是在扯淡。
I I kinda thought that that was bullshit.
我是说，换个角度想，你大可以把这些时间省下来好好享受生活，而不是一直在做那些瞎忙活的事务性工作，你懂吧。
I mean, the other view was like, you could just take that time and, like, enjoy your life, you know, instead of doing busy work tasks.
谁知道呢。
Who knows?
太离谱了。
Ridiculous.
对。
Yeah.
没错。
Exactly.
我认为，在人工智能时代，我们的工作方式、工作时间和工作方法都将发生变化。
I do think the way that, like, we work and when we work and how we work is going to change in the AI era.
一个很好的例子是语音输入在企业中的广泛应用。
Like, one great example is voice dictation has blown up in enterprises.
这最初起源于语音编程，工程师们只需对着麦克风说话，Cursor就能自动生成代码。
So it started with vibe coding where engineers would just talk into a mic and it would like produce software for them in cursor.
现在，这种技术已经扩展到了销售、市场和商业领域。
And now it's spread to like sales, marketing, business.
但在开放式办公环境中，每个人都能听到别人在说什么，这并不适合这种工作方式。
And that is not well suited to like an open office where everyone can hear what everyone else is saying.
因此，我认为为了适应人工智能时代，将会出现一些文化和环境上的变革。
So I think there's gonna be some like cultural and even environmental changes that are gonna happen to adapt to kind of the AI world.
关于Open Claw还有最后一点
One more thing about Open Claw and
然后
then
我认为，如果我没理解错的话，它的一个显著优势是具有持久记忆。
we'll I move think that one of the compelling advantages of it, if I get it right, is that it has persistent memory.
是的。
Yeah.
它会记住你是谁、你的偏好
So it will remember who you are, your preferences
对。
Yeah.
而且不会像金鱼记忆那样，每次刷新聊天就丢失所有信息，你用Yes时就会看到这种情况。
And doesn't lose that every time you refresh the chat like Goldfish Brain like you'll see with Yes.
ChatGPT 和 Claude。
ChatGPT and Claude.
尽管它们
Although they are
越来越好。
Getting better.
把这个加进去。
Putting that in.
你还在X上发了一篇帖子。
And you also had a a post on X here.
你说，记忆是当前消费级AI中最引人入胜的主题之一。
You say memory is one one of the most fascinating topics in consumer AI right now.
如果做得好，带有记忆功能的应用程序能提供比以往任何软件产品高出百倍的体验。
Done well, apps with memory can provide a 100 x experience on any prior software product.
它了解你，并能适应你。
It knows you and adapts to you.
再详细说说这一点吧，我觉得这非常有洞察力，而且完全正确。
Just expand upon that a little bit because I think that's really perceptive and and right.
是的。
Yeah.
我认为这就像拥有一个伴侣、导师或教练，他们始终陪伴在你身边，理解你所经历的一切，并能提供更优质的建议或意见。
I think it's like the concept of having say if you had like companion mentor or coach who was side by side with you and understood everything you were going through and then was able to provide like much better advice or opinions.
我甚至在想一个例子，比如我正在和Claude交谈，它帮我写备忘录，而它了解我对这家公司的看法、我通常写备忘录的方式，这些都非常有帮助。
I'm even thinking of an example of like, if I'm talking to Claude and it's helping me write a memo, the fact that it knows like how I feel about this company, how I typically write memos, all of that is very helpful.
我经常用ChatGPT处理健康相关的事情，我发现在这方面也极其有用，因为长期跟踪这类信息是很困难的。
I use ChatGPT for a lot of health stuff and I found that that is incredibly useful there too because keeping track of that kind of thing over time is hard.
我认为记忆功能仍有一些问题需要解决，是因为人们用这些产品处理的都是如此私密的个人事务和专业事务。
The reason I think that memory, there's still things to figure out is because people are using these products for such intimate personal things and professional things.
比如ChatGPT的Pulse产品，它会根据你讨论的内容每天给你发送一份简报。
So for example, the ChatGPT pulse products, which is basically like it sends you a briefing for the day based on things that you're talking about.
对我而言，它会把最严肃的工作事项和最私人的事情混在一起，把这些内容集中呈现在一个界面里，会让人感到困惑。
For me, it will combine, like, the most serious work thing with, like, the most personal thing and having that surfaced in one interface is confusing.
因此，我对模型公司如何根据用户与它们交谈时的语境、话题来区分记忆和上下文，非常感兴趣。
And so I'm interested in how the model companies maybe segment memory and context based on when a user is talking to them, what they're talking about, that kind of thing.
因为你把这么多个人信息
Since you're putting so much personal stuff
是的。
Yeah.
在聊天机器人中，ChatGPT 有一个设置。
In the bot, there is a setting within ChatGPT.
如果你允许我用你的对话来改进模型，是的。
If you use my chats to make the model better Yeah.
这允许系统根据你输入的内容进行训练吗？
Which allows it to train on the material you input there?
对。
Yes.
你是开启还是关闭了这个选项？
Do you have that toggled on or off?
我开启了。
I have it on.
我是个AI极致主义者。
I'm an AI maximalist.
知道吗？
Know?
我们干脆在这上面全力以赴吧。
We might as might as well go all out on this.
你就不担心你的私人对话会被
You don't worry that your personal conversations will show up
我的意思是
I mean
别人看到吗？
read by somebody.
我其实没那么担心，因为我了解模型公司在这方面有多谨慎。
I'm less scared about that just because I know how how careful model companies are around that.
我的ChatGPT开启了双重验证。
I do have two factor offset on my ChatGPT.
好的。
Okay.
所以，希望这很难实现。
So hopefully, it's hard to have.
基础。
Ground.
是的。
Yeah.
简要谈谈变化的速度。
Briefly about the pace of change.
嗯。
Mhmm.
从你作为风险投资人的角度来看，这也很有趣。
This is also interesting from your point of view as a as a VC.
你提到两年前、三年前，或者两年半前。
You talked about just two years ago or three years ago or two and a half years ago.
试着回忆一下。
Try to remember.
2023年9月是两年前左右的事了。
September 2023 was two and a years ago.
在你列出的100款顶级应用里，有7款创意工具都是图像生成器。
Seven of the nine creative tools on your list of 100 top apps were image generators.
三年后，仅剩3款图像生成器还在列表中。
Three years later, only three image generators remain.
我是说，回到我们刚才的话题，这些工具基本上都被吞并了，是的。
I mean, they basically, going back to our conversation, got gobbled by Yeah.
ChatGPT。
ChatGPT.
我不知道有没有人，或者说有没有任何人预见到这件事的发生。
I don't know if everybody, anyone saw that coming.
不过也有可能我们是有心理准备的，毕竟他们之前推出过Dolly。
Or maybe we did because they had Dolly.
但能不能请你稍微聊一聊，你是怎么理解这种变化的速度的？
But just talk a little bit about, how do you wrap your head around the pace of change here?
因为一些看似强劲的趋势，比如MidJourney这类工具，可能转眼就消失了。
Because something that can seem like a strong trend, like the sort of mid journeys place and all this can just be gone.
是的。
Yeah.
我同意。
I agree.
我认为，特别是在图像领域，正如我在报告中提到的，我们并没有看到像视频或音频等领域那样，模型公司大规模碾压初创企业的情况。
I think with image in particular, and and I mentioned this in the report, but we haven't seen the same kind of model companies crushing startups in like video or audio or other things.
我认为对于谷歌和Gemini来说，进入图像领域非常自然，因为他们拥有所有的YouTube数据。
I think for Google, Gemini, it was very natural for them to go into image because they have all the YouTube data.
他们还有其他可用于训练的数据。
They have all the other data they can train on.
ChatGPT，你说得对，因为他们有Dolly，所以可能比原本会做的更早、更深入地进入了这个领域。
ChatGPT, think you're right that because they had Dolly, they went in in there maybe harder than they would have otherwise.
对我来说，总体趋势是，像Nano Banana和ChatGPT在处理较为简单的提示时非常擅长生成图像，比如制作表情包、通用宣传单或广泛的营销素材之类的东西。
I think the general trend for me is like Nano Banana and ChatGPT are great for image generation if it's like a fairly straightforward prompt and you're getting out like a meme or a general like a flyer, a broad based marketing asset, something like that.
但我们仍然看到一些图像生成公司出现在列表中，它们更像是复杂的 workflows，比如 Civitai 用于 comfy UI 模型构建，或者像 MidJourney 这样的平台，依然存在于那些对审美有更高要求的用户群体中。
But we are still seeing some image generation companies on the list that are either more like sophisticated workflow like Civitai for, you know, comfy UI model builders or something like a mid journey, which is still on the list for people who are more kind of aesthetically opinionated.
但我确实认为，如果你直接处在大型模型公司正在构建的产品路径上，你就必须对如何包装模型以及如何交付输出有更明确的主张。
But I do think that some of these if you're directly in the path of what the big model companies are building, you have to be a lot more opinionated about how you package the model and and how you deliver an output.
而且，希望你能针对某一类愿意为这种特定工作流支付高价的用户来实现这一点。
And hopefully, you do it for a specific type of user that's willing to pay a lot for that specific workflow.
我们会不会达到这样一个阶段：你可以输入更复杂的提示，比如建筑设计，然后它就能无差错地生成结果？
Are we gonna get to a place where you can prompt something more sophisticated like an architectural design and it will spit it out without errors?
是的。
Yeah.
我的意思是，Nano Banana 已经相当不错了。
I mean, Nano Banana is already quite good.
我们之前在报告中做了一个全球 AI 采用率的热力图，按国家划分。
We did there's a chart in the report where we did kind of a heat map of global AI adoption by country.
于是我给了 Nano Banana 一份国家列表和对应的热力图评分，它完美地为每个国家填上了正确的颜色。
And so I gave Nano Banana the list of countries and the heat map score, And it perfectly filled in every country with the right shade of Yeah.
哇。
Wow.
基于这种采用率，我认为这真的非常惊人。
Based on the adoption, which is I think really spectacular.
我认为，你或我今天用ChatGPT或Nano Banana就能提示生成一个出色的建筑模型。
I do think that, you know, you or I could prompt a great architectural model on ChatGPT or a Nano Banana today.
我不认为那些非技术背景的普通建筑师能够或愿意这样做。
I don't think that the average architect who is nontechnical can or wants to do that necessarily.
因此，我认为我们将继续看到那些将提示作为产品一部分的、在更专注的使用场景中真正成功的产品。
And so I think we're gonna continue to see products that, like, the prompt is part of the product kind of really succeed in those more focused use cases.
在视频方面？
On video?
是的。
Yeah.
Sora 去年曾是短暂的年度爆款。
Sora was the, like, runaway hit of the year last year for, like, a half a second.
是的。
Yeah.
这就是你在Sora上看到的。
This is what you have on Sora.
Sora花了二十天，这可不是个小数目。
Sora spent twenty days, which is not insignificant.
太多了。
It's lot.
登顶美国App Store榜首，下载量增长速度比ChatGPT还快。
Top of The US App Store and reached 1,000,000 downloads faster than ChatGPT.
从那以后，下载量已经下降了。
Since then, have have decreased.
我觉得这么说可能有点轻描淡写了。
I think that's sort of putting it likely.
它已经彻底销声匿迹了。
It's fallen off the face of the earth.
嗯嗯。
Mhmm.
那里发生了什么？
What's going on there?
是的。
Yeah.
那个令人震惊的案例真的很有意思。
The the sordid is really interesting.
我认为，那个实验中蕴含了大量值得学习的经验。
There's there's like a ton of lessons, I think, embedded in that one experiment.
所以，我首先想说的是，这个模型实际上非常出色。
So the first thing I would probably say is the model is actually very good.
我认为它在音频和视频的真实感方面，已经接近类似V3的水平。
I think it's close to something like a v o three in terms of like realism on both the audio and the video.
他们最聪明的突破是Cameo功能。
Their big unlock which was super smart was the Cameo feature.
所以关键在于，为什么每个其他Sora都像是Jake Paul，因为他授予了他们使用Cameo功能的权利。
So the fat that was why like every so every other Sora was like Jake Paul because he granted them the right to like use the Cameo.
这使得它迅速走红，因为人们都在制作朋友的梗图。
And that's what made it go viral because people were making memes of their friends.
但由于这些视频可以导出，最好的Sora视频会被上传到TikTok或Reels，然后与最优秀的人类创作视频竞争。
But because the videos were exportable, what would happen is that the best Sora videos would get uploaded onto TikTok or Reels, and then they would compete against the best human made videos.
因此，在其中一个平台上，整体的浏览体验比单独使用Sora要好得多。
And so the overall feed experience was just kind of strictly better on one of those platforms than on Sora alone.
到那时，下载量已经大幅下降。
Downloads are way down to that point.
我认为它并没有成为他们原本希望的那种社交网络。
I think it hasn't become the social network that they maybe hoped it has.
它取得成功的地方在于作为创意工具，因为这个模型相当出色。
Where it is succeeding is as a creative tool because the model is quite good.
所以他们仍然拥有三百万个DAO，并且这个数字实际上正在随着时间缓慢增长。
So they still have 3,000,000 DAOs and it's actually, you know, slightly climbing over time.
没错。
Mhmm.
也就是日活用户。
Daily active users.
对。
Yes.
日活用户。
Daily active users.
没错。
Yes.
所以大家还是把它当作一个创作工具来用，但不会把它当成社交网络产品来使用。
So people are still really using it as a creative model, but they're not using it as like a social graph product.
目前还没有谁能真正做成AI社交平台。
We haven't seen anyone crack AI social yet.
我觉得这件事会非常棘手。
I think it's gonna be really tricky.
现在，当我们即将结束时，我想和你谈的最后一件事。
Now, you know, last thing I wanna talk to you about as we come to a close here.
在我们之前的对话中，你提到你认为人工智能基本上会重新定义商业。
Earlier in our conversation, you mentioned that you envision that AI will basically be this reimagination of business.
是的。
Yeah.
如果我理解得对的话，所有企业告诉我。
That all biz tell me if I'm getting this right.
所有企业都将被重新塑造为一家人工智能公司。
All businesses will be reinvented as an AI company.
是的。
Yeah.
那么现有的企业会怎样呢？
What happens to the incumbents?
是的。
Yeah.
过去六个月里，这种情况也发生了很大变化。
This has changed a lot in the last six months too.
我认为许多现有企业之所以没有及时应对，是因为它们规模大、成功，当AI刚出现时，难免有些松懈。
I think a lot of incumbents were understandably because they're big and successful companies, like a little bit of sleep at the wheel when AI first came out.
我们确实看到它们开始反击了。
We're definitely seeing them start to fight back.
比如谷歌有四个独立产品上榜，如果你在24个月前，也就是Bard刚推出、Gemini早期版本的时候告诉我这件事，我根本不会相信。
Like Google has four standalone products on our list, which if you had told me that twenty four months ago when like Bard came out, the early version of Gemini, I would not have believed you.
Gemini，NotebookLM。
Gemini, NotebookLM.
NotebookLM、AI Studio，还有Google Labs。
NotebookLM, AI Studio, and then Google Labs.
Google Labs 是你访问Flow和创意模型的地方。
So Google Labs is where you access Flow and the creative models.
AI Studio 是为开发者准备的。
AI Studios is for developers.
好的。
Okay.
是的。
Yeah.
我认为我们正在看到所有现有企业都出现这种情况。
And I think we're seeing that across incumbents.
比如，许多垂直领域的软件公司，像ServiceTitan或Workday，都在集成AI功能。
Like, a lot of these vertical software players, things like, you know, service titan or Workday are kind of building in AI features.
我认为问题在于，特别是如果它们有风险会蚕食自己的产品，就必须改变商业模式。
I think the question is, especially if they're at risk of kind of cannibalizing their own products, you have to change your business model.
比如，它们会比那些正在构建AI原生版本的新创公司更快地抢占所有使用场景吗？
Like, are they going to eat all the use cases faster than the new startup that's building the AI native version of them kind of eats them?
特别是考虑到现在正在成立的公司数量，它们很可能选择AI原生的软件产品，而不是那种二十五年前的遗留软件产品。
And especially for if you think about how many companies are being founded now, they're probably gonna pick the AI native version of a software product, not like the twenty five years old legacy version of a software product.
所以我认为这种转变不会立即发生，这就是为什么我觉得SaaS末日论有些夸大其词。
So I think it's not gonna be immediate change, is why I think the SaaS pocalypse is a bit overblown.
但这确实是一个真正的风险。
But, like, it's definitely a real risk.
是的。
Yeah.
去年年底的时候，我和萨姆·阿尔特曼在一起。
And I was with Sam Altman at the end of the year last year.
他谈到，他认为下一代胜出的软件将是那些从零开始就以AI为核心构建的软件，而不是后期附加的。
He talked about how he believes that the software that will win in the next era will be those that are built ground up AI Yeah.
而不是后期附加的。
Not bolted on.
是的。
Yeah.
是的。
Yeah.
所以这种情况是有可能发生的。
So that could happen.
我很大程度上同意这一点。
I largely agree with that.
我认为在某些领域，现有企业更容易通过掌握你的数据来锁定你。
I think it's harder in some categories where where the incumbent can kinda lock you in because they have your data.
他们拥有所有的集成系统。
They have all integrations.
切换起来太麻烦了，但这种情况终将发生。
It's such a pain to switch, but it's going to happen.
我只是觉得在某些行业中，这需要数年时间，而不是像Citrini报告所说的那样几分钟内就能实现，任何东西都能用氛围编码。
I just think in some of these industries, it's gonna be years, not like the Citrini report was like in minutes, anything can be vibe coded.
我觉得那离我们现在的状况还有点远。
I think that's a little far from where we
Citrini报告只是有点夸大其词。
The Citrini report was just, like, a little bit overblown.
用DoorDash作为例子是个错误。
DoorDash was a bad example to use.
对。
Yeah.
我同意。
I agree.
我本来就不太觉得那份报告站得住脚
I didn't really think it was
别的例子不说，为什么非要举DoorDash的例子？
Out of anything, why DoorDash?
但我确实觉得，从某些层面来看，这场“SaaS末日”或许暗藏更多内情。
But I do think that in some ways, maybe this sasspocalypse has more to it.
嗯哼。
Mhmm.
如果我们认同你刚才的观点，那这些公司确实都面临着——可能不是迫在眉睫，但至少是中期的风险。
If you're if we believe what you do what you're arguing here, then these companies didn't have these like, maybe not immediate, but even middle term risks.
对。
Yeah.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。