重新思考 AI 交互：从文本框到更好的设计

在任何人发送提示词之前，目标应该是让提示词足够简洁、准确、有用、详细、有上下文，从而最大程度降低得到非常通用且不太有帮助的回复的可能性。所以我想或许应该让人们慢下来，不要急于提示。这并不难，绝对不是。这只是因为我们被 AI 热潮冲昏了头，忘记了使用这些方法。我们可以做得更好。欢迎来到 Dive Club，我是 Rid，这里是设计师永不止步的学习之地。本周的嘉宾是 Vitali Freriedman，他长期以来一直是 UX 领域的领先思想家之一，也是 Smashing Magazine 的创始人——过去 18 年里你几乎肯定访问过这个网站。这次对话将深入探讨 AI 如何影响我们与数字产品的交互方式，以及我们的流程和模式如何随之演变。话不多说，让我们开始吧。你还记得第一次体验 ChatGPT 的那个神奇时刻吗？可能有人给你发了那个聊天工具的链接，然后你想，好吧，我可以聊天。于是你进去，在文本框里输入一些内容，那个神奇的文本框居然真的能理解你。这是一个你难以忘怀的神奇时刻。然后你发送了内容，它开始“思考”——虽然它其实从不思考——然后返回给你一些看起来有意义且合理的东西，你可以问任何问题。这感觉就像一个魔法盒子。我认为，这种第一次体验魔法盒子的经历，对很多人来说是一生一次的瞬间。就像有一个“之前”和“之后”的分界线。兴奋感就来源于此。但当我们稍微退一步想想，它不过是一个文本框。我们已经使用文本框很多年了。我们知道如何设计出令人难以置信、完美无瑕、美观的文本框，对吧，Mike？看看你，我确信你能设计出可以放在 Mona 和 Lou 旁边的文本框。然后你有了那个神奇的文本框，你在里面输入内容，这有很高的交互成本，因为人们非常不擅长表达意图。他们非常非常不擅长说出自己想要什么。即使他们说了想要什么，也不代表他们真的那么想。这非常复杂。人类是奇怪的生物。然后我们还需要等待。每当你向 ChatGPT 或任何 AI 工具发送内容时，有时要等 20 秒、30 秒、40 秒，甚至一分钟。如果你使用深度研究功能，可能要等 5、7、10 分钟。但如果你坐电梯按了 4 楼，你不想等 20 秒，甚至 5 秒都不想等。我每次按了按钮后都会很紧张——为什么门不关上开始移动？我等不了 40 秒。然后有时它会一直重复自己。我经常感到困惑，我们很多人对 AI 很不耐烦：为什么它花这么长时间？为什么我要重复自己？为什么它总是忘记事情？为什么它这么烦人？为什么它这么不准确？为什么它带我去虚假的地方？为什么它总是带我去我不想去的地方？为什么我要一直纠正它？那不是我的问题。然后我现在还需要学习如何与 AI 对话的语言。例如，让我非常沮丧的是，我们看到所有那些精彩的提示工程指南，我认为这可能是一种错误的方法。为什么我要学习如何提示？为什么世界上所有人都要学习如何提示？难道 AI 不应该更好地理解我吗？交互成本、表达意图的代价，为什么落在用户肩上？为什么不能直接整合到 AI 的工作方式中？所以，与其用聊天机器人、文本框，不如让它消失？我们仍然需要表达意图。但你知道我们可以做什么吗，Mike？也许，我不知道，按钮。我们可以设计出色的输入框，但也可以设计出色的按钮、单选按钮、滑块、复选框等等。为什么总是只有我来问 AI 问题？为什么它从不问我问题？

Before anybody sends a prompt, the purpose should be to make it so succinct, so accurate, so useful, so detailed, so contextual that the chance of getting a very generic and not very helpful response is minimized. So I want to maybe slow down people in prompting. This is not difficult stuff. I mean, by no means. This is something that we just forgotten to use because they're getting a little bit too excited about AI hype. We can do so much better. Welcome to Dive Club. My name is Rid and this is where designers never stop learning. This week's episode is with Vitali Freriedman who's been one of the leading thinkers in UX for a very long time and he's the founder of Smashing Magazine, which you've almost certainly been on at some point in the last 18 years. So, this conversation is a deep dive into all of the ways that AI is impacting how we interact with digital products and how our processes and patterns are evolving as a result. So, without further ado, let's dive in. Do you remember that magical moment when you experienced your GBT for the very first time? Somebody maybe sent you a link to that chat thing and like, okay, I can do chat. And so you went in and you could send something in a text box and then the text box that magical text box would actually understand you right this was like this one of those magical moments that you just don't forget that easily and then you send it something and then it thinks although it never thinks really and then it sends you back something that seems to be even meaningful and reasonable and you could ask anything. This felt like a magic box. And I think that this thing, this experience of having this first experience with magic box, this to many people was this one once in a lifetime moment. Like there was a before and after. And this is where excitement came from. But when we zoom out for a second, just think about it like a text box. I mean, we've been having textbooks for years. This is like I mean, we know how to design incredible, impeccable, beautiful text boxes, right, Mike? I mean, look at you. You can design the text box that would sit next to Mona and Lou. I'm pretty sure about that, right? And then you basically have that text box, that magical text box, and you type into it, which has a high interaction cost because people are very bad at articulating intent. They're very, very bad at articulating what they want. And even if they say what they want, it doesn't mean that they mean it when they say it. It's very, very complicated. People are strange creatures. And then we need to wait. So whenever you send something to CH GBT or AI tool of any kind, you're waiting sometimes 20 seconds, 30 seconds, 40 seconds, sometimes maybe even a minute. Sometimes if you go into deep research, you might wait for like 5, 7, 10 minutes, right? But if you ever clicked like went into an elevator and you clicked on floor 4, you don't want to wait for like 20 seconds, even 5 seconds. I mean, it always I'm always getting so nervous when it's just I pressed you. Why don't you just close the doors and start moving? I can't wait till like 40 seconds, right? And then sometimes it's just repeating itself forever. It's like I always getting so confused by a lot of us being very impatient with the eye like why does it take so much time? Why am I why do I have to repeat myself? Why does it keep forgetting things? Why is it so annoying? Why is it so inaccurate? Why does it bring me to fake places? Why does it keep like bringing me to places where I don't want to be? And why do I have to correct it all the time? That's not me. And then I now need to learn the language of how to speak to AI. Like for example, what I really get frustrated about that we see all the wonderful prompt engineering guides and I think that maybe that's kind of a wrong approach to this. Why should I learn how to prompt? Why should all the people in the world learn how to prompt? Shouldn't AI understand me better? Like the cost of interaction, a code of articulating intent, why does it leave on the user shoulders? Why shouldn't be just integrated into how AI works? And so instead of chatbot, instead of text box, what if we could just have it disappearing? We still need to articulate intent. But you know what we could do, Mike? Maybe, I don't know, buttons. I mean, we can design incredible input boxes, right? But we can do wonderful buttons and radio buttons and sliders and check boxes and stuff like that. And why on earth is it only me who always have to ask AI something? Why does it never ask me back?

天啊，当你角色互换，让 AI 问你问题时，能创造的价值太惊人了。但几乎所有我交谈过的人，尤其是科技圈外的人，都不知道这甚至是可以做到的。这不在他们对交互方式的认知框架内。我就想，

Man, just the value that can be created when you flip roles and AI asks you questions is amazing. But the almost everybody I talk to, especially people outside of tech, have no idea that that's even possible. Like that's not in their frame of reference for how this interaction can look like. I'm like, man,

这大概是我使用 AI 时超过一半的做法：给它一个提示，我说“问我一堆问题，然后我们一起做点东西”。那么，我们如何设计可供性，帮助人们甚至理解这是可能的？

that's probably over half of what I do with AI is just give it a prompt. I say, ask me a bunch of questions and then we'll make a thing together. And so how can we design affordances to help people even understand that that's possible?

在开始之前，对我来说非常重要的一点是，这些事情中有很多都非常简单。没有什么大魔法。似乎我们过于痴迷于“AI 优先”，以至于忘记了那些人们已经习惯、知道并理解的好东西，因为它们已经存在于心智模型中。

What's really important for me before we even start is that many of those things are remarkably simple. There is no big magic. It's like it seems like we are so obsessed with being AI first that we're forgetting about the good old things that people are used to and people know and understand because it already lives in the mental model.

如果你正在用耳机收听，我强烈建议你转到 YouTube 或 Spotify 的视频播放器，因为 Vitaly 分享了大量非常实用的例子，我们还会一起浏览不同的产品，非常有趣。所以，你可以继续听，但为了获得完整体验，我绝对推荐这一部分配合视频观看。

If you're listening to this in headphones, I highly recommend hopping over to YouTube or the Spotify video player because Vitali shares a ton of really practical examples and we walk through different products and it's a lot of fun. So, you can totally keep listening, but to get the full experience, I definitely recommend a video component for this part.

那么，如果我们进去，比如说探索一下「为 AI 界面寻找有用的设计模式」，它就会开始运行并给我一个列表。通常，在深度研究模式下，如果它思考时间更长，可能会切换。好吧，它大概不会问我什么。让我改成推理模式。虽然这不一定是我想在这里做的。它会开始思考，最终也不会。但如果你进入深度研究模式，它通常会问你几句话，对吧？问你要这个、要那个，等等。通常你得到的是一系列问题，看起来就像这样。然后你需要做的就是把这些全部复制粘贴回文本框，然后对每个问题，大多数人会说「是，我要那个」，对吧？或者「不，我不要那个」。对吧？就是这种故事。为什么？

And so if we go in let's say and explore something like find useful design patterns for AI interfaces and so off it goes and it gives me that list. Typically when it comes to deep research right if it thinks longer maybe switch over. Okay that's not probably going to ask me anything. Let me just change it to reason. Although it's not necessarily the something I would like to do here. It's going to start thinking and eventually no it's not. But if you go into deep research mode, typically it will ask you kind of a bunch of a couple of sentences, right? Asking you do you want this, do you want that? And so on so forth. And usually what you get there is basically a list of questions that would appear very much like this. And then what you need to do with that is you copy paste all of that back into the text box and then to each of them most people would say yes, I want that, right? And no, I don't want that. Right? This kind of story. Why?

是啊。我是说，我们到底为什么要那样做？

Yeah. I mean, why on earth would we do that?

如果你去用 Perplexity，我做同样的事情。我在这里问同样的问题：为界面寻找有用的设计模式。所以，如果我进去点击「现在研究」，它会告诉我，嘿，你想在它实际工作时添加一些细节或澄清吗？因为人们经常忘记一些关键细节，然后他们必须等到 AI 完成才能提供额外细节。在这里我可以输入「为了可访问性」，对吧？然后它实际上会完成工作，并添加那层额外信息。这已经解决了在 AI 思考时添加内容的问题，尽管它并不真正思考。但当然，我们还可以做其他事情，因为我认为在很多方面，我们在这里必须做的这种表述完全是多余的。这又是一个来自 Nikki 的精彩例子，他写了这个模式，基本想法很简单：我们有一个任务构建器。你想在这里做什么？我想提问、搜索、解释或其他什么，也许我想把它集成到 Slack、Gmail、Salesforce 之类的，然后我想从中生成一些东西。所以，如果你想想这个，你可以把它框定为人们经常做的一些任务。你可以说，总结一个 Slack 频道并转换成 Word 文件，或者分析 Salesforce 的一些数据并转换成 PowerPoint。点击点击点击，中间发生的是，当你选择了，比如搜索 Notion 并制作 PowerPoint，你可以说，让我为你创建一个模板。这是一个提示，当然可以扩展和增强。搜索我的 Notion 中的特定主题（这是你必须提供的），并创建一个总结发现的 PowerPoint 演示文稿。点击点击点击，你写下主题，完成，然后你就得到了结果。对吧？所以它有点像任务规划器。我认为这太棒了，对吧？

If you go say for perplexity and I do the same thing. I'm going to ask the same thing over here. Find useful design patterns on the interfaces. So, if I go in and say research now, it tells me, hey, do you want to add some details of clarifications while it's actually working on it because very often people forget some critical details and so then they have to wait until AI is done to then provide extra details. And here I can say for accessibility, right? And so it actually does the work and then it adds that layer that extra information right to it right so that's already solved this problem of actually being able to add something while AI is thinking although it doesn't really really think right so there is that but there are of course other things that we could do at this point because I think in many ways like this articulation that we have to do here it's just totally unnecessary because again this is a wonderful example that comes from Nikki uh in which he wrote about this pattern here and basically the idea is very simple we kind of have a task builder. So what do you want to do here? I want to ask or search or explain or something else and maybe I want to integrate it into I know Slack or Gmail, Salesforce and something like that and maybe I want to make something out of that. So if you think about this, you can actually frame it into some of the frequent tasks that people are doing. You can say you know what summarize a slack channel and turn it into a word file or maybe analyze some data from Salesforce and turn it into a PowerPoint. So click click click and what happens in between though is that while you have selected that let's say you know what was that one search notion PowerPoint. So I want to search in notion and make a PowerPoint. So you can say well let me create a template for you. So this is a prompt and of course it could be extended and augmented. Search my notion for a particular topic that's something that you have to provide and create a PowerPoint presentation summarizing the findings. Click click click you write your topic done and then you basically get the results. Right? So it's kind of like task planner if you like. I think that's incredible right?

是的。

Yeah.

另一方面，我还要说，我们可以做得更多。一个很好的例子是 Consensus。我非常喜欢 Consensus，在很多方面都很喜欢。如果有人还不知道 Consensus，这真的是一个非常好的 AI 体验例子。让我来演示，我可能有点太兴奋了，但假设我们在找一些东西，比如做某种驱动研究，我显然不是任何方面的专家，但我们就选一个。它进去后，会尝试找到这些问题的答案，这并不令人惊讶，然后为我做所有事情，很漂亮。我绝对喜欢的是，当它实际工作时，让我回来展示一下。我喜欢的是，它实际上采用了我们喜爱和关心的老式东西，比如过滤器，并在这里提供给你。这难道不神奇吗？

In other ways I would also say uh that we can do much more than that. One really nice example of that is consensus. I love I mean I love consensus in so many ways. If somebody's not aware of consensus yet, this is really really great example of really good AI experience. So let me guide I mean this is like going all I'm getting too excited about this I guess right but let's say we are looking for something I don't know like let's go for some sort of drive research I'm not obviously expert in any of that but let's just go for that and so just pick one and so it goes in and it actually going to try to find answer to those things that's not very surprising right and does all that thing for me beautiful what I absolutely love while it's actually working on that let me come back and show it over here. What I love is it actually takes the good oldfashioned stuff that we love and care about like filters and it offers them for you right here. Isn't that amazing?

对我来说，这就像为什么我们不把它放在所有地方？

I mean, for me, this is like why don't we have it everywhere?

这太有趣了，因为这个模式如此熟悉，但我认为我实际上从未见过它与……相关。

It's so funny cuz the pattern is so familiar and yet I don't think I've actually seen it in relation to

我以前从未在任何 AI 体验中见过它。所以，当然你可以问任何问题。你可以问任何类型的问题，但如果我只想指定呢？我想说过去五年，也许我至少需要五篇引用，也许特定的期刊排名，也许特定的方法论，我只想看观察性研究，也许病例报告，等等。你应用这些，然后用它执行查询，这比仅仅说「问我任何问题」要有用得多，对吧？

I have never seen it in any kind of AI experiences before. So, sure you can ask anything. You can ask any kind of question, but what if I want to just specify it? I want to say last five years and maybe I'm looking for at least five citations and maybe a particular journal rank and maybe a particular methodology right I want to see only I don't know observational studies right and maybe case reports and you know whatever things like that and you apply it and then you execute a query with that that's infinitely more useful than just saying ask me anything right

这是简单的事情，但实际上有很大的不同。一旦你得到结果，显然你会得到这些引用等等，这并不令人惊讶。但他们有一个我觉得非常令人印象深刻的东西，就是共识计量器。因为 AI 的问题在于，它通常只给你一个答案，你问一个问题，它给你一个陈述或概述，但为什么不给我概述的分布，即陈述的分布呢？

it's simple thing but it actually makes quite a difference and once you get the result here now obviously you get this citation ations and all that and so on so forth right that's uh not very surprising here but one thing that they do have which I find quite impressive is consensus meter because the problem with AI is that usually it gives you just an answer so you ask it a question it gives you a statement or it gives you an overview but why doesn't it give me distribution of overview so distribution of statements

所以它背后的想法是，比如「城市热岛效应是否会降低电力峰值预测的准确性？」你会得到论文告诉你的分布，95% 的论文表示是，少数一篇表示可能，我想没有表示否的。这些东西很有帮助。然后我喜欢这种分层，一方面，在 AI 体验中，我们通常有来源和参考文献之类的东西，但在这里它们还用了颜色编码，绿色表示已确认，黄色表示未确认或存在混合意见。

so the idea behind it is okay does urban heat island effect reduce accuracy in electricity peak forecasting so you kind of get the distribution of of um what papers tell you right 95% of papers indicate yes and a few uh one of them is indicates possibly right and there is I guess none that indicates no those things are helpful and then I kind of like this layering so on the one hand typically in AI experiences we have this sources and you know references and stuff like that but here they're also colorcoded so green meaning this is okay confirmed right and yellow if it comes up somewhere means that it's not confirmed or there are mixed opinions on them.

嗯，所以这些小事情非常令人印象深刻，然后你还可以过滤那部分。你也可以按特定情感过滤输出。我们不应该忘记过滤、排序、搜索等等这些东西都很棒。我们不应该仅仅为了 AI 优先而抛弃它们。如果有的话，我们可能应该更倾向于 AI 第二。

Uh so those things are little things that are really really impressive and then you can also filter that part. You can filter that output by a particular sentiment as well. Like we should not forget that filtering, sorting, searching and all of that stuff is great. We shouldn't be just dismissing it for the sake of you know AI first. If anything we should probably be I don't know leaning more towards maybe AI second.

就用老式的筛选器吧。筛选器很棒。我希望在 AI 体验中看到更多筛选器，这并不难实现。

Just bring the good old-fashioned filters. Filters are great. I would love to see more filters in AI experiences and it's not hard to do.

在底层，你基本上是用一些你希望被尊重的特定细节来增强提示词，对吧？但这不能只是生成文本，对吧？它必须是那种仅生成文本的 AI 与真正尝试理解、分类和归类的系统之间的结合。所以这比常规系统要复杂一些，但我认为这非常棒。这就是我想要的。

You just under the hood, you basically augment the prompt with some specific details that you want to be respected, right? But it cannot be just generating text, right? It must be a combination between an AI that just generates text and something that actually tries to understand, categorize, and classify. So that's a little bit more complicated than just regular systems, but I think that it's incredible. This is what I want.

让我印象深刻的是，这甚至不是简单的要点列表。我们不是在那种来回对话或文档的心理模式下操作。这里有控件和 UI，一切都非常熟悉，但我在这些新的 AI 产品中并不常见到这种设计。

That was the thing that stood out to me is even just differentiating from this isn't a bullet list. We're not operating within this mental model of a back and forth or a doc. It's there are controls and UI and it's all very familiar and yet again I haven't seen it very often in these new AI products.

还有一件经常缺失的事情，因为记住我们面临的主要挑战之一是准确性很差。甚至不光是准确性，我觉得很多人都会假设会有一些幻觉，而且不太清楚你是否能信任某些东西。所以很多时候，这种修正错误或优化过程或验证过程，不管我们怎么称呼它，都很耗时。但接下来，我特别喜欢 Elicit——顺便说一句，如果你没听说过，它是一款非常棒、设计精良的 AI 工具。它有几件事做得非常好，首先它真的力求准确。所以它会获取来源，然后根据你可能选择的具体标准筛选这些来源，并从中提取数据。但我最喜欢的是这些星号。Mike，这些星号是什么？

There's also one more thing that is often missing because remember one of the main challenges that we have is that we have pretty poor accuracy. Not even accuracy, I would say that a lot of people are assuming that there will be some hallucinations and it's not very clear if you can trust something or not. And so a lot of time this sort of fixing errors or refinement journey or verification journey whatever we want to call it takes time. But then what I absolutely love about Elicit and it's a fantastic tool by the way if you never heard of it, it's a really fantastic also well-designed AI tool. It does a couple of things really well and first of all it really tries to be accurate. So it gets sources and then it screens those sources for specific criteria that you might select and then extracts data from that. But what I love most is this asterisks. Mike, what are these asterisks?

嗯，我不知道。

Yeah, I don't know.

它们很神奇，对我来说这真的很神奇，因为通常这只是一些论文的引用。但这会给用户带来负担，因为他们现在必须去论文里找到提到这一点的地方，对吧？相反，它们直接链接到论文的特定段落，对吧？这指明了来源出处。实际上是直接链接到段落，甚至不是链接到整篇论文，对吧？然后你还可以通过那里的几个引用点来查看它实际来自哪里以及为什么出现在这里，对吧？

They are magical and that to me that's truly magical because typically this would be just references to some papers. But that creates a burden that leaves on the user's shoulder because they now have to go to the paper and find a place where this is mentioned, right? Instead, they're linking directly to a particular segment of that paper, right? That indicates where this comes from. Literally direct linking to the segment, not even to the paper, right? And then you can also go through a couple of mentions there to see where it's actually coming from and why it's coming here, right?

因为我只是看着那些引用，然后想，嗯，我认识这些 URL 吗？是的，可能就够了。就像，我信任它。我不会费劲去点击它们。

Because I'm just looking at the references and like, well, do I recognize the URLs? Yeah, it's probably good enough. Like, I trust it. I'm not going to put in the effort of actually clicking on them.

是的，我认为这里还有一点非常重要，那就是没有强调聊天机器人等等。是的，这是一种中心舞台的体验，你需要打字，对吧？或者你需要选择你感兴趣的任何主题，但这里没有聊天机器人，因为此时你正在探索面前的数据，对吧？当然，你可以以某种方式在某个地方调出那个文本框。我很确定这一点。但此时你更像是处于适当的探索研究模式，对吧？

Yeah, and I think that's also one thing that's really important here is that there is no emphasis on chatbot and so on. Yes, this is like center stage experience where you have to type, right? Or you have to choose whatever topic that you're interested in, but there is no chatbot here because at this point you are exploring the data that is in front of you, right? And sure, you can bring back that text box in some way somewhere. I'm pretty sure about that. But you're kind of more in the proper exploration research mode at this point, right?

我也喜欢这一点。当然，这也适用于所有其他方面。我认为这太棒了。这就是为什么我认为这类体验完全不同。它们与传统聊天机器人截然不同，因为它们为那些只需要理解某些东西但没有时间自己翻阅论文的人提供了巨大的加速。但随后这确实支持了某些在 Elicit 中解释的想法、概念和陈述。所以这很棒。

And I like that as well. And it goes also, of course, for all the other things. And I think this is incredible. This is why I think those kind of experiences, they're totally different. They are unlike the traditional chatbots at all because they really provide an enormous speed up for people who just need to understand something and don't have the time to go through papers on their own. But then this really backs up certain ideas and concepts and statements that are being then explained here in Elicit. So that's great.

我记得使用 Elicit 的时候，大概是一年半以前，它真的令人印象深刻，因为感觉就像，哇，他们不只是简单加个聊天功能，AI 已经深深嵌入到这个产品的本质中。我真正感受到这一点是在——可能现在变了——但当你甚至设置一个研究项目时，你可以用 AI 来自定义列。所以你得到的是结构化数据，但你用 AI 来构建你的报告，而且你不需要在输入框里想出一个完美的提示词。就像，不，不，这里有电子表格的熟悉感，但现在我只是在回答一些独立的问题，比如在列级别我想要什么数据。我看到这个的时候就想，‘哦，天哪，这太有道理了。’

I remember going through Elicit, I mean it was probably close to a year and a half ago and it was really impressive because it did feel like wow they did not just slap a chat on like AI is in the very fabric of what this product is. And the moment that I really felt that was I think they had something it might have changed but when you were even setting up a research study you could use AI to customize the columns. So you're getting structured data but you're using AI to structure your report and you're not having to think about it in terms of this perfect prompt inside of an input. It's like no no there's this familiarity of a spreadsheet but now I'm just answering isolated questions of what data do I want at the column level. I saw that and I was like, 'Oh man, that makes so much sense.'

是的，我们也没有那个。我想我开始做的是，如果我看我的 Perplexity，因为我开始设置这种偏好和自定义设置。哦，自定义设置，我想就在这里。当我基本上要求它尽可能以数据表形式显示结果时。哦，酷。我觉得这非常有用。所以这只是一个提示词，然后随每次查询一起提交。你可以在 ChatGPT 中找到，在 Perplexity 中找到，到处都有。这是一种告诉 AI 它需要了解你的哪些信息才能更好地回答你的查询的方式。我非常重视这一点。所以我尽量经常检查它，但空间有限。所以我们在这方面做不了太多。甚至这个文本框也是一个有趣的例子，因为我看着它，我想，是的，这都很有道理。我想要大部分这些。但很多时候我的偏好框是空的，因为我真的不知道里面该放什么。我确定我能想出来，但我可能得和另一个 AI 交互才能弄清楚这个聊天框里应该放什么。

Yeah, we don't have that either. I think that what I started doing and I think if I look at my Perplexity because I started setting up this sort of preferences and customization settings. Oh, customization I think it's here. When I basically ask it to show results in data table whenever it can. Oh, cool. I just find it very useful. And so this is just a sort of a prompt that is being then submitted with every query that you send. So you can find it in ChatGPT, you can find it in Perplexity, you can find it everywhere. A way to tell AI what it needs to know about you to respond to your queries better. I take it very seriously. So this is something that I try to review as much as I can, but it's limited in space. So we can't do much there. Even this text box is an interesting example because I'm looking at this. I'm like, yeah, that all makes a lot of sense. I want most of that. And yet so often my preferences boxes are empty because I genuinely like I don't know what goes in there. I'm sure I could figure it out, but I would have to probably interact with a different AI to even figure out what should go into this chat box.

是的，我的意思是这让我又回到了这里。这是一个很棒的资源。我不知道为什么这么多人不知道它。这太棒了。这是来自 Luke Bennis 的。他提出了一些关于什么是真正好的 AI 体验的想法。他里面有一些非常好的想法，比如这个可能是你会欣赏的。所以这个想法确实有点手把手指导的意思。

Yeah, I mean it's kind of really brings me back to also this here. This is a wonderful resource. I don't know why so many people are not aware of it. This is fantastic. This is coming from Luke Bennis. He has been coming up with a couple of ideas about what could be a really nice AI experience. He has some really nice ideas in there like for example something like this is probably something that you would appreciate. So the idea is a bit of handholding indeed.

所以当你写提示词时，在开始写之前，系统可以问你类似这样的问题：‘好的，你想要什么？有没有参考？你期望什么样的专业水平？’ 这些又只是老式的 UI 控件。但天哪，为什么我们觉得它们在 AI 时代不再需要了呢？我认为这样的东西会很棒。另一方面，他甚至建议了这样的东西，我觉得非常非常酷：提示词的工具辅助创作。所以你在这里开始写提示词，然后有一个 AI 助手在你提供的文本层面上操作，说：‘嘿，让我让它更简洁一点。让我添加一个具体的上下文。’ 也许 AI 可以问需要什么样的上下文，比如：‘这是给谁的？我要给高管做演示，给设计师……’ 也许你只想指定这些，就问几个问题。我认为我的最终目标是：在任何人发送提示词之前，目的应该是让它如此简洁、准确、有用、详细、有上下文，以至于得到非常通用且无帮助的响应的可能性降到最低。所以我想也许让人们慢下来写提示词，对吧？让人们慢下来实际向 AI 系统发送内容，因为显然这不是免费的——首先是因为可持续性，但另一方面它也需要时间，比如 20 秒，也许 30 秒，看情况。然后你必须浏览这堵文字墙，然后你发现少了什么，因为答案根本不是你想要的东西，你漏掉了某个关键词之类的。然后你重新开始。人们浪费了大量时间与 AI 输出反复来回，仅仅因为他们漏掉了东西。所以也许我们确实可以做一些引导。或者直接说：‘嘿，等一下。你这里是想说这个还是那个？’

So when you're writing a prompt, before you even start writing, you could be asked something like: 'Okay, so what do you want? Is there a reference? Is there a particular level of expertise you're expecting?' And again, these are just old-fashioned UI controls. But oh my, why do we feel like they are not needed anymore in the age of AI? I think something like that would be incredible. On the other hand, he's even suggesting something like this, which I think is really, really cool: tool-assisted authoring for prompts. So you maybe start prompting here, and then you have an AI assistant that operates on that level of the text you have provided and says, 'Hey, let me make it a bit more succinct. Let me maybe add one specific context.' And maybe AI can ask what kind of context it needs, like, for example, 'Who is this for? I'm looking for a presentation to executives, to designers...' Maybe you want to just specify that, just ask a few questions. I think that my ultimate goal is: before anybody sends a prompt, the purpose should be to make it so succinct, so accurate, so useful, so detailed, so contextual that the chance of getting a very generic and not very helpful response is minimized. So I want to maybe slow down people in prompting, right? Slow down people in actually sending something to an AI system, because obviously it doesn't come for free—first of all because of sustainability, but then on the other hand it also takes time, like 20 seconds, maybe 30 seconds, depends. And then you have to look through this wall of text, and then you realize that something is missing because the answer is not at all what you wanted, and you're missing like one important keyword in there or something like that. Then you start all over again. People are wasting an enormous amount of time going back and forth with this AI output just because they missed something. So maybe we could do a bit of handholding indeed. Or just say, 'Hey, hold on for a moment. Do you mean this or that here?'

是的。

Yeah.

在提示词层面，而不是输出层面。我很喜欢。我完全——因为这是一个完美的例子——我不知道，那个提示词对我来说是一个非常初级的提示词，你知道，但它仍然是大多数用例。人们进入 AI 然后说：‘给我做个东西。’ 如果你以动词开头，AI 就会做点什么，你知道，它总是会做点什么。而这将是一个完美的例子：如果你对提示工程有一点了解，你可能会说：‘我的目标是创建一个年度 D。先别构建任何东西。换行。让我们来回交流。问我问题，获取你需要的所有上下文。最后，你明白如何成为一个提示工程师。所以成为提示工程师，写出对你来说完美的东西，给你所有需要的上下文来做出好的东西。’ 也许全世界只有 1% 的人明白，要高度产品化就必须进行这种互动。

On the level of the prompt rather than on the level of the output. I'd love that. I totally—because this is the perfect example of—I don't know, that prompt to me is a very novice prompt, you know, but it's still the majority use case. People go into AI and they say, 'Make me a thing.' And if you lead with a verb, the AI is going to make something, you know, it's always going to make something. And this would be the perfect example where with a little bit of understanding of how prompt engineering works, you probably would say something like, 'My goal is to create an annual D. Don't build anything yet. Line break. Let's have a back and forth. Ask me questions and get all the context you need. And at the end of it, you understand how to be a prompt engineer. So be the prompt engineer and write the thing that's perfect for you that gives you all the context you need to make something that is good.' Like maybe 1% of the worldwide population understands that that's the interaction you have to have to highly productize that.

绝对同意。我认为故事是这样的：当我们看所有提示词框架时，它们在很多方面都有特定的结构，对吧？那我们为什么不直接复制这个结构，让人们不必记住它呢？比如，这里还有一个模式，我觉得很棒。也许我们就该这么做。也许这应该是 AI 体验的开始，而不是一个带有开放文本框的聊天机器人，对吧？‘嘿，主要提示词、上下文、背景、输出细节。’ 你知道，有时你会说‘扮演 UX 设计师’或‘扮演财务顾问’之类的。所以也许我们应该在这里以某种方式反映这个结构。这样当你写的时候，你知道该写什么，因为如果它问你‘问我任何事’，那你到底该做什么？你怎么组织？所以我认为这非常非常强大。我的意思是，这些小事情并不难。只是 UI 的东西。真的不难。但这样你就以更结构化的方式增强了人们寻找的东西，给他们一个可以操作的结构，然后你最终也会获得更好的体验，因为输入更好，所以输出也会最好。

Absolutely. I think that it's like the story there is: when we are looking at all the frameworks for prompting, in many ways they have a particular structure, right? So why don't we just replicate the structure in a way that people don't have to remember it? Like, for example, this is another pattern from here as well, which I think is great. Maybe we should do just that. Maybe this should be the start of the AI experience rather than a chatbot with an open text box, right? 'Hey, main prompt, context, background, output details.' And you know, sometimes you say 'act as a UX designer' or 'act as a financial advisor' or anything like that. So maybe we should have that structure somehow reflected right here. So as you write it, you know what to write, because if it asks you 'ask me anything', so what are you supposed to do exactly and how would you structure that? So this I think is really, really powerful. I mean those little things are not difficult. It's just UI stuff. It's really not hard at all. But then you're augmenting whatever people are looking for with this stuff in a more structured way and give them sort of a structure to operate within, and then you end up with a better experience as well because the input is better and so the output will be best.

在我们进入更广泛的 AI 话题之前，关于聊天机器人领域还有什么要说的吗？但这是我很想听你谈论的主要事情之一。所以我很好奇我们有没有漏掉什么。

Anything else on chatbot land before we kind of go into a broader spectrum of AI things that I want to get your take on? But this is like one of the main things that I'd love to hear you talk about. So I'm curious if there's anything we haven't mentioned.

我认为对我来说，真正重要的是让人们慢下来，当他们试图表达意图时。如果他们一开始很快，那么 AI 的慢速就会破坏他们的体验。但如果我们让他们停留片刻，问一些我们需要知道的东西，以便给他们更好的回应，那实际上是一个更好的状态。我唯一想说的是，当你问类似‘设计师在使用 AI 产品时应该考虑哪些常见模式和启发式方法？’时，不是让你打字，而是给你单选按钮或复选框之类的，对吧？我的意思是这样才对，对吧？而不是说‘嘿，回答那个问题，然后回答那个问题，然后回答那个问题’，你可以直接‘砰砰砰砰砰’，如果你想跳过，也应该能跳过，对吧？所以如果它真的以那种方式让你慢下来，你不必一直打字，而是直接选择。因为我们经常看到相反的情况。比如，当你开始 ChatGPT 的深度研究时，在它进入深度研究模式之前，它会告诉你：‘在我开始深度研究之前，这里有几个问题我想让你回答。’ 然后可能有六七个或八个问题，然后你必须把它们全部复制粘贴到文本框里，然后逐一回答。但为什么不直接给我，我不知道，每个问题的单选按钮、文本框、滑块呢？这样我就可以直接回答，而不用复制粘贴任何东西。我基本上可以选择和定制我的路径，然后你就可以去做你的事了。

I think to me it's really slowing down people when they are trying to articulate their intent is really important. If they are fast in the beginning, then the slowness of AI really breaks this experience for them. But if we kind of keep them for a moment and ask them something that we need to know in order to give them a better response, right, then it's actually a better place. The only thing I would say is that when you're asking something like 'What are common patterns and heuristics designers should be thinking about when working with AI products?', instead of asking you to type, it gives you radio buttons or checkboxes, whatever, right? I mean this should be it, right? Instead of saying 'Hey, answer that question, then answer that question, then answer that question', you can just go 'boom boom boom boom boom' and if you want to skip, you should be able to skip as well, right? So if it kind of really slows you down in that way, it's not that you have to type all the time, but you just go. Because very often what we see is the opposite. See that when you start Deep Research in ChatGPT, for example, before it even goes into deep research mode, it will tell you: 'Well, before I even start doing the deep research, here are a few questions that I would like you to answer.' And then there are maybe six or seven or eight questions, and then you have to copy-paste them all into the text box and then answer each of them. But why don't you just give me, I don't know, radio buttons, text boxes, sliders instead for each of those, so I can answer directly without copy-pasting anything? I can basically select and choose my journey, and then off you go and you can do your thing.

我认为这在很多方面都会很了不起，对吧？经常出现的一个想法是动态界面，AI 即时生成的界面，它曾经风靡一时，然后有点降温了，人们说‘你知道吗，实际上可预测性很好。’ 然而对于你展示的那种 Perplexity 界面，这正是动态生成界面的最佳点，其唯一目标是提取上下文。

I think in many ways this would be remarkable, right? Something that comes up a lot is this idea of a dynamic interface, something that AI is generating on the fly, and it was like all the rage and then it's kind of trended down a little bit, and people are like 'you know what, actually predictability is nice.' And yet for that type of interface that you were showing with Perplexity, this is where it's like kind of the sweet spot where dynamically generated interfaces that have the sole goal of context extraction.

这太棒了，因为你只需要用户已经熟悉的最基本的原子组件，然后定制一套控件，让人们更具体地表达他们想要什么。感觉很快几乎所有工具都会这样。

That is like killer because you're always going to only need the most basic atomic components that users are already familiar with and you're just tailoring some kind of a set of controls to get people to be a little bit more specific about what they want. That feels like yeah, you might see that in almost all of these tools here soon.

是的，我希望如此。我的意思是，这也关乎我们如何捕捉用户的上下文，对吧？因为最终一切都围绕这个。在用户旅程的不同阶段，人们需要表达，需要阅读大段文字，然后需要对这些文字做些什么。他们经常想提取、压缩或扩展内容。还有很多微调，因为他们只想得到正确的东西，不管是什么，然后可能发给经理或制作演示文稿之类的，但他们想要有价值的东西。我在测试中发现，很有趣的是，人们拿到 AI 的输出后，会查看、阅读、检查，大致验证是否合适。然后他们挑选内容，可能取第一段、第二段、第七段和第十二段，把它们放到一个独立的地方，比如文本编辑器里。然后在那里编辑，再带回 ChatGPT 之类的工具，让它总结并以有意义的方式重组。这很奇怪。当用户有专门的编辑空间时，这并不好。我们需要缩短人们想做某事和实际做到之间的距离。这非常重要，因为理想情况下，我希望回到 ChatGPT 时能说“等一下，我不需要这个，能删掉这部分吗？”但我不能，我只能就这个问题对话。所以我必须复制到别处再处理。

Yeah, I hope so. I hope so. I mean it's also kind of the story about how exactly we capture users' context, right? Because in the end it's all about that. I mean there are different parts of the journey where obviously people need to articulate, then they need to be going through this wall of text, and then they need to do something with it. Very often they want to extract or they want to compress or they want to expand. There's a lot of tweaking happening too, because they just want to get to the right thing, whatever that is, and maybe send it to a manager or create a presentation or something, but they want something valuable. And I see in testing, it's so funny because people take the output from AI and then they look at it and read it and check it and verify if it's all right more or less or not. And then they cherrypick. They take maybe the first paragraph and then the second and maybe the seventh and maybe the 12th and they pull it all in a separate place, like maybe in a text editor or something like that. And then they do the editing there and then they bring it back to ChatGPT or so to summarize it for them and restructure it in a meaningful way. This is weird. Whenever you have a dedicated space for whatever it is that they're doing here, that's not great. We need to reduce the distance between where people want to do something and when they do that. This is really important because ideally I would love to be able to have an option to just if I go back to ChatGPT here to say hold on for a moment, I don't need this. Can I just remove this part? Well, I cannot because I can just have a conversation about that. So I need to bring it, copy it somewhere and then move from there.

但回到 Perplexity，我挺喜欢它的很多功能。可以说“等一下，让我进入编辑模式”，可能只是不太显眼或不够清晰。可以说“转换为页面”。一旦你这样做，它就会处理并创建一篇关于该主题的文章或页面。但对我来说重要的是，它有目录，你可以导航，我们稍后会看到。所以你可以跳转到特定区域，同时它还在工作。顺便说一句，在 Consensus 上你也能看到类似的功能，你可以这样导航。如果我问一个单独的问题，我可以跳转到那个问题或之前的问题。我可以在响应内部导航。然后在这里，回到 Perplexity。我得到生成的文本。你很快就能做一些我认为其他地方做不到的事情，因为它的设计方式很好地映射了人们使用 AI 的方式。给你。是的，但有点问题。这不是应该发生的。这应该是一个上下文菜单，你可以……

But again, coming back to Perplexity and I kind of like Perplexity for a lot of things. Can say hold on for a moment. Let me kind of go into editing mode. It might be just not as visible or not as clear. Can say convert it to a page. So once you do that, it kind of goes through it and kind of creates an article or a page on that topic. But what's important for me is that it has a table of contents that you can navigate within as we'll see in a moment here. So you can jump to a specific area while it's working. Maybe this kind of something that you also see here on Consensus, by the way, you have this way to navigate. So if I ask a separate question, right, I can go and jump to that question or the previous question. I can really kind of navigate between that response right within that response. And then in here, back to Perplexity. So I get this text being generated. You will be in a moment able to do something that I didn't think you can do anywhere else because of the way of how it's actually designed, because it kind of I think it really maps well into how people use AI as well. Here you go. Yes, but it's broken somehow. This is not what was supposed to happen. This is supposed to be context menu where you can actually...

是的，你能看到。只是缺少背景。

Yeah, you can see it. It's just missing the background.

是的，我不太确定为什么。我想可能是因为我把它弄坏了。但重点是，我现在可以说“让我改一下，也许删除这部分或扩展这部分或做点什么”。对吧，我能写“扩展更多”吗？好的，它可能会扩展。是的，它会。所以真的在输出层面操作，这在其他地方做不到，对吧？这很好。然后你当然也可以说“我想移动列、添加一些部分，还有这些不同的视图。我想在某个地方放一个表格”，对吧？为什么不是到处都有？

Yeah, I'm not quite sure why. I guess because maybe because I broke it. But kind of the point is of this right, that I can say now go and say let me just change that and maybe remove this part or extend this part or do something with it. Right, can I write extend more? Okay, so it's probably will be extending at this point. Yes, so it will. So kind of really operate on the level of that output, which is something that you cannot do much in other places, right? So this is nice. And then you can of course also say I want to move columns around and add some sections and also have this little, you know, different views. I want to have a table on something, right? Why isn't this everywhere?

是的，我没见过。

Yeah, I have not seen that.

这很简单。所以我想，与其说“请以列表格式或表格格式写下来，并让它紧凑一些”，我认为这应该无处不在。

It's simple. So I want, instead of saying please write down in a list format or in a table format or anything like that and make it compact or something, I think that this should be everywhere.

我的意思是默认就无处不在，就像每个 AI 体验中都有。因为，这取决于用在什么地方，但我认为它有巨大的价值。也许不是这个，因为你添加了图片之类的，但这个和这个可能很棒，对吧？所以我有点痴迷于此，因为我觉得我们在某些方面可以做得更好。

I mean just everywhere by default, like in every single AI experience. Because in, I mean it depends on where it's used, but I think it has incredible value. I mean, maybe not this because you're kind of adding images or so, but this and maybe this is great, right? So I'm a little bit obsessed with this because I think in some way I feel like we can do so much better.

我的意思是，这并不难。这绝不是因为我们被 AI 炒作冲昏了头脑而忘记使用的东西。我可以想象一开始有个自动选项，比如“好吧，如果你不想选择，没问题，我们会尽力而为”。但这是一个非常直观的小控件，我会每天都用。

I mean, this is not difficult stuff. I mean, by no means this is something that we just forgotten to use because we're getting a little bit too excited about AI hype. I could see maybe like an auto at the beginning where it's like, okay, if you don't want to make the choice, fine. We'll do the best that we can. But this is such an intuitive little control that I would use it every single day.

我也是。是的。所以这是故事的一部分。我认为有很多创新，我必须说。所以发展非常非常快。比如在德国，当你搜索“风车如何工作”时，对吧？我点进去。也需要一点时间。你总是看到等待，我们有点习以为常了。人们非常不耐烦。如果说有什么变化，那就是过去几年他们变得更加不耐烦了。

Me too. Yeah. So that's kind of the part of the story. I think there is a lot of innovation, I have to say. So it's really moving very very quickly. I like for example in Germany when you're searching for, I don't know, how do windmills work for example, right? And I go in here. Also takes a bit of time. You see like always waiting and we're kind of taking this for granted. People are very impatient. If anything, they become way more impatient over the last couple of years.

我几乎在想，这是否也是提取上下文的好地方。你知道，你几乎假设你有第一个轻量级提示，然后你能否在处理第一个提示时就已经获取信息，用于下一个提示？我打赌人们实际上完全愿意等更久。这让我想起 Gamma 的设计主管说过的话，他谈到生成演示文稿时，他们会提示用户处理主题设置，你可以使用主题控件。人们突然不关心加载时间了，因为有事可做。也许这就是获取额外上下文的地方。

I'm almost now wondering if this is the right place for some of that context extraction too. You know, like you almost assume that you have this first lightweight prompt and it's like can you already get information that you would feed into the next prompt while you're processing the first one? And I bet people would be totally fine actually waiting much longer. It reminds me of something that the Gamma head of design said and he talked about how when presentations were being generated, they then prompted people to work on the theming and you can like theming controls. People all of a sudden they didn't care about how long it was loading because it gave them something to do. Maybe that's the place to get some of that extra context.

我的意思是，在 Perplexity 的例子中，它只是要求更多上下文，对吧？所以这就像我们可以让人们等待，或者我们可以给他们一些事情做，以产生更有意义的输出，对吧？所以这对我来说完全合理。我真的很喜欢 Gemini 的一点是，他们有这个功能。在哪里？我想可能不是每个模型都有。哦不，他们确实有。

I mean in the case of Perplexity it's just asking for more context, right? As well. So this is like we can keep people waiting or we could give them something to do to create a more meaningful output, right? So that makes perfect sense to me. I mean one thing that I really liked about Gemini is that they have this one thing. Where is it? I think it was maybe not in every model. Oh no, they do.

他们有一个“双重检查回复”按钮。我一开始还在想，这是什么意思？如果你点击它，它基本上会遍历自己的输出，试图找到任何能支持当前检查内容的来源。如果有的话，我这里没看到高亮。通常，如果 Google 找到了与陈述稍微相似的内容，它会变成绿色；如果内容略有不同，就会这样高亮显示。通常会有高亮，但这里没有。我会尝试验证来源，确保它是对的。幻觉问题我们目前还无法解决，但如果你能对 GPT 和 Perplexity 说“去检查所有这些链接，确保它们存在”，也许这应该成为默认模式，尽管可能会多花点时间。这些都是非常简单的调整和改进，但我认为它们会随着时间累积。一旦你引入了格式选择、更多内容、结构化提示、更好的输出导航方式、结果分布而非仅仅摘要，这些东西真的会累积起来。我希望看到人们爱上 AI 产品。我没有看到人们爱上——我知道 Perplexity、Claude 和其他很多工具的用户非常喜欢它们，但我希望人们觉得“哇，这真是一次绝妙的 AI 体验”，或者只是体验，因为人们不会那样想。我希望他们真正得到帮助，找到很多价值，这样他们就不会在屏幕前浪费时间，去浏览输出、寻找复制文本、编辑、调整、修改、再问别的问题、添加更多上下文。这是一个故事。通常需要花费大量时间，而很多时候可能并不必要。

They have a double-check response button. For a moment I was thinking, what does it mean? If you click on it, it basically tries to go through its output and find any sources that actually back up whatever the check is doing. If there is anything, I don't see any highlight here. Typically, it becomes green if Google found content that's slightly similar to the statement, and something that's slightly different is highlighted this way. Usually it's highlighted, but here it doesn't. I try to verify sources to make sure it's actually right. Hallucinations is something we cannot fix yet, but if you could say to GPT and Perplexity, 'Go through all these links and just make sure they exist,' maybe it should be a default mode anyway, although it probably would take a bit more time. Those things are very simple adjustments and refinements, but I think they all compound over time. Once you bring selection of a format, more content, structured prompting, a better way to navigate the output, distribution of results rather than just a summary, those things really compound. I want to see AI products that people fall in love with. I don't see people falling in love with—I know that people working on Perplexity and Claude and many others absolutely love the tools, but I want people to feel 'Wow, that's an absolutely amazing AI experience,' or just experience, because people don't think about it that way. I want them to really be helped, to find a lot of value so they don't waste time in front of the screen navigating through the output, finding ways to copy, edit, tweak, bend the text, ask something else, add more context. It's a story. It usually takes an enormous amount of time, which is often maybe not necessary.

如果目标是维持对 AI 体验的信任，设计师还应该考虑其他什么吗？

Anything else that designers should be considering if the goal is to maintain this trust in an AI experience?

我认为获得信任的最佳方式是提供准确性。不幸的是，我们无法像对待软件那样绝对可靠地做到这一点。范围界定（Scoping）非常重要。范围界定与收集足够上下文是同一个概念。但我所说的范围界定，基本上是要确保人们明白他们身处何处。通常的问题是，他们可能提出了一个问题，但要获得信任感，他们需要知道答案来自哪里。通常是互联网，而互联网不一定可信。所以，如果你想激发信任并建立信任，我们需要展示来源。能够指出该查询所尊重的领域或范围也很有帮助。不一定非得是像这里这样的文件。可以是小过滤器，比如“拥有 20 年经验的专家”之类的。事实上，大多数时候什么都没有显示。我没有看到任何关于具体探索或研究了什么的参考。看起来几乎是随机的。也许确实如此。我不太确定某些 AI 引擎内部是如何运作的。但也许可以这样说：“这些不仅是答案的来源，更是更全局的范围。对于这个查询，我们考虑了来自可信专家的 279 个页面，他们具有这样的专业水平，并且在医疗保健领域至少拥有这个学位。”类似的东西会很有帮助。另一方面，我还想说，建立信任的最佳方式可能是向人们反馈他们已经理解了。做到这一点的方法是通过突出显示我们的假设或信念。

I would say the best way to get trust is to provide accuracy. Unfortunately, we just can't do that really absolutely reliably with AI like we do with software. Scoping would be very important there. Scoping is kind of the same idea as gathering enough context. But what I mean by scoping is basically we want to make sure that people understand where they are. Very often the problem is that they might be asking a question, but to get that notion of trust, they need to understand where the answer is coming from. Typically it's the internet, and the internet cannot be necessarily trusted. So if you want to elicit trust and build trust, we're going to need to show sources. It's also quite helpful to be able to indicate what is the domain or the scope that is respected for that query that the person is submitting. It could be not necessarily a file like it is over here. It could be little filters, like 'experts with 20 years of experience,' or anything like that. In fact, nothing is being displayed most of the time. I don't see any reference about what specifically was explored or studied. It seems almost random. Maybe it actually is. I'm not quite sure how everything is working under the hood in some of those AI engines. But maybe it would be a good idea to say, 'These are not just the sources of where it's coming from, but more global scope. For this query, we considered 279 pages from experts who seem to be credible and have this level of expertise and have at least this degree in healthcare.' Anything like that could be quite helpful. On the other hand, I would also say probably the best way to build trust is to reflect back to people that they understood. The way to do that is through highlighting what we assume or what we believe in.

举个例子，ChatGPT 以及现在其他所有 AI 都有记忆功能。如果你去定制你的 GPT，你可以告诉 AI 一些关于你自己的事情。我想是在偏好和设置里。所以你可以有记忆。你可以引用保存的记忆，也可以引用聊天历史。你还可以管理它。所以你可以说：“亲爱的 GPT，记住我喜欢胡萝卜。”你喜欢胡萝卜吗，Mike？

Just to give you an example, ChatGPT and everybody else as well at this point have memory. If you go to customize your GPT, there are some things that you can say to AI about yourself. I think it's in preferences and settings. So you can actually have memory. You can basically reference saved memories and also reference chat history. You can also manage that. So basically you can say, 'Dear GPT, remember that I like carrots.' Do you like carrots, Mike?

还行吧。

They're okay.

这不太令人兴奋。好吧，我们开始。“记住我喜欢胡萝卜。”然后它当然记住了。所以下次如果我问“给我一个晚餐食谱”，它可能会给你——看那里。我甚至没告诉 GPT 这么做。但它现在知道了。这很好。太棒了。但它也可以告诉我，也许这里可以有一个批次，比如素食、胡萝卜或其他我喜欢的东西，只是为了向我反馈“我了解你，这就是答案的来源”。不一定非得是文字。“哦，这是一个以胡萝卜为主的素食晚餐食谱，简单、美味、令人满足。”也许你不需要写所有这些。只写“素食”因为你是素食者，“胡萝卜”因为你喜欢胡萝卜，以及其他东西，仅仅作为一个指示器，一个信号表明它理解原因。它解释了为什么给出这个答案。我也喜欢这样，因为那样你就可以关掉“胡萝卜”，然后突然就有了一个一键操作。

That was not very exciting. All right, here we go. 'Remember that I like carrots.' And then of course it does. It remembers it. So next time if I ask something like 'Give me a recipe for dinner,' it might give you—look at that right there. I didn't even tell GPT to do that. But now it knows. That's nice. That's great. But it can also tell me, maybe it could be like a batch right here, a vegan, carrots or whatever else that I like, just to reflect back to me that 'I know you, this is where it's coming from.' It doesn't have to be like text. 'Oh, here is a vegan carrot-focused dinner recipe that is simple, flavorful, and satisfying.' Maybe you don't have to write all of that. Just write 'vegan' because you're vegan, 'carrot' because you like carrots, and whatever else, just as an indicator, as a signal that it understands why. It kind of explains why it's coming with this. I like that too because then you could turn off 'carrot' and all of a sudden you have a one-click affordance.

实际上，我觉得今晚不想吃胡萝卜。

Actually, I don't think I want to eat carrots tonight.

没错。所以你可以逐步构建，或者也许可以是一个下拉菜单。我们可以选择其他东西。但相反，我们只是在处理文本。现在如果你想说“好吧，我不喜欢——我今天不想吃胡萝卜。我确实喜欢胡萝卜。不要从我的记忆中删除胡萝卜。但我今天不想吃胡萝卜。”这就成了一个故事。你不能只是调整说“让它更蘑菇味”。我甚至不知道那是什么意思。“我想要更多蘑菇。让它更蘑菇味。”没办法。没有那样的按钮。然后它问我一个问题，我又得去打字。

Exactly. So you can kind of build it up, or maybe it could be a dropdown. We can select something else. But instead we're just dealing with text. And now if you want to say 'Okay, I don't like—I don't want carrots today. I do like carrots. Don't remove carrots from my memory. But I don't want to have carrots today.' That becomes a story. And you cannot just tweak and say, 'Make it more mushroomy.' I don't even know what that means. 'I want more mushrooms. Make it more mushroomy.' There is no way. There is no button for that. And now it asks me a question, and then I have to go again and type.

而且如果你真的想迭代，你可能会走上这条路，最终得到八个完整的回复垂直堆叠，突然之间变得非常难以解析。

And if you do want to iterate on it, you would probably go down this path that would lead you to eight full responses stacked vertically, which all of a sudden becomes very difficult to parse.

在理想情况下，我几乎想要这样：好，这是配料清单，我希望能够进行逐项交互，也许像复选框那样，对吧？

In a perfect world, I almost want like, okay, here's the ingredient list and I want to be able to perform like item level interactions and maybe we do this like a checkbox, right?

是的，完全正确。或者甚至探索一下。也许这是空间探索的完美用例，但我现在对基于画布的工作流非常感兴趣。这正是一个完美的例子：我会想，好，这看起来不错，但有一部分——也许我没有这些食材。好，现在让我单独拿出那部分，然后迭代：好，我能用什么替代？也许我想尝试三种不同的东西，然后选一个插回去，而不是一遍又一遍地得到重复的输出。

Yeah, exactly. Or like even exploring things. Maybe this is the perfect use case for spatial exploration, but I'm really really interested in canvas based workflows right now. And this would be a perfect example where I'm like, okay, this looks pretty good, but there's like this chunk. I don't maybe I don't have these ingredients. Okay, now let me just isolate that and iterate on like, okay, what could I do instead? And maybe I want to try three different things. and then this is the one and I insert it back in rather than having this like really repetitive output over and over again.

是的。所以这种精炼过程，我认为是最痛苦的，因为如果你需要调整，你得看看：好，这部分我需要什么、想要什么、不想要什么？然后你说“更像这样”“更像那样”——这输入方式太糟糕了，因为你得复制粘贴你喜欢的内容，然后说“我有这个，没有那个”，这纯粹是浪费时间。所以我们可以通过分解成几个主题来提供更交互的体验。当然，这会在之后发生，因为输出是逐 token 生成的。生成时你不知道接下来会是什么，也不一定知道最终结果。但之后你可以做后处理，比如“以这种形式呈现”，或者加个小按钮说“待办清单，从中生成待办清单”等等。但表格转列表也可以这样呈现。这些东西感觉就像“拜托，别这样”，只是些小细节，对吧？我认为它们累积起来会产生不同效果，让任务从几分钟变成 20 分钟，尤其是在做深度研究时。还有这种轮次之间的交互或导航。如果我再问别的问题，每次都要面对两堵文本墙。如果我发现这根本不是我要的，就回去。如果有三四次，那就成故事了。但这里不是这样，对吧？因为如果我追问一个问题，我只需把它加在这里。比如“专门针对冰岛的研究”。哦，顺便问候一下所有在冰岛观看我们的人。有 16 个来源，太好了。但现在我可以说：“嘿，这里有两个不同查询之间的导航。”

Yeah. So this kind of refinement journey I think that these are they are the most painful ones because if you need to tweak now well you need to kind of see okay what else what from this do I need or do I want or what do I not want and then you say maybe more like this maybe more like that and it's this is a horrible input more like this more like that because I mean you have to copy paste whatever it is that you like and say I have this I don't have that this is just waste of time so we could actually have a more interactive experience here right by breaking this down into some um topics. I mean, of course, it will be happening later because once the output is generated, it's generated token by token. So, at this point when you're generating it, you don't know what's coming up and you don't know maybe necessarily what it's going to be like. But then you can actually do this sort of post-processing potentially and say uh presented as this or maybe you could have this little button here saying to-do list, make a to-do list out of that or whatever, right? But then when it comes to a table to list, it could be presented this way. Those things really feel like okay, give me a break. These are just little nicities, right? I think they compound and make a different compound and make a difference between people just spending a few minutes on a task and people spending 20 minutes on a task that can make a difference especially when you do some sort of profound research. And again also this kind of interaction or navigation between these turns. So if I ask something else now I have two walls of text I need to go through every time. And if I find that, okay, this is not what I want at all. So, let me go back. We have like three or four of them, that becomes a story. But not here, right? Because if I go and I ask a follow-up question, let me just add it in here. Something like um specifically related to research in Iceland. Okay, greetings to everybody who's watching us from Iceland by by the way. Oh, there is 16 sources. Excellent. Right. But now I can say, "Hey, here is a navigation between these two different queries."

对。

Right.

所以如果我觉得不相关，我还可以删除一个。对，让我回到这里。

So I can also delete one if I don't find it relevant to me. Right. Let me go back here.

哇，这么一个小细节。

Wow. Such a little detail.

我喜欢。对。

I love it. Right.

真的非常棒。而且这个删除选项也超级有用，这样你就能只得到你需要的。

It's really, really nice. And then again, this option to delete is super helpful, too. So you just get what you need.

是的。因为如果你没做对，或者得到一堵文本墙，它不像你想象的那么相关，现在就成了记录中永久的一部分。这让我抓狂。

Yeah. Cuz if you don't do something right or you get this wall of text that isn't as relevant as you thought it was going to be, it's now this permanent part of the record. drives me crazy.

哦，你得重新开始整个对话，差不多问同样的问题。这些都是真正累积起来的小事。嗯，也许有一件事我想展示一下，可以吗？好的，请。

Oh, you have to restart the conversation all over again. I kind of asking the same thing. These are all just really kind of small thing that do add up. Um maybe one thing you one one thing I would like to show if that's okay. Yeah, please.

这是 ExoA。Exa 也很酷，我想你会看到不少工具这么做，因为有很多不同的交互模式。我们刚才只聊了聊天，对吧？但当然还有语音，语音通常很难处理。还有一种方式可以稍微不同地呈现数据，比如数据表格，但几乎像数据网格。这里有网站，你可以查找东西，它会给你这个列表，对吧？我想展示这个是因为它实际做的事情非常有趣。比如我输入“西雅图 AI 初创公司的创始工程师，获取技术实力和资历”。我喜欢的是，它也会把提示分解成几乎像 UI 控件一样的东西。它说：好，我要把它分解成主题。首先，我需要找到开发 AI 产品和服务的公司的创始工程师。这差不多。公司被归类为 AI 初创公司。这差不多。人在西雅图。看起来也对。我还可以添加其他条件，比如“工程师”“创始工程师”，对吧？我有什么偏好吗？我不知道该加什么条件。

So this is ExoA. Exa is also pretty cool and I think you can see quite a few tools doing that because there is a lot of different kind of interaction modes. I mean we spoke just about chat, right? But of course there is also voice and voice is usually very difficult to deal with in general. But there is also a way to present data in a slightly different way like a data table but maybe almost like a data grid. So there is websites here and websites here you can actually look for things and it kind of gives you this list right? The reason why I wanted to show this is because it's actually really interesting of what it actually does. So if I took let's say something like founding engineers at AI startups based in Seattle get me the technical strength and seniority. What I love about this is that it also takes the prompt and then it breaks it down into kind of almost like UI controls. So it says, okay, I'm going to break it into themes. So first of all, I need to find founding engineers at the company developing AI products and services. That's about right. Uh company is classified as an AI startup. It's about right. Person is based in Seattle. Seems to be right. And I can also add some other criteria here and say engineer founding engineers, right? Any preference that I have? I have no idea what criteria to add here.

精通 React。哦，有了。精通 React，对吧？有了。对。我还可以排除一些东西，这又是其他 AI 做不到的。所以这基本上就是字面意义上的筛选，对吧？你可以说，哦，等一下。根据我在找什么，它给我一种添加丰富信息的方式。那么，你还想从他们那里了解什么？只是技术实力？也许工作年限、毕业日期，任何你觉得相关的，以及你想要多少结果。所以它基本上创建了一个表格，提取了所有这些信息，因为我告诉它我需要 25 个结果，它就会尽量找尽可能多的来源来给我 25 个结果。所以我会得到一个包含 25 个结果的电子表格，它们都可以根据我觉得相关的每个数据项进行筛选。所以我可以比较，也可以排序。

Proficient in React. Oh, here we go. Proficient in React, right? Here we go. Right. I can also exclude some things which again no AI allows you to do. So, it's kind of like literally filtering, right? You can say, oh, hold on for a moment. Depending on what I'm looking for, gives me a way to add enrichments. So, what else do you want to know from them? Is it just tech strength? maybe years of experience, graduation date, anything else that you find relevant and how many results do you want. So it basically creates a table with all this information pulled out and because I told it that I need 25 results, it will actually try to find as many sources as it can to give me 25 results. So I will get a spreadsheet with 25 results and they all can be filtered with all the data for each of those things that I find relevant for me. So I can compare it and I can sort by it.

为什么我们就是没有选项来排序和筛选从 AI 得到的结果？这不是最显而易见的事情吗？我觉得我可能是最暴躁的人了。对。但那些研究——

Why on earth don't we have an option to sort and filter results that we're getting from AI. Isn't it the most obvious thing ever? And I mean I I'm coming across as probably the most grumpiest person ever. Right. But those research

是的。

Yeah.

但这些都是真正能极大改善体验的小事。极大改善。而且这些是我们在软件中习以为常的功能，但出于某种原因，它们从未真正进入 AI。我不太确定为什么。我非常相信视频在解释设计师思路方面的力量。所以需要反馈时，我会在 Slack 里丢一个 Loom 链接，再丢一个 Figma 原型链接，然后反馈就散落得到处都是。我的意思是，一团糟。所以我正在构建我一直希望存在的产品，它叫 Inflight。你可以把它想象成一个异步评审工具。

But those are little little things really that can tremendously improve experience. Tremendously. And they are things that we are used to in software, but for some reason they never really make it to AI. I'm not quite sure why. I'm a big believer in the power of video to explain my thinking as a designer. So when it's time to give feedback, I'll drop a Loom link in Slack and another link to a Figma prototype and then feedback will be scattered everywhere. And I mean, it's a mess. So I'm building the product that I've always wanted to exist and it's called Inflight. You can kind of think of it like an Async Crit.

这是一个简单的方法，可以分享视频演示以及交互式原型或你正在设计的任何东西，然后 AI 会采访你团队中的人，获取你需要的反馈，并将所有内容整理到一个漂亮的洞察页面中。目前我只向 Dive Club 的听众开放访问权限。所以如果你想成为第一批使用 Inflight 的人，请访问 dive.comclub/inflight 来抢占名额。在你走之前，我还有几个问题。我想稍微把视角拉远一点。作为从事涉及大量 AI 的产品的人，我经历过另一个设计挑战：你有一个光谱，一端是更隐蔽的 AI，所有事情都在幕后发生；另一端则是到处贴满闪光图标。基本上每个设计师都必须弄清楚自己在这个光谱上的位置。你对此有什么建议？你是如何思考这个挑战的？

It's an easy way to share a video walkthrough along with an interactive prototype or whatever you're designing and then AI interviews the people on your team to get you the feedback that you need and organizes everything for you in a beautiful insights page. So right now I'm only giving access to Dive Club listeners. So if you want to be one of the first to use inflight, head to dive.comclub/inflight to claim your spot. Couple more questions maybe before I let you go. And I kind of want to zoom out a little bit. There's another design challenge that I've experienced as someone working on a product that involves a lot of AI where you kind of have this spectrum. On one end, you have the more disguised AI and everything's happening behind the scenes and on the other, you know, you're just slapping sparkle icons everywhere. There's like this spectrum and basically every designer has to figure out where do I fit onto that spectrum? What advice do you have for that person? How are you thinking through that challenge?

我不知道为什么，但这也是我观察到的一个现象，我们称之为“安静 AI”与“可见 AI”。有一些工具，比如 Dovetail。Dovetail 是一个面向研究人员的工具，可以让你录制会话、发现洞察、创建报告，以及研究人员需要做的所有好事。让我非常惊讶的是，那里没有闪光图标。

I don't know why but this is something that I also observe and this is something that we call quiet AI versus visible AI. So there are a few tools like for example Dovetail. Dovetail is a tool for researchers which allows you to record sessions and find insights and create reports and all the good stuff that researchers need to do. What was really surprising to me is that there is no sparkles there.

没有。

No.

基本上，你要做的是仔细审视用户旅程，看看人们需要做什么。他们需要能够录制会话，需要做招募。好吧，也许 AI 可以帮忙，对吧？他们需要找到一些洞察。好吧，也许 AI 可以帮忙。也许你还需要编辑人们提供的敏感信息。嗯，AI 可以帮忙。所以你基本上看看现有的旅程，然后在其中撒上一点 AI，要么减少挫败感，要么改善或加速成功，对吧？这就是安静的方式，确实像你说的那样，在幕后发生。但另一方面，我们也有这些体验，感觉几乎是 AI 优先，我一般对 AI 优先有点过敏，因为我觉得这就像说 JavaScript 优先，或者我不知道，船优先、容器优先，对吧？或者任何东西，对吧？技术是用来服务并帮助人们达成目标的。所以，如果有的话，我们不应该那么痴迷于 AI，而应该绝对痴迷于人类，因为最终使用技术的是他们，对吧？所以我们真的需要做很多研究，理解他们用 AI 做什么，而不是我们用 AI 做什么。我的意思是，我们可以用 AI 做事情。我们是技术人员，如果你愿意的话，对吧？但我们需要痴迷于人们如何使用那个东西，对吧？所以对我来说，真正重要的不是说 AI 优先，而是也许 AI 第二，不是 AI 最后，但我们需要思考 AI 在我们交付的产品中擅长什么，对吧？它带来什么价值，然后我们看到人们在获取该价值时在哪里挣扎，以及我们实际上可以在哪里进一步提升该价值。如果有一个特定领域，人们浪费大量时间，来回折腾，也许他们不一定知道自己想要什么。所以，我们需要引导他们到某个地方，对吧？嗯，也许 AI 可以在那里帮忙，这很好，但它不是那个东西。它不是价值生成器。它更像是通往 AI 提供的价值的路径。它有一个不可思议的机会。我们只是有一个还没有真正好好解析的东西，对吧？

Basically, what you have is you take a close look at the user journey and you look at what people need to do. Well, they need to be able to record sessions. They need to do recruiting. Okay, maybe AI can help with that, right? They need to find some insights. Okay, so maybe AI can help with that. Maybe you also need to redact some sensitive information that people are providing. Well, AI can help with that. So you basically take a look at the existing journey then you sprinkle a bit of AI all across it to either reduce frustrations or improve or speed up successes, right? And so that's the quiet it's kind of all indeed as you saying like kind of happening under the hood. But on the other hand we have these experiences where it's kind of almost feels like AI first, and I'm a little bit allergic to AI first in general because I feel like it's like saying JavaScript first or I don't know ships first and containers first, right? Or anything, right? Technology is here to serve and to help people get somewhere. So, if anything, we should not be that obsessed about the AI, but to be absolutely obsessed about humans because they are the ones who are kind of using that technology in the end, right? So we really need to do a lot of research, understand what do they do with AI, not what we do with AI. I mean, we can do things with AI. I mean, we are technologists if you like, right? But we need to be obsessed about how people use that thing, right? And so for me the really important thing is not to say AI first but maybe AI second, not AI last, but we need to think about what is it that AI is good for in our product that we're delivering, right? So what value does it bring and then we see where people struggle in getting that value and where we can actually boost that value even further. If there is a certain area where people lose a lot of time, they waste a lot of time going back and forth and maybe they don't know what they want necessarily. So, we need to guide them somewhere, right? Well, maybe AI can help there and that's great, but it's not the thing. It's not the value generator. It's sort of a path to the value that AI provides. It has an incredible opportunity. We just have a thing that we haven't really unpacked properly yet, right?

所以，就个人而言，我会说我会看用户需求。我会看人们在哪些地方挣扎，然后用一点 AI 来提升，看看我们是否真的能在那里帮助人们。但在这段旅程中，我认为重要的是我们允许人们减少我们之前谈到的交互成本，比如表达他们想要什么等等，并且在人们与输出作斗争并想要改进时帮助他们，因为 AI 不是免费的。它感觉像魔法，可以为我们做一切。但它伴随着巨大的成本。我的意思是，有很多不同的层面，从间接成本和修复错误的成本，到算力成本和能源成本。就像没有技术是免费的一样。我们只需要意识到这一点。所以当有人提交一个提示时，我希望那个提示几乎完美，几乎理想。这样我就知道所有需要知道的东西，才能真正生成有意义的东西。否则，那就是浪费时间、精力和一切。

And so, personally, I would say I would look at user needs. I would look at where people struggle and then boost it all up with a bit of AI to see that we can maybe really help people there. But also on that journey, I think it's important that we allow people to reduce that interaction cost that we were speaking about before, like articulating what it is that they want and so on, and also helping them in the cases where they kind of are struggling with the output and they want to refine it because AI doesn't come for free. It feels like it's magical and can do everything for us. It comes with a tremendous cost. I mean there are many many different layers from indirection cost and fixing errors cost to compute cost to energy cost. It's like no technology comes for free. We just need to be aware of that. And so when somebody submits a prompt, I want that prompt to be almost perfect, almost ideal. So I know everything that I need to know to really generate something meaningful. Otherwise, it's just waste of time and energy and everything in between.

就个人而言，我不是想刁难，但我想我就是这样，我可能会选择安静 AI 的方向，而不是喧闹的 AI。还有一件非常有趣的事，我记得 Norman Nielsen Group 做过一项研究，他们发现实际上当人们看到 AI 作为标签或徽章出现在某处时，这不一定是一件好事，因为人们不是在寻找 AI 功能，他们在寻找能用的功能。它们可能恰好是 AI 或者不是，对吧？但这不一定是客户真正欣赏的东西，比如“哦，他们现在有 AI 功能了。那个 AI 功能在哪里？”对吧？它甚至可能产生相反的效果，因为它是 AI，所以可能不可信，可能产生幻觉，可能做这个做那个。所以，我总是会做一个测试，说“好吧，这是一个 AI 功能或 AI 驱动的，而这只是一个功能或产品”，对吧？然后看看哪个效果更好。如果完全没有区别，我也不会感到惊讶。

Personally, and I'm not trying to be difficult but I guess I am, I would probably go with a quiet AI direction rather than loud AI. Also which is very funny, there was a research done I think by Norman Nielsen Group and what they discovered is that actually when people see AI as a label or badge somewhere, it's not necessarily a good thing because people are not looking for AI features, they're looking for features that work. They might happen to be AI or not, right? But it's not necessarily something that customers really appreciate like, oh, this they have AI feature now. Where is that AI feature? Right? It can even have this opposite effect where because it's AI, it's maybe untrustworthy, maybe it's hallucinating, maybe it does that and does that. So, I would always run a test to say, okay, this is an AI feature or AI powered and this is just a feature or a product, right? And see what works better. I would not be surprised if there was no difference at all.

我自己已经能感觉到了，我点击 AI 功能时就会假设会被追加销售，因为这种情况发生了太多次。我看到小闪光图标，心想“哦，这很有趣。”然后一点击，就提示你必须升级到下一个套餐。我就想，“天哪，我受够了。”

I can already feel it for myself where I just assume that I'm going to be upsold when I click on an AI feature and because it happens so many times I see little sparkle I'm like, "Oh, that's interesting." They click and it's like you have to upgrade to our next plan. I'm like, "Man, I'm so tired of this."

是的。但你知道吗？这很有趣，Mike，因为我认为这是很多人遇到的问题，因为他们意识到，即使是你正在使用的任何工具，人们很快就会达到积分的上限，比我们想象的要快得多，因为他们想，“哦，我只是想再进一步，再多玩一会儿，再多玩一会儿，再多实验一会儿”，对吧？然后不知不觉，你的积分就用完了。我的意思是，它可能负担得起，但也可能是一个非常昂贵的工具。

Yeah. But you know what? This is interesting, Mike, because I think that this is an issue that a lot of people have because they realize that even, you know, whatever tool you're using. People are running at the limits of the credits very quickly, much faster than we think, because they think like, oh, I'm just going to go further and just play a bit more and play a bit more, experiment a bit more, right? And before you know it, you're running out of credits and then I mean, it can be affordable, but it can be very expensive tool.

不是每个人都愿意为 Pro 之类的计划每月支付 200 美元，即使它能产生令人难以置信的深度研究等等，对吧，那很贵，我的意思是那真的很贵。所以当你遇到某个层级的限制时，这就是终点，旅程结束了。我认为很多公司现在意识到，我们需要让用户在现有的积分计划内获得一些成功，但这并不意味着给他们一个自由格式的文本框，让他们为所欲为。

Not everybody wants to pay like for pro of like plan $200 a month or so even if produces incredible deep research and whatever right that's expensive I mean that's really expensive right and so when you're hitting limitations of a tier this is it this is the end of the journey and the goal is I think at a lot of companies are realizing that now we need to get people to a to some success within the credit plan that they have and that does not mean give them the free form text box to do whatever they want

因为那样他们会非常快地达到那个限制。我们需要引导他们把提示词写得更小、更准确、更简洁。所以他们不应该写更多提示，而应该写更少但更好的提示，对吧？所以我们需要在用户达到免费层级的限制之前，让他们在第一次会话中取得一些成功。但如果他们一直继续、继续、继续、继续，那对公司来说可能非常昂贵，因为他们消耗了太多积分，或者他们会太快遇到限制，然后无论哪种方式都没有价值。

because then they hit that limit very very quickly. We need to guide them to make that prompt small, right, and accurate and concise. So they don't prompt more, they should prompt less but better, right? So we need to give them some success in the first session before they hit the limits of the free tier. But if they just keep going, keep going, keep going, keep going, it might be either very expensive for the company because they're just using too many credits, right? or they will be hitting the limitations too fast and then there is no value either way.

在你走之前，最后一个问题，我想稍微展望一下未来，深入你的想法。你显然在思考所有这些不同的 UX 原则，它们如何演变，以及我们看到的这些涌现行为，你比我所知的几乎任何人都想得多。那么，当你展望这一切的走向时，也许是更智能体式的行为，也许只是与 AI 交互的不同方式。你现在关注哪些事情，特别期待看到它们展开？

Before I let you go, one final question and I kind of want to just look a little bit further into the future and get inside your brain a bit. You're obviously thinking about all these different UX principles and how they evolve and these emergent behaviors that we're seeing as much as basically anybody that I know. So when you kind of look in terms of where this is all headed, maybe it's more agentic behaviors, maybe it's just different ways of interfacing with AI. What are some of the things that you have your eye on right now that you're particularly excited to see unfold?

是的，我认为人们对智能体有很多兴奋，但至少在目前实践中，这个领域的可靠性非常低，也因为智能体总是带有一些你需要建立的护栏，这意味着权限，意味着你需要某种审批层等等，我的意思是，没有什么是免费的，对吧。

Yes, I think that there's a lot of excitement about agents but in practice as of now at least uh there is very little reliability uh in that space also because agents always come with some guardrails that you need to establish that means permissions that means you need to have some sort of approval layers and things like that it's I mean nothing comes for free right yeah

首先把它组合起来就相当有挑战性。

it's can be quite challenging to put together in the first place

但我真正相信的是，如果有人在 20 年后看到这个，他们会想，他们当时到底在想什么？我认为我们可能不会再看到很多提示工程指南了。我认为最终它只会成为 UI 的一部分。我认为在这一点上，所有的小事情都会变得自然。就像有些人说的，我想可能是设计复杂性负责人说过，在很多方面，AI 可以被视为一种花哨的新东西，就像现在的自动补全一样，到处都有一点，这里一点，那里一点，但并没有真正被宣传为重大新事物。至少对我来说，一个重要的转变是，我们可能正在走向一个世界，我们在很多方面变得更少战术性、更多战略性。这意味着，你知道，没有一天过去，互联网边缘的某个人不会告诉我们，我们将被 AI 取代，如果不是 AI，那么了解、理解和掌握 AI 的人将取代我们。我认为我不同意，因为我认为人类带来了巨大的价值，不仅是在批判性思维、情商等方面，这不用说，但对我来说，必须有人对这一切的走向有愿景。我不相信，也许现在我只是天真，但我不认为产品会由 AI 设计和开发来以最好的方式服务人们。我的意思是，能够创造这种联系，几乎像与产品强烈的情感联系，但我必须看到有人在上面工作，他们花了很多心思在我的体验上，我不确定是否有任何 AI 能够超越人类带来的对细节的关注。最后，我认为我们人类，对吧？正在远离我们正在做的工作。是的，它会改变，而且已经改变了，还会进一步转变。我对此毫无疑问。但这只意味着我们正在做的事情会不同。我们将以某种方式编排那些 AI 体验。也许用智能体，也许用拥有子智能体的智能体，等等，对吧？但必须有人引导、编排并为人们创造这种体验，因为 AI 非常擅长创造体验。但我不确定它是否在为人们创造体验。我希望看到人们来到一个网站或一个 AI 产品，说这太棒了，我绝对喜欢它，我想每天都用它，他们对此充满热情，他们真的非常想用它，因为他们爱上了这个界面，因为它理解人们的需求。它理解人们关心什么，它知道什么时候以某种方式说话，什么时候不，什么时候调整特定的语气和声音，也许 AI 可以做到。我并不是想成为那些说不要 AI、忘记 AI、它不重要的人之一。你知道，我们是人，对吧？但我认为我们带来了巨大的价值。我们不应该忘记这一点。我期望的是，我们需要看到 AI 非常擅长很多事情，人类也非常擅长很多事情。我想看到的是美妙的人类优先体验，其中恰好包含一些 AI 组件。这就是我想看到的，对吧？两者结合的地方。

but what I really believe in and I think That's, you know, if somebody's watching this 20 years, they'll be thinking, what the hell were they thinking back then? I think that we will probably not end up seeing a lot of prompt engineering guides anymore. I think that in the end it will be just a part of UI. I think that at this point like all the little things will be just natural. We're just doing like some people say I think it maybe was a head of design complexity who said that in many ways AI could be perceived very much like fancy new thing that would be very much like autocomplete now just everywhere a little bit of that a little bit of this a little bit of AI here a little bit of that here but not really properly advertised as the big new thing one significant shift at least to me is that we're probably moving to the world where we are becoming more less tactical and more strategic in many ways that means means that you know not a day passes by without somebody on the fringes of the internet telling us that we're going to be replaced by AI and if not AI that's going to replace it then people who know and understand and get AI will replace us and I think I disagree because I think that there is an enormous value that humans bring to the table not only in terms of like critical thinking emotional intelligence and so on that goes without saying but to me there must be somebody who has a vision about where the hell it's going I don't believe maybe for now maybe I'm just naive but I don't see products being designed and developed by AI to serve people at the best way possible. I mean to be able to create this connection almost like a strong emotional connection with the product but I must see people working on it who put a lot of thought into what my experience is like and I'm not sure if any AI can outcare and outlaw this attention to details that people bring to the table. In the end, what I think is kind of us humans, right? Moving away from the work that we're doing. Yes, it will change and it has already changed and it will be shifting further. I'm not I have no questions about that. But that only means that what we're doing will be different. We'll be just orchestrating those AI experiences in some way the other. Maybe with agents, maybe with agents who have sub agents and whatever, right? But somebody must be guiding and orchestrating and kind of creating this experience for people because AI is very good at creating experiences. But I'm not sure if it's creating experiences for people. I want to see people who kind of come to a website who a AI product and say this is amazing. I absolutely love it. I want to use it every day that they feel like extremely passionate about it and they really really want to use it because they fell in love with this interface because of how it understands people's needs. It understands what people care about. It understands when to say things in certain way and what not. When to adapt a particular tone and voice and maybe AI can do that. And I'm not trying to be like one of those people saying no AI, forget about AI. It's not important. It's uh you know we are the people, right? But I think that there is an enormous amount of value that we bring to the table. We just shouldn't forget that. And what I'm expecting is that we need to see that there are a lot of things that AI is very good at. There are a lot of things that humans are very good at. What I want to see are wonderful human first experiences that happen to have some AI components in them. That's what I want to see, right? Where you kind of have the marriage of both.

是的。嗯，我完全同意，非常感谢你来做客，并且把事情说得非常具体。我的意思是，看到你如何解读这个领域，你在关注什么，你在研究什么，这真的很有趣。所以非常感谢你花时间分享。

Yeah. Well, I I couldn't agree more and really really appreciate you coming on and getting super specific about things too. I mean, this was just fun to see how you are interpreting the space and and what you're paying attention to and what you're studying. So really really grateful you took time to share.

哦不。非常感谢你邀请我。我的意思是，我明天还可以来。嗯，就是明天。完全没问题。我认为对我来说，关键是找到那些小细节有巨大的价值，因为我认为它们真的可以成就或破坏一个界面或体验。

Oh no. Thank you so much for having me. I mean I can be here tomorrow. Uh and that's tomorrow. That's no problem at all. I think that's the the main point for me is that it's there is enormous amount of value in finding those little details because I think that they can really break a make an interface or make an experience.

就像当这些东西真正结合在一起时，它们确实会复合，然后你就会有一种非常不同的体验。我的意思是，我认为我第一个爱上的 AI 产品实际上是这个，Consensus。即使我在那里没有找到任何有意义的东西，我可能也会用它。它只是一个很棒的 AI 体验。我希望看到更多这样的产品。

It's like when those things really come together, they really do compound and then you have a very different experience. I mean, I think like the very first AI product that I fell in love with is actually this one, which is consensus. I probably would use it even if I didn't find anything meaningful for me there. It's just a great AI experiences. I would love to see more of that.

我的意思是，我希望这期节目能为大家提供一个更精细的视角，让设计师可以用它来解读他们正在体验的 AI 产品，同时也能看到那些微小的细节和机会，从而让这些产品真正变得可用，而不是仅仅停留在利用新技术能力的浪潮上。

I mean, I hope that's what this episode does for people is even gives a little bit of a finer lens that designers can use to interpret some of the AI experience that they're playing with and also to see the tiny subtle details and opportunities that they can take advantage of to make these products something that are truly usable and not still just riding the wave of these new technological capabilities.

让我们拭目以待吧。我想也许 20 年后你再看这个节目时会想，嗯，他们那时候还有界面呢。

Let's wait and see. I guess maybe once you watch this like 20 years later or so thinking, well, they had interfaces back then.

我喜欢这个说法。这很有趣。谢谢你分享。

I love it. This has been fun. Thanks for telling.

非常感谢你邀请我。

Thank you so much for having me.

在你走之前，我想花一分钟介绍一下我最喜欢的产品，因为我经常被问到我的技术栈是什么。Framer 是我用来建网站的工具。Genway 是我做研究用的。Granola 是我在评审时做笔记用的。Jitter 是我用来给设计做动画的。Lovable 是我用代码实现想法用的。Mobin 是我寻找设计灵感的地方。Paper 是我像创意人员一样设计用的。而 Raycast 是我每一步的快捷方式。我精心挑选了这些公司，这样我才能全职做这些节目。所以，支持这个节目的首要方式就是去了解它们。你可以在 dive.comclub/partners 找到完整列表。

Before I let you go, I want to take just one minute to run you through my favorite products because I'm constantly asked what's in my stack. Framer is how I build websites. Genway is how I do research. Granola is how I take notes during crit. Jitter is how I animate my designs. Lovable is how I build my ideas in code. Mobin is how I find design inspiration. Paper is how I design like a creative. And Raycast is my shortcut every step of the way. Now, I've hand selected these companies so that I can do these episodes full-time. So, by far the number one way to support the show is to check them out. You can find the full list at dive.comclub/partners.