AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
This interview has been lightly edited for length and clarity. Nick Statt: Hello and welcome to Decoder, Nilay’s show about big ideas and other problems. This is Nick Statt, senior producer. And I’m joined by our brand-new supervising producer, Greg Ott. Greg Ott: Good day, everyone. And Hi, Nilay. Nilay is here too. He is the person who hosts the show, and his name is also in the show. So it makes sense. Nilay Patel: It’s true. We don’t consistently say my name in the show enough. We should do it all the time. I should do it. Welcome to Decoder with Nilay Patel. GO: Like Nick was saying, this is Nilay’s show and we are doing a mailbag episode. This is where we go through all the feedback. Because you listen to the end of every episode, we know you do. We do mention that we go through every single email that we get. Nick, you can verify that. We get a ton of email, and we do read through all of it. Is that true? NS: We now get YouTube comments. We get many, many Verge comments. We get Spotify comments and posts on Bluesky, Threads, and elsewhere. So yeah, we get a lot of feedback. I do feel like promising we’ll read all the YouTube comments is an emotional commitment that we’re making, but we do read everything. GO: We do. NS: I do read the Spotify comments. I read every one because those are specifically pretty pointed. GO: Yeah. They’re kind of like uncut gems. Because they’re trying to build this little community, somebody might read this because it only gets like six comments as opposed to 600 for a YouTube video. NS: We last did a mailbag episode in April, and we said we were going to do more of these. So we’re back again. This is September. We’ve had a few months. We’ve had a lot of really big interviews. So yeah, let’s just get into it. GO: We’re going to kick things off with the “software brain” video you put out. This was one of the biggest things that people have commented on. It is far and away the most talked-about episode in the past six months. Hundreds of comments on The Verge website, hundreds of comments on YouTube, single to double-digit comments on Spotify, and lots and lots of emails. We want to start there because a lot of the feedback was positive and people actually rarely email us a ton of negative feedback. We get a lot of, “Attaboy.” Some guy I knew in high school just emailed me — Hi Dean — saying how much he liked our Kathy Hochul episode. So people like the show, but for this one, we got some negative stuff on The Verge that we’re going to get to in a second. One thing we wanted to talk about though is something a reader brought up about the idea of the Nilay rants. And of you effectively pulling out your inner Dennis Miller and really giving us your take on something. For this one, one listener wrote, “Just listened to ‘People do not yearn for automation,’ and all I can say is the people do yearn for more Nilay rants.” So we want to know, what is the origin of this type of rant? And then how much of the show’s DNA do you think is built around this idea of you giving us the rant, and the idea of you giving us your hot take on something really, really well thought out, as opposed to the interviews we usually do? I feel like I should make another riff on the “People do not yearn for automation,” but as Dennis Miller — just full rant. GO: We’re going to have to pull out some really good old references. The power of AI means Dennis Miller can deliver software right now. GO: “Let me tell you something, babe. Sam Altman, the last time he was sitting in a Whataburger…” Right. We could just do it. It’d be great. I’m sure we wouldn’t get sued for that. They’re just essays. I know they’re rants and they’re delivered on YouTube and they’re very pointed in particular ways. But that’s just what I would’ve written as a column 10 years ago before we had podcasts this way. It has been very obvious to everyone that you can’t just write a column and expect to have an impact. You have to stare at the camera and do it. We’re very good at publishing that as a column and then having a transcript and polishing it and delivering it all the other ways. That’s the origin of it: For all the podcasting I do and as little writing as I do anymore, in my heart, I’m still a writer. Taking all of the reporting and all of the ideas that we hear and all the conversations we have and synthesizing them into ideas that you can grapple with is the point. It is really the point of running a magazine. That’s where that comes from. Now, can I do it as often as I want to do it? I obviously can’t. The reason I did “software brain” at that moment is that the excitement around AI is so high. The hype is so high and it’s all connected to how good AI systems are at writing software and manipulating databases. But the reason that I thought, “I got to make the time to do ‘software brain’ right now,” is that I could see the hype cycle around AI failing to contend with the lack of verifiability in every domain except software. That’s getting a little blurry now. Lots of companies are just databases. Software engineering is verifiable, you can run it through the compiler and see if it works. AI has obviously gotten really good at making software. If you believe the world is just software, that means AI has gotten good, it’s done. You can see the labs are talking about how they’ve achieved AGI or they will very soon and it is very much connected to the ability of AI to write software. I’m just not discounting that that’s important. I’m just saying once you get outside of software, verifiability gets way harder. You can see it in the episode we did with Verge reporter Robert Hart about AI and math. There’s an amount of verifiability in math and you can burn a lot of tokens, we don’t know how many, and solve some unsolved math problems. There’s probably verifiability in various other science domains, so it’s really interesting. Is there verifiability in generating novel drugs? No, you got to inject a bunch of people with novel drugs. What are we going to do? GO: That’s why they got so many rats in cages somewhere, it has to be done. Right. There’s just some other thing that’s going to happen in the next 10 domains that is going to be really, really challenging. That’s where I wanted to put the marker down and say, “I have been swimming in this water for a long time with these people and the power of software is not to be denied, it has remade our world. But if that’s the only framework or you think that is a universal framework, you’re going to run into these walls.” There are lots more of those we can do, but I want them to hit, I only want to write hit singles. Importantly, they’re all born of the reporting. That’s always been the dynamic on The Verge. Certainly the dynamic on Decoder is the data can get to your head. The episodes where I just talk sometimes outperform the episodes with CEOs. That’s not good. That’s a bad incentive loop for me because I want everything to come out of the fact that I’ve talked to all of the people and you can go check my work. You can go see what those conversations are like and you can see that they are genuine and honest in all the things they should be. That’s where I feel like I not only have the depth and the foundation to write something like “software brain,” but I can send it to those people and say, “This is how I feel about it.” They come on the show and they talk to me about Google Zero or “software brain” and all that stuff, and that conversation gets much bigger than me just ranting. NS: My reaction to that, Nilay, is that you feel like your thesis for “software brain” has borne out over the last few months. We did get feedback from people who disagreed with some of the premises. It wasn’t that they’re saying people aren’t unhappy with AI, but they think that we’re wrong about why they’re unhappy with AI. One reader, Josh Vanderberg, wrote in to say that he thinks it’s not that people hate AI as a technology, but rather that they don’t like their current economic situation and they’re projecting that onto AI. He’s pretty sure that we’ll eventually embrace natural language as the superior way of doing things on a computer. He added, “All modern computer interfaces suck hard. We’re forced to map our fluid non-deterministic creative selves into hierarchical menu systems. No matter how intuitive and discoverable, we constantly fail to find the tools that we need.” He uses Canva’s CEO as an example of somebody that deeply understands the problem and they realize that AI can provide the user interface we’ve all been yearning for. He goes on to say, “People definitely yearn to just tell Photoshop what they want done in words and forget everything they ever learned about that horrible user interface.” We had a few other similar comments like, “People do not yearn for automation, they yearn for cheap stuff. And when AI gives them cheap stuff, they’ll be happy.” A third person wrote, “’Software brain’ misses three things regular people actually care about: control, agency, and meaning. We like doing things. The doing is where the meaning comes from.” I’m curious just from those particular comments and also the other feedback you got from “software brain,” and how you respond to those criticisms? Let’s take them in turn. This is a very thoughtful comment, Josh. I don’t disagree that all modern computer interfaces suck hard. I’ve used Liquid Glass. I agree with you. I refuse to upgrade this computer to macOS Tahoe. I’m just not doing it. There’s some confusing thing happening in user interfaces, probably related to how much all of our devices have turned into shopping malls or SaaS products or whatever it is. There’s something broken there and I agree that that’s bad. I disagree that natural language is a perfect substitute for everything or can become a universal interface. I say this as somebody who is writing a Decoder book based on all the interviews on the show. I’ve used Wispr Flow to get as much out of my head onto the page as I can. That is incredibly powerful and you can see why people like it. I also have to edit every single word it generates because it insists on formatting my thoughts in ways that I find completely bananas. All these systems say, “You think in bullets.” I say, “I am not thinking in bullets right now. I’m just pausing between sentences.” There’s something there where you can take the natural language input and what we as people are reasonably good at doing is taking everything else with it, the context of our body language, the tone of our voice, and creating other meaning out of it. Computers are really bad at that. Maybe they’re going to get better at it, but they’re currently very, very bad at it. I’m using my Wispr Flow example. It’s just one that you can see how bad it is at that. It’s not just you’re being creative to a computer or talking to a computer that’s going to do what you say and you don’t have to push all the buttons, it’s lossy. It’s lossy in a very important way that using a mouse to click on a button is not lossy at all. It is very, very directed and it will do exactly the thing that you want it to do. Ordering from Grubhub with your finger on a phone screen is 100 percent accurate — unless you’re drunk, which I’ve been. But talking to an agent to have it to go do something is lossy. It is error-prone in very specific ways. These things are going to exist together. That, to me, is the balance that we’re going to have to strike, that all of the hype around AI assumes that what I want is another person instead of a specific thing to happen. You can talk to lots of designers or professionals who use these tools and they will all complain about these tools. We use this app called Riverside to produce the show several times a week. Every single person on this team can issue a PhD thesis in all of the problems with Riverside because it’s the tool [truncated for AI cost control]