Nilay Patel on What AI Means For Expertise at Work

Morra: I want to talk a little bit about expertise because one of the things that I think about as a content creator and a writer, and I've had it happen to my work, is that Gen AI is going to steal my expertise from me, or it's going to democratize expertise so much. I mean, you already see this with things like Canva, right?

You had the CEO of Adobe, Shantanu Narayen, on, which was a great interview. He talked about the fact that everyone has a story inside them, but they don't always have the tools to properly express that story. And he said that Adobe with AI will basically give everyone the tools to express the story. And that really got my feathers in a ruffle, because I thought, “God, I've spent years honing my craft, working so hard to tell a story.” The CEO said, “I want everybody to just be able to tell a good story.” How do you think about not just AI, but the general democratization of access to tools in the context of people who are driven and want to become experts and want to become the best at what they do? I assume you're one of them that makes it.

Nilay: My conception of myself is not a podcaster, an editor, or a manager. It's a writer. I write sentences for a living. And boy, do I work hard at sentences…My goal is to make every sentence valuable… My ideal sentence communicates one thought clearly and that's always kind of what I'm going for and that's really hard to do — It's very tempting to jam multiple thoughts into a sentence. 

There is yet to be an LLM that can write like me. It is just really difficult…. I've been writing sentences and thinking about sentences very intently for 15 or 20 years. I don't think everybody has the same relationship with sentences. That's just a personal quirk….But my expertise is also in knowing what to write about. And so this is a split that I would give you. I care very deeply about the craft of communicating, how I communicate, what my voice is used for, and how to deploy it in literal sentences. The actual problem is what am I communicating for? For new ideas, fundamentally new ideas, like taking two new pieces of information and saying, “Okay, here is the story that these two pieces of information are trying to tell you.” That's hard. And LLMs literally are statistical representations of the past. And so they will not be able to do that work. They will not be able to do that work as well as I can, even if they get 80 % as good at writing sentences as I can. They will never be able to identify the new idea. And so I don't worry so much about the democratization of access.

Morra: Actually, I'm sorry, pause. Say more about that because I think a lot of people think of LLMs as crafting new ideas based on a volume of information. But you just said they're statistical models of the past.

Nilay: An LLM, its base level of work is to ingest an enormous amount of data. Mostly from the internet, which is publicly available. That's the thing that they have ingested the most of. The work of training the model, which is very energy intensive and very difficult, is figuring out the statistical relationship between different kinds of information in all the ingested data.

So whether that is the statistical relationship between different words across the printed internet or the statistical relationship of different colors and values of pixels, brightness values, color values of pixels, and images that it has ingested, it's that big statistical body of relationships that is an AI model. And then when you ask it, when you prompt it in some way, the process of inference is generating what amounts to a very fancy autocomplete answer to your prompt. And so you're like, one of my favorite ones is you say, make me a famous video game plumber. Every model, regardless of its copyright or trademark protections or its safety guidelines, will produce a picture of Mario every time. Because there's no other statistical relationship between that prompt and some pixels except Mario. There are no other famous video game plumbers. There's one, and it will always make you a picture of Mario because it's a one-to-one statistical relationship. And so you just look at that. Take a minute and just think about what that means. It means that unless there's a pre-existing statistical relationship in the world, the model cannot generate the information. And you can kind of force it into, you can kind of force these models into doing something else. You can say, here are two things, smash these ideas together and you can watch it try to smash them together, but it does it without any understanding. It’s always kind of dull. Human beings are much more creative than this. 

And maybe you've never played a video game and you're not burdened by the past in that way and you'll just do something else. So I just think there's an awful lot of opportunity in being original simply because on a technical level of how these models work.

Morra: So when you think about expertise and craft, is your advice to people to dive even deeper into what they're passionate about and hone a craft? Is it to become critical thinkers? How do we think about becoming experts in the age of AI?

Example of AI Generator

Nilay: That's a really good question. A lot of it is an expert in what? In so many jobs, the expertise or the craft isn't using the tool. Every company has people in it who are just really good at using Microsoft Excel. Most companies have people who are really, really good at using PowerPoint.

For these folks the craft isn't using the tool, in taking a natural language input from someone else and then using some complicated tool to deliver some out, some comprehensible output on the other side. That's the, I think that's the place that's most at risk, right? Our interface paradigm for computers is changing. And so you see a lot of those moments tend to have a lot of disruption with it.

And there's been a lot of interface paradigm changes in the past 30 years that you can just point to. You went from text-based computers, command line interface computers, to Windows and ICE, and a whole generation of people, their jobs changed. Punch cards to command line interfaces, a whole generation of people, their jobs changed. We went from that to the Mac, what you see is what you get. Page layout, a whole generation of people's jobs changed. From that to desktop publishing, to...

Touch screens, da da da da da da, and photography, right? The whole universe of jobs changes. The internet then happened. And you just see these stacks of change, I wanna say stacks of change, you just see these moments of huge change. It's a big stack of things that change, and the foundation of it is usually a computing paradigm change. We're gonna go from running all the applications locally to running all the applications on the web. A whole bunch of people got different kinds of jobs the second that happened.

So now here we are at another one, which is very clearly at another moment for that. Some things will change. I don't think that's the big disruptive change. I think the big disruptive change comes when some people get really good at that, and that becomes a skill unto itself. I think this big disruptive change comes when the people who have no access to somebody who's really good at PowerPoint begin to communicate at a baseline level of quality that's as good as the person with the PowerPoint has today.

And I'm generally excited about that. What we end up with is we push the state of the art in communication forward in some way, because some people get really good at the tools, and we bring up the floor because people who had no access to the tools suddenly can do it better, I think that's an app benefit. So Canva is an example you gave. Canva is very democratic. Canva is so democratic that they instruct people that they cannot trademark the logos that they make in Canva because everyone has the same logo, like sort of definitionally if you use one of their templates, someone else has used one of their templates and you cannot get a trademark on the logo template. That's incredible. That's just like, that's just fundamentally incredible that the world of high-quality logos has gotten to the point where the people who make the software like this is, we know it's good. We know you like it. You can't trademark it because someone else has used the same logo.

Morra: And we're okay with that. That's the other thing— is we're like “Cool. We're still gonna use it.”

Nilay: Yeah, I'll tell you, we get a lot of responses to the Decoder podcast. The professionals were all very angry at the Adobe CEO conversation. They were upset. They have a lot of gripes about Photoshop and the cost of Creative Cloud in Adobe in general. So that all came out. We have essentially the same conversation with the CEO of Canva who's said all the same things. those conversations, you put them next to each other. It's people who make Creative software, talking about how AI will make people creative faster. That's what they're promising. And the response to the Canva episode was like, “Cool. I use Canva at my job. I am a social media manager. I use Canva to make Instagram layouts. And I would love it if those Instagram layouts were a little bit better because I'm already under all this time pressure at work to do more, more, and more. And of course, I want Canva to do better.” And so you just see the audience has this wildly different reaction based on the tool they're using. Canva is not meant for people who charge high rates for their work because of their expertise in Photoshop. It's meant for people who otherwise would have no access at all. And so the promise they're making to folks is we'll make you even better. We'll raise the floor. And no one ever really gets mad at raising the floor. They get mad at commoditizing the ceiling.

Morra: I agree with you. And in a way, Canva is one of the most revolutionary things that we never talk about as being revolutionary. But here's what it's done that's bad, is that in capitalism, people are no longer willing to pay as much for the skills that require Adobe because they're like, just get a 22-year-old and they'll do Canva. It's good enough.

And that's a problem. And that's how I often think of AI for a lot of us, especially when I'm getting older. I'm expensive. It's like, well, we could have good enough and save a lot of money.

Nilay: Yeah, and I don't know that the platforms have any incentives to show people anything other than good enough. If you're Google and you run YouTube, it's really weird for the YouTube content moderation team or the recommendation algorithm team to say, “I think AI content is bad.” Because a huge other part of Google is ferociously making the AI content. And it's the same problem at Meta. You name a tech company. They all have this sorting problem where some enormous part of the company has to say that the work that the other part of the company generates is bad and those things are just gonna crash into each other and I don't know who's gonna win. I really don't know who's gonna win. The reason all the big platform companies, particularly social companies, are investing so heavily into video, AI-generated video, it's not some noble mission to democratize Adobe Premiere.

It's so that you can plug in assets and have it generate video advertising in social video. So you're like, here's the shoes I wanna sell, here's some copy, push a button, and it'll just make you a compelling video ad. And it won't make five for five cohorts, it'll make five million individually targeted video ads. That's the dream, and that's literally the thing that they're chasing after. This thing already glimmers of it already exists.

Morra: My own clients at agencies are already doing that. Gone are the days of the art director spending time. It's now like, we can hyper-target, so let's just make some Canva graphics and let the targeting do its thing.

Nilay: Right. And that, and even that is fairly manual. Like you have to have that idea and you have to like assemble the stack of technologies. have to get, you know, an army of offshore Canva users to generate. I'm saying this is like the glimmer of this is already built into the TikTok ad product where TikTok is selling this capability to start for the advertisers. And so like, you just see everyone sort of terrified at this moment. Like here's why we're barreling towards this. It's not as noble as we're all saying out loud. It's to generate more good enough advertising at a massive scale and hopefully sell more shoes. 

I am hopeful that the market responds to this by saying there will be a thing that provides some higher quality, that some people will get paid higher rates. And I think the challenging part when you democratize a capability is that something that used to be hard for a few people to do and guaranteed high rates for that whole set of a few people, the bottom's gonna fall out and only the very best get a higher rate. And that rate might be actually higher than it was before. But the sort of like “make a living” high rate might just disappear. And that, man, the knock-on effects of that are crazy. Like how do you become the person who spends years grinding away in the middle to become the person who's at the top? Where does that market go? Who are those people going to be? Who is going to subsidize that work? How do you get someone, I was very fortunate, I've written a lot of sentences. I was very fortunate to get paid a bunch of money to write sentences for 20 years, and take the time to think about all the sentences, where's that gonna come from? If when I started, I was just writing AI-quality slop sentences, someone's gotta say, well, I think there's some potential in your sentences. And I don't know who that person's gonna be. So I think that stuff is worth worrying about. 

But overall, that Canva example, right, the floor is gonna get higher. It's very hard to say to a bunch of small businesses, we want your floor to stay low. It's very hard to go to people who don't have the ability to communicate in the formats that we now demand communication to be in and say, actually, you can't. And that is a real tension.

And so you can already see it with the internet at scale in a real way… What I don't know is whether AI can actually do all the things people say it can.

And so if you listen to Decoder, I ask every executive, “Can AI actually do all this shit? I don't know if it can. And their answer's like, I don't know. It's like, “good question.”  And I wonder, there's this concept that I heard at a conference called “pretender technologies.” And it is a thing that gives you a glimpse of what might happen but can't actually get you there. 

I truly do not know if this is a pretender technology or not. But I know that everyone is imagining what would happen on the other side of it being real. And so that's the thing that I'm thinking about.  On the one side, I know what's gonna happen to the internet. It's bracingly clear already. On the other side, I don't know if this technology can iterate its way to the finish line.