TikTok’s “Transparency and Accountability Center” + ChatGPT’s Origin Story + Kevin Systrom’s Artifact App

This transcript was created using speech recognition software. While it has been reviewed by human transcribers, it may contain errors. Please review the episode audio before quoting from this transcript and email [email protected] with any questions.

Kevin, this week I went on a reporting trip to Los Angeles. I was there to meet with some people from TikTok, and I brought you some trinkets.

Oh. what’d you bring me?

Well, we have the branded TikTok jotters, which are some ballpoint pens.

Those are cool.

Yeah. There is some refillable nourishing hand sanitizer with a TikTok logo on it. We have the classic for you notepad.

Wow.

Of course, a riff on the TikTok for you page. And then here is some information about how many trust and safety professionals the TikTok corporation has hired.

[LAUGHS]: This is a real goodie bag, yeah.

And finally, there’s the Wi-Fi password for TikTok if you need it.

[LAUGHS]:

Anyway, based on this, would you say that TikTok no longer poses a threat to our national security?

Yes. Based on the notebook, hand sanitizer, pens, honestly, pretty good password for the Wi-Fi. I’m not going to read it out, but I wouldn’t have guessed this.

You would never have guessed it.

And this crumpled sheet of paper, I believe that my concerns have been sufficiently alleviated, and I no longer believe that TikTok poses a threat to the security of the United States.

Well, you see, that’s why you invite the media down. It’s because that’s the kind of shift in perception you can get.

Well, mission accomplished. Let’s move on to the next subject. I have no more concerns about TikTok.

[LAUGHS]:

[MUSIC PLAYING]

I’m Kevin Roose. I’m a tech columnist at “The New York Times.”

And I’m Casey Newton from Platformer.

This week, Casey takes a field trip to TikTok, I report on some behind-the-scenes developments at OpenAI and ChatGPT. And Kevin Systrom, the confounder of Instagram, tells us about his new AI-powered news app.

OK, let’s talk about this trip. So you went down to LA to do what at TikTok?

So a couple of years back, TikTok announced that they were going to open up this building that they were calling the Transparency Center. And this was around the time that the Trump administration was trying to force ByteDance, which owns TikTok, to sell it off to some conglomerate led by Oracle and Walmart.

And TikTok said, whoa, whoa, whoa, whoa, whoa. Before we go that far, we want to be more open with you than any social network has ever been with anyone. And we’re going to build a room where journalists and lawmakers and regulators can come in and you can basically stare the algorithm dead in the eye and see what it’s made of.

And so I was very excited about this. I was ready to go down, and then there was a global pandemic. And so we took a virtual tour. And I’m just going to say it was not that great. It’s very hard —

Here’s our conference room.

Yeah. It’s very hard to take a virtual tour of a algorithmic transparency center, is what I learned. But a couple of years went by, and conditions on the ground changed. And a few weeks back, I got an invitation to go to TikTok’s office in Culver City.

Wow. So you show up. So set the scene a little bit.

Yeah.

What’s it like? I’ve never been.

It looks like what you would expect a TikTok office to look like. There’s giant LED screens showing TikToks, life-sized representations of the logo, conference rooms named after viral moments on the app. If you’ve been to Facebook or Twitter, this is kind of how they all look.

Totally, yeah, which is always weird. The conference room name thing has always weirded me out. You go to a tech company, and usually, the conference rooms are like —

It’s like, Chewbacca Mom.

Right. I remember — yeah, named after big moments on the platform or Ice Bucket Challenge, which I always think is — you know you’re eventually going to have to lay someone off in there, right? You know someone is getting canned in the Chewbacca Mom room. And it’s like —

That’s right. The results of our investigation into your embezzlement case are in, and we need to talk to you over in the Ice Bucket Challenge. [LAUGHS]

Meet us in Grumpy Cat for a discussion of your performance evaluation.

Yeah. We’ve got some bad news for you over in the Tide Pod Challenge zone.

So anyway, it’s got —

Bring your lawyer.

It’s got conference rooms with funny names. What is the actual Transparency Center? Describe it.

Yeah, we were in two rooms. One was just a big conference room, where we heard some speeches from some TikTok folks, which I can talk about. And then after that moment, we crossed the street and went to the Transparency Center itself.

Oh, it’s a different building?

It was in a different building.

Was it transparent? Did it have glass walls?

Yes, you could look in on it from the street. No, you couldn’t. I would actually say it was heavily secured. So we go in there, and we were sort of led on a presentation. And it’s called a Transparency Center, but the way that I would think about it is a little bit more like a children’s museum about TikTok, where you have a docent, who is a very friendly, bright, articulate, host.

And she brought us through a series of exhibits that were interactive. And there were a series of these large, smartphone-like screens, where she could swipe up and swipe down, almost as if she was swiping through TikTok itself.

OK.

And then we’re led to the real heart of the exhibits. One of them is essentially a guided tutorial, where you could — it was like —

god, it sounds so boring, and this part kind of was.

It was like, how does TikTok keep people safe? And you tap a button. And then it would show you some visual media and try to answer that question. And then you could go back and they would say, well, how do you ensure that this doesn’t happen? And so basically a fancy Q&A zone.

Right. The terms of service in museum form.

Exactly. The more interesting of the exhibits was a room where you could pretend to be a TikTok content moderator. And you know I was in heaven doing this, right?

Yeah.

I love the subject of content moderation. I write a lot about content moderation, but I have not really been in a position to do any moderating of content. And what TikTok has set up is a facsimile of their content moderation system, where you could choose any of the bad things.

So you could choose hate speech or violent extremism or nudity or bullying. And they would then show you a handful of videos. And next to those videos, they would show you relevant policies from their own community standards like about bullying or hate speech or extremism. And then you would decide, does this violate the content or not? And then you would either say like yes, it does, and if so, which policy does it violate? Because that’s something that content moderators have to do. They can’t just say, good or bad. They have to say, no, you broke this rule. And then you would get the result, and it would say, well, were you right or were you wrong? And I was right a lot of the time —

Hey, not to brag.

But I was also wrong, too, in ways that surprised me. And I’ll just tell this story quickly because I think that we have so much heartburn over content moderation in this country. And people get so mad about having their posts removed, but we also get so mad about posts that are left up that we think should be removed. And I think everyone would benefit from spending an hour in a chair, just trying to decide whether posts belonged on TikTok or Facebook or Twitter or whatever.

There was this one where — I’d been reading these policies about bullying and harassment. And there was this TikTok that was taken from inside — you know a refrigerator case at a bodega where they have the cold drinks type of thing?

Yeah.

So somebody was sort of behind that and was pointing a squirt gun. And it was a bright green squirt gun. It was clearly a squirt gun. And some kid was going to reach in to get a can of Coke or something. And the squirt gun goes off and blasts the kid in the face, and the video ends with him recoiling and falling backwards. And I’m like, this seems pretty clear harassment of someone.

Right.

Maybe the kid was in on it, but I can’t tell from the video. And instead, it just looks like somebody being mean to innocent people shopping at a bodega.

So you clicked the, “this violates the rules. Take it down” button.

Yeah. And they were like, no, that actually doesn’t violate our rules.

Fascinating.

Yeah, so look, if you want to go squirt a kid with a squirt gun anywhere in the world, put it on TikTok, go viral, make money —

Well, I know what I’m doing tonight.

Yeah, there’s your weekend plan.

OK, so they see this simulator, this flight simulator but for content moderation. What else?

So those were the big two exhibits. And then on the way out they said, now, there is this third exhibit, but we’re not going to take you to it. And it’s actually in this room over here. And so she sort of pointed to this area where we were not allowed to go.

And she said, in this room, you would pass through some metal detectors. You would sign a non-disclosure agreement. And then you would sit down and you would see the TikTok source code. And it’s a room where the source code can be inspected.

Wow. Like the Ark of the Covenant vibes?

Yeah. And one of the big challenges that TikTok has is that no one trusts anything it says on a handful of subjects. One of the things that people don’t trust it on is, could the Chinese government interfere with ByteDance and insert code into TikTok, either to surveil Americans or maybe promote pro-China content? And could they put that into the source code itself?

And so there is a room where people like regulators academics or whoever else TikTok allows can go in and look at that code. I don’t write code, so I wouldn’t be able to read it, but there is such a room.

So you didn’t get to go in the secret code room, but it does exist. And other people have gone in there and can go in there, who do understand how to evaluate code and could go look for evidence of Chinese interference?

Yeah, that’s right. Now, of course, this gets tricky in a hurry because so you look at the code, and it looks fine. Then you walk out the door. Then what happens to the source code?

Right.

And so TikTok has a lot of these challenges, but they are starting to think through, OK, how could we try to organize ourselves in such a way that we address those concerns more permanently?

Right. So let’s just back up for a second and talk about why you were at TikTok in the first place. So as we’ve talked about on the show, TikTok is under a lot of scrutiny from regulators and politicians. It’s being banned in certain state governments and, actually, federal government employees now are no longer allowed to have it on their phones, their work phones.

And part of the reason for all the suspicion, as we’ve discussed, is that it’s owned by ByteDance, which is a Chinese company. And Chinese companies can be influenced by the Chinese government.

Yeah. And I mean, “influence” is probably not strong enough a word. When the Chinese government tells a Chinese company to do something, the company either does it or they’re shut down. There’s not a lot of room for negotiating there.

And so TikTok has been on this furious quest to prove to American regulators, mostly, but also journalists that it is independent, that it is not being steered by ByteDance in ways that would make it a threat to US interests.

That’s right. And to the extent that American’s data is being stored in China or that Chinese employees would have access to that data, TikTok is trying to say, we are going to undertake a huge effort to ensure that doesn’t happen.

And of course, the whole goal for this is that ByteDance can continue to own TikTok and minimize the risk that someone else is either going to shut it down completely or force them to sell it.

Right. And I think it’s very interesting that the way that they have decided to be transparent is with this children’s museum, as you described it, but also this allowing people to go and inspect the source code or a version of the source code at a moment in time.

And I think this is an area where I’ve changed my mind in the last few years. I used to think that transparency would solve a lot of problems with social media, that if they just opened up the algorithms and made them viewable so you could see why are certain posts ranked higher than others, that that would actually go a long way to increasing trust in these social networks.

And I still think that’s true in some cases. I think transparency is usually a good thing, but on this algorithmic transparency bit, it just seems like that’s not — like it’s only a partial solution because only one of the worries that people have about TikTok is that ByteDance or the Chinese government could be inserting malicious code into its code base.

A lot of the other fears are about content moderation decisions, frankly, like which posts are allowed to stay up and go viral? And which posts are taken down or demoted so that no one sees them? And that kind of thing typically wouldn’t be in the source code. It’s not like there’s a line of code that says, take down all videos of Tiananmen Square or something like that.

That’s right. The first exhibit, one of the things that it did, one of the questions that it would answer for you is, why am I seeing the videos that I see in my feed? How are these videos chosen for me?

And so the program explains that to you. And as part of the explanation, they show you the code snippets. And the written explanation will say something like, well, we take a set of videos, and then we run it through our machine learning systems. And a score is generated, and then we reduce that to a smaller set of videos, and we run that through some more loops.

And on one hand, I think they have explained, with a high degree of accuracy and a pretty good degree of depth, how TikTok works. So they are being transparent about that. And yet, at the end of the day, can I really say with any specificity why I saw a video? It’s like, no. It’s like, some math happened in a computer, and now I’m seeing this. And I think that’s the truest explanation for why you’re seeing what’s in your feed. And I don’t think that’s a very satisfying explanation for a lot of people.

And it makes me think of this idea that Elon Musk had when he took over Twitter, that he was going to open the code base and look at the code and find this smoking gun evidence that Twitter had been suppressing certain kinds of posts.

And instead, what he found is kind of a normal algorithm, where certain posts are promoted, not based on their ideology, but how likely it is to keep you on the app. That’s what I think I’ve learned by talking with engineers who build these algorithms is it’s very rare that the controversial thing is hardcoded into the app itself. The app itself is usually just trying to get you to stay on the app.

Yeah, but businesses make business decisions. And we don’t often talk about that when we’re talking about content moderation decisions. These people are all trying to make money.

Totally.

Yeah.

So what else did you see? What did they say? What did the TikTok executives talk about? And I’m curious. Just what else happened?

Yeah. I think the centerpiece of the day, aside from the visit to the children’s museum, was a set of talks from some TikTok executives, where they tried to give us the lay of the land and talk in some detail about Project Texas. Project Texas is TikTok’s effort to move all Americans’ data to the United States, to put it in what they call a secure enclave. Which will be managed by Oracle.

Oracle sort of being the US babysitter of TikTok in this case.

That’s right. They’re setting up a subsidiary called US Data Security, and that’s going to be in charge of making sure that everything TikTok stays in the United States. They’re saying that they have already spent $1.5 billion on this and that it will be a massive ongoing expense for them, but that at the end of it, there will be an independent board of directors who will be able to check in at any time and say, hey, did anything untoward happen with data?

So that was what they wanted to talk about was the Project Texas and their very expensive-sounding efforts to appease regulators and politicians in the US. Were there any things that they notably didn’t talk about or didn’t want to talk about?

Well, I wanted them to get into this stuff on the record a little bit. They wanted to do that particular portion of the day on background, which is a term in journalism, which is basically like, you can report what we said, but we don’t want you to attribute it to specific people.

And I imagine that’s because these are very sensitive negotiations that when they are going to talk to the Biden administration, it’s going to be some extremely high-ranking official and not the more, I guess, mid-level managers that we were talking to, in some cases. So I get it. It’s sensitive, but at the same time, you bring reporters down to the Transparency Center to talk about your transparency efforts, and you’re like, well, but don’t attribute any of this information and we don’t want you to record this part. It does get a little frustrating.

Right my transparency center t-shirt is raising questions that have already been answered by my transparency center t-shirt.

I feel the same way about US Data Security as the subsidiary name. It’s a little try hard.

So after this visit to TikTok and it’s transparency center, are you feeling any different about TikTok and its prospects to avoid punishing regulation or an outright ban in the US? Are you feeling more comfortable about TikTok?

I mean, on one hand, I am glad I got to go down there. I did learn a lot about how the recommendation algorithm works, about the way that they moderate content. And also, I do think it’s good to just meet the executives who are working on this stuff. And you sit down with them, and they’re mostly American citizens, and they come across like any other employee of any tech company I’ve ever written about.

They’re trying to do their job. They have their challenges. They hope that it’s going to work out. They’re not sure. And there’s a certain amount of comfort there. But at the same time, there are the other issues that we have talked about on this show. We had Emily Baker White come in and she said, yeah, ByteDance employees tried to figure out who her sources were by using her IP information. ByteDance was not particularly forthcoming about Project Texas until Emily wrote about it.

And, on top of all of that, the fact remains that at the end of the day, if China wants something from ByteDance, it’s almost certainly going to get it. And that’s always been the trap that TikTok is in.

So I don’t know, man. I don’t want to sound like I’m weaseling out of answering your question, but I do find the case of TikTok really hard because I think there probably are good reasons to take really, really strict action against it. And I’m not sure that just being extra transparent about everything is going to answer our concerns. What’s your take on all of this? Do you think a more transparent TikTok is one that should stick around or do you think we need to scrutinize it more?

I would say, yes, I believe both those things. It seems like as much as they might want to portray themselves as being independent, when it really matters, when there’s something that the Chinese government wants, whether it’s user data or to make a certain kind of video go more viral or less viral, the question is, what capacity does TikTok in the US have to push back on that? And so far, I have not heard them talk about that in a convincing way.

Yeah.

I would like to see them, in addition to doing all this transparency work, I would like to see them meaningfully and honestly reckon with some of what’s out there about how they operate and how ByteDance operates, with TikTok sort of, kind of at arm’s length, but not. I just haven’t really seen them address that stuff.

When they get asked about it, they usually say, we’ve never been asked to do anything by the Chinese government and we would say no if we were asked. So that’s what they say and. OK, but to your point, it’s like, you can’t really tell the Chinese government no. And that is the specific thing that I feel like that they have never really reckoned with.

At the same time, the reason that I struggle to say “ban them” is because I feel like I know, with a fair degree of confidence, what will happen if TikTok gets banned, which is that YouTube and Facebook will just benefit tremendously and that you’ll see all of the people making TikToks right now will just be making shorts and reels. And it’s like, meet the new boss, same as the old boss.

Right.

And 99 percent of our lives will be exactly the same. It’s just that one more competitive force has been taken out of this already pretty anti-competitive market.

Right. And as far as what they could do to restore trust in the US, at least for me, if they came out and were just radically honest and said, look sometimes ByteDance does stuff that we don’t approve of, and we don’t know about and it makes us really uncomfortable, and it’s kind of awkward, but they’re are owners. And so we just have to deal with that, and here are some things that we’re doing to mitigate that. And here’s some of the fights that we have internally. I actually think this is a scenario where radical honesty and not this glossy corporate transparency stuff could go a long way, at least with me.

And the absence of it makes you feel like there is something that they are afraid to say.

Yes.

And that is actually, I think, at the root of my discomfort here, is that sense that there is an underlying fear at that company that is always being talked around. And it’s why I can’t sign off fully on my brain on some of the stuff.

Totally. Well, I’m glad you got to go. That sounds very exciting. And thank you again for the gifts.

Yeah. Did you do any reporting this week?

[LAUGHS]: Are you my boss? Yes, I did some reporting this week.

Oh, exciting.

I want to tell you about it right after the break.

Let’s do it. [MUSIC PLAYING]

All right, Kevin, you have finally done some reporting.

I don’t like this attitude. Listen, it takes a long time to make a delicious meal, right?

It’s true.

If you’re just whipping out microwave Trader Joe’s frozen meals, it’s going to be quick, but if you want to prepare a filet mignon, it’s going to take you some time.

So this week, you delivered a true prix fixe menu of journalism. And it was about one of our favorite subjects, AI.

Yeah. So I’ve been looking into OpenAI, the company that made ChatGPT and DALLE2 and a lot of these other AI products that we’ve talked about on the show.

And I talked to some people who know what’s going on inside the company, including current and former employees and really just tried to lay out the origin story of this product, ChatGPT, that has now taken over the world in this incredible way. I mean, it’s scaring Google. Microsoft is investing $10 billion into OpenAI.

And it felt like it came out of nowhere when it landed too, right?

Totally.

So I think it’s a really good question. It’s like, where did this thing come from?

Totally. So I’ve been looking into this for a couple of weeks, and I found what I would say are three big takeaways from my reporting. The first is that ChatGPT is just way more popular than I thought. It’s got more than 30 million registered users.

Wow.

More than 5 million people use it every day.

Wow.

And for a product that is only, really, two months old, that is a phenomenal number of people.

That’s huge.

So just by contrast, Instagram in its first year got 10 million users. And that was seen as one of the fastest-growing things of all time.

Yup.

So getting 30 million users within two months, I just don’t know that I’ve ever seen a software product grow that fast.

Put that in context. That’s actually bigger than the “Hard Fork Podcast.”

[LAUGHS]: It’s slightly bigger than the “Hard Fork Podcast.” Another really interesting thing I found out is that this was a total accident. So OpenAI, its plan for most of last year, the thing that it was working on was GPT-4, this new language model that they were developing. They were very excited about it.

And so the plan that they had been going on was that they were going to release GPT-4, or whatever it’s going to end up being called, early in 2023, along with a series of chat bots that were more narrow. They were more aimed at business users and that that was going to be how a lot of people interacted with GPT-4, was through these limited chat bots.

Sort of very tentative experiments.

Totally, totally. But in November this announcement goes out to employees. And it basically says, OK change, of plans. We’re going to release a chat bot now, before GPT-4 comes out, in part because we don’t want another I company to beat us to the punch and release a chat bot before GPT-4, but also because we think releasing a chat bot now will let us gather some feedback and ultimately make GPT-4 better.

So they decide to dust off this old chat bot that had been sitting on the shelf and update it and then launch it to the public for free as something called Chat with GPT 3.5, which I think we can all agree is —

Well, I didn’t think there could be a worse name than ChatGPT, but we found it.

So they announced that they’re going to do this, and they give it a two-week deadline.

Wow. What is this, Elon Musk’s Twitter?

[LAUGHS]:

That’s extremely hardcore vibes.

I don’t know if they actually literally slept on the floor of the office, but it was a sprint. And internally at OpenAI, there were some people who were just kind of confused by this.

They said, we’ve been working on GPT-4. It’s getting to a place where we feel like it’s almost ready. Why are we scrambling to release a chat bot based on our last language model, which has been out for two years and which really isn’t the state-of-the-art anymore? But OpenAI’s leaders really wanted to put something out there. So they make this announcement, and then 13 days later ChatGPT comes out.

So one of the reasons this is interesting is that now it is February, and none of these other chat bots have emerged. So what happened? Did OpenAI just misread the market? And what do we know about these other language models?

So we know that other companies are working on similar projects. There’s a company called Anthropic, which has its own chat bot in the works called Claude. It’s been rumored that DeepMind, which is a Google AI subsidiary is close to coming out with a chat bot too, but none of these chat bots have come out publicly yet, which makes it seem like this fear that these OpenAI executives had may have been overblown.

I remember that the original premise of OpenAI was that they were not going to move that fast. They wanted to make their work, quote, “open and safe.” That’s not the sort of thing that you think about being built in 13 days. So how did they reconcile those ideas?

Yeah, so OpenAI is a strange beast. It was started in 2015 as a nonprofit. It was started by this all-star group of tech people, Elon Musk, Peter Thiel, Sam Altman, and Reid Hoffman. It had all of these initial sponsors who chipped in the money to get this thing off the ground.

And the whole point was that it was not going to be driven by these narrow commercial interests. They pitched it as the anti Google or the anti Facebook, where those companies were developing AI to suit their business needs, whereas OpenAI was going to be half research lab and half humanitarian AI organization that was going to make sure that its AI was safe and responsible and steer this whole area of technological progress in a better direction.

Yeah. So what changed?

Well, it’s a long story. Basically, in 2019, they decide, we need to raise all this money because it’s very expensive to build and train these giant AI models. So they started a for-profit subsidiary. So basically, it’s a confusing structure, but right now, there’s OpenAI, the for-profit subsidiary and then there’s the nonprofit umbrella organization that still has a board that oversees the whole thing.

But they started behaving, in some ways, a lot more like the companies that they were competing against. They started trying to commercialize their research, trying to figure out how to make money from it. They struck a deal with Microsoft that gave Microsoft some exclusive rights to GPT-3.

And this was all controversial within the AI research community where people said, wait a minute. Weren’t you guys supposed to be not behaving like a startup that was just trying to raise a bunch of money and sell things and I push things out into the world, maybe before they’re ready?

And so if you talk to OpenAI, they’ll say, well, we still have some of the best safety research going on in the world. We’re still taking our time, doing things deliberately. We’re not just being reckless and putting things out before they’re ready. But I think there are also people inside the company who are uncomfortable with how fast things like ChatGPT are coming into existence and maybe worry that the company is cutting some corners in an attempt to get to the market.

So that’s the second thing, that ChatGPT was not expected to be nearly as big as it is and that it’s kind of been an accidental hit. The third thing is that ChatGPT’s success within OpenAI has created all of these interesting problems for them. They did not expect this to be as big as it was.

Oh, come on. That’s silly. Of course this was going to be popular.

No, it’s not because, look, Meta had released a chat bot last August called BlenderBot that you’ve probably never heard of because no one used it.

Well, I’ve heard of it, but I cover Facebook, but what I will say is I didn’t use BlenderBot. In my memory, though, it’s like, did BlenderBot even do half the stuff that ChatGPT does?

Well, no. In part because Meta nerfed it to prevent it from being misused because right after it was released, people did get it to say crazy things, like that the election was stolen and some anti-Semitic stuff. And this kind of thing had happened before. That was the whole deal with Microsoft’s Tay bot a few years ago.

So BlenderBot from Meta comes out. It doesn’t really make a splash. The reception is very lukewarm. And so there are people inside OpenAI who think, OK, we’re slapping a new interface on this two-year-old model. It’s already been out there. People have already seen this kind of thing before. It’s not going to be that impressive. So they were totally caught by surprise when it became a total cultural phenomenon.

This feels like a case of tech people living in a bubble and not understanding how cool their own lives are. You know what I mean?

Mm-hmm.

And, of course, I’m guilty of this too all the time, but when you have access to a tool that can do as many things as ChatGPT does and feel somehow bored or unimpressed by it, that’s funny to me.

Totally. And it speaks to just the way that the employees of these companies, I think, get desensitized to the power of the things that — look, because inside OpenAI, this is the second-best model.

It’s built on this two-year-old model, GPT-3, and they already have, internally, access to tools that are better than this. So for them, this was kind of an afterthought. And they didn’t realize that it would totally upend the technology industry, be used by millions of people, wreak havoc in high schools across the country. They just did not see this coming.

Yeah.

And so they’ve been scrambling to address some of these issues and patched some of the flaws. So for example, this week, they put out an AI detection tool that they’re aiming at teachers and other people who might want to be able to detect if a given text is written by AI or not. It’s not super clear that it’s all that effective, but they have been putting this out there. They’ve been talking with educators and trying to cool temperatures around this. So they’re playing catch up, but it’s fair to say, I think, that this has been a chaotic couple of months inside OpenAI.

Yeah, it’s like a classic moment of catastrophic success. What I wonder is how much has that changed the trajectory of the company. Now that their second-best, afterthought product has galvanized so much attention, how much does that change what they want to do with the company?

Well, in some ways, it’s made their lives easier because they now have this $10 billion investment from Microsoft. So I think in some ways, it’s been very good for them, but in other ways now, there’s a lot of targets on their back.

Going first in such a competitive and controversial area, as large language models and advanced AI, has drawbacks as well. And you hear that not just from people at Google and other AI companies, who are clearly jealous of the success and attention that ChatGPT is getting, but you hear that from less obviously biased sources too.

They just say, with something like this, you have to really be careful because if you’re rushing this out and you’re not paying enough attention to the possibility that it could become used by millions of people overnight, you’re not going to build in some of the guardrails that maybe you need.

Right. And that was, of course, the lesson from the 2010’s and the social media boom was that if you go really fast and you just try to grow at all costs, inevitably you’ll make some mistakes, but my sense is that the people at OpenAI know that. I guess the question is just if the people making those points are being listened to.

Totally. And this not only has changed OpenAI’s trajectory, but it’s set off this total arms race, where now you have Google declaring code red and trying to fast track a lot of its AI products. You have Baidu, the Chinese tech giant, which is reportedly coming out with its own AI chat bot. And then you still have DeepMind and Anthropic and all these other companies that are gearing up to release their own.

Now they’re racing too. So this decision that OpenAI made to rush this chat bot out, as much as it may have seemed like an afterthought at the time, really did set off a race in AI that I think is dangerous.

And I think OpenAI believes that it’s dangerous too because in their charter, from many years ago, they specifically say that as we approach AGI, this artificial general intelligence, the intelligence that’s equal to that of a human, they specifically say that we’re not going to feed into this race dynamic, that if someone else, if another tech company or research lab gets closer to AGI, than we are, we will stop what we’re doing and help them.

So it really is a case of — I don’t want to say mission drift, but almost like they started off as a nonprofit. They started this for-profit subsidiary several years later. And there’s these competing impulses that are still inside the company, where a portion of the company thinks, whoa, whoa, whoa. Why are we doing what we said we wouldn’t do and racing to get this stuff out into the public’s hands?

And then there’s this other force that’s like, well, now we’re a big tech company, and we’re competing with big tech companies. And if we don’t make this stuff, someone else will, and they’ll get the credit and the funding. And maybe their values won’t be in line with ours.

What you have me wondering, though, is wasn’t all of this inevitable? Wasn’t someone always going to go first? And, assuming it was good, wasn’t that always going to cause an arms race?

I think there’s an element of that, sure. I think that it’s plausible that if OpenAI had not released ChatGPT as quickly as they did that someone else would have come along and released something that had fewer guardrails or was more easily abused.

And that’s certainly what they would tell you at OpenAI is that they did years of safety work on the base model GPT-3 that was used to make ChatGPT. So it’s not as if they had no safeguards in place. They had lots of them. It’s just that they weren’t expecting this many users.

So I think it was always inevitable, but I also think it’s very important in the early stages of this next phase of AI to set norms in the industry, that before you release something like ChatGPT, you do months or years of testing on it, in a very limited setting, where it’s not going to be used by every high schooler in the country overnight. You allow society to slowly adapt and put in place some of the structures that allow this to not be so disruptive. You don’t just flip a switch overnight.

Yeah. So what’s the next turn of the screw here? They’ve got all of this technology cooking. There was some reporting this week that Microsoft is about to announce a bunch of new integrations, which would bring OpenAI’s tech even further into the mainstream. They seem like they’re in the lead now, but what do you think is next?

Well, they’re still working on GPT-4, which is their next-generation language model. That’s still slated to come out, from what I hear, later this year. And they are using a lot of the feedback that they’re getting on ChatGPT to make GPT-4 better.

So I expect that there will be another big launch in the next few months, that there will be this — I don’t know whether it’s going to be — what we don’t know yet is what GPT-4 is going to look like, whether there’s going to be an easy-to-use chat bot interface like there is with ChatGPT, but that is still the company’s big project this year.

And in some ways, I think it’s going to have an easier time. I think if GPT-4 had come out of nowhere the way that ChatGPT came out of nowhere, we would be seeing an even bigger societal response to it.

And it’s something that I’m really going to be looking at when GPT-4 does come out, especially this question of bias. So I’m sure you’ve seen but in the last few weeks, there’s been this conservative backlash to ChatGPT. There was an article in “The National Review” called “ChatGPT Goes Woke” that was all about how —

Oh my god. No, I didn’t. That sounds like an article ChatGPT would write if you said, write a conservative article about ChatGPT.

Right. So very predictable, but also interesting in the sense that people are trying to figure out, does this chat bot have ideology? Has it been programmed in a way that makes it favorable to one side or the other? And so they’ve been testing it and saying, write an admiring poem about Donald Trump. And it’ll say, we can’t do that. And then it’ll say, write an admiring poem about Joe Biden. And it’ll do it. And so there are people who are saying —

Let’s not question the wisdom of the AI here. This is state-of-the-art technology. If it can’t write an admiring poem about Donald Trump, maybe it can’t be done.

Right. It’s becoming a new frontier in the content moderation, free speech wars. And I think that that’s something that I’ll be watching very closely this year and I think for probably several years to come is just the emerging battle over centralization and control and censorship in these models. I think this is going to be just as controversial as some of the conversations around social media were in the last decade.

Yeah.

So get ready. We’re going to have hearings.

Love a hearing. Very good story, Kevin, and I know we already talked about one of my stories this week, but I wrote another one too.

Dammit.

And we’ll talk about that after the break.

[MUSIC PLAYING]

All right, Kevin. So this week, we’ve talked about TikTok, and we’ve talked about AI, but now I want to talk about something that brings both of those things together. And that thing is called Artifact.

What is Artifact?

So Artifact is this new app that the co-founders of Instagram are working on. And I think it’s interesting in part because it is charting a path for what I think a lot of apps are going to do in a world where these large language models are now available.

So the co-founders of Instagram famously, they sell their app to Facebook for something like $1 billion.

$715 million because Facebook stock went down after the closing of the deal. It’s a sensitive subject with them.

[LAUGHS]: Wait. I got to find the sad trombone.

Yeah.

[SAD TROMBONE SOUND EFFECT]

There you go. [LAUGHS]

Only $715 million. So they sell this app. It famously takes, off becomes a huge cultural phenomenon, leads Facebook to a new era of growth. They leave the company.

That’s right.

And so how do we get to Artifact?

So on Friday, I got an email from Kevin Systrom.

One of the co-founders of Instagram.

Yeah. And in a lot of ways, it felt like an email that I’d been waiting to get for five years, since he and his co-founder, Mike Krieger, quit Facebook. We haven’t heard much from them since then, although they did work on a really cool website called RT Live, which helped track the spread of the pandemic.

It was a free website that anyone could visit, but aside from that, they hadn’t tried any new business. And I always assumed they would. And Kevin emailed me and said, hey, we’re actually ready to talk about what we’re doing. And he’s actually here right now to talk to us about it. Hey, Kevin.

Guys, how are you?

Good. How are you?

Good. I’m super excited to talk, yeah.

Great name, by the way.

I know, right? It’s good.

[LAUGHS]: So welcome to “Hard Fork.” Tell us what Artifact is.

Yeah, so Artifact is a hyper-personalized newsfeed driven by the latest in machine learning. It’s like TikTok for text. There’s a lot of text out there on the web, most of it news, blogs, articles. And we take all of that, we understand it, using machine learning. And then we say, hey, user, you signed up for Artifact. What are you into? And then we start matchmaking, and we present you a feed of stuff. And at first you’re just, like, OK, it’s a newsreader. But if you use it for — I don’t know — a week or two or depending on how heavily you use it, it becomes really different. And that’s because it starts to understand what you like and you don’t like.

So I’ve now been using it for well over a year, and it knows I’m into Japanese modern architecture. And if a really cool example of Japanese modern architecture pops up, it’ll serve that to me. It also knows things that I don’t tell it. Obviously, I’m into Instagram. I’m into all the drama around Instagram and where it’s headed. And I don’t think I would ever say that on a survey.

Right.

But it clearly knows I’m into that, so it serves it to me. So I get to follow along with the history of Instagram as it gets built. So that’s what Artifact is. It’s a newsfeed for you. And over time, it morphs from just a newsreader into — presidents back in the day used to have these things called valets. And valets would —

Casey still has a valet.

You do? OK. Fancy.

[LAUGHS]: I’m hiring for one. The job posting is up.

But they read everything for you. They find the most interesting stuff in the morning, and then it’s on your desk. There was an executive at Facebook who had this, one named Nick. I actually thought it was smart.

Can I ask some questions just about the app itself? Because I have not played around with it yet, although —

Of course.

— I’m on the wait list, so as you as soon as you feel like —

Wait, wait, wait. Casey, you didn’t offer him an invite?

You know what, I’m inviting him right now. But Kevin asked you a question.

Casey never thinks of me in these moments. He’s very selfish.

Well look, Kevin and I compete. We compete each other. It’s not in my interest, always, for him to have access to what I’ve access to.

Right. We’re frenemies. So this is not generative AI, right? These articles that you’re showing users, they’re not being written by AI. They’re just being ranked by AI. Is that correct?

That’s right.

So they’re coming from publishers. They’re coming from “The New York Times,” “The Washington Post,” “The BBC,” wherever. The app that you’re talking about, it just sorts them into a feed based on personalized guesses about your interests.

Yeah. We do not generate text right now, though I will say I’m very excited about generative text in the future for us, less about replacing writing and less about replacing authors or publishers, way more about synthesizing events and being able to point you to the right sources on the events.

If a large event happens, Artifact launches, it’s like, OK. These are the people covering the tech side of it. These are the people covering the competition side of it. Here’s what they think. And it allows you to quick slice an event much more quickly. So to answer your question directly, Kevin, no, we’re not doing generative text yet, although, of course, that’s an option.

I’m interested in the social stuff. So right now, if you start using Artifact, you’ll see that core news feed, and you can start to teach it about your interests. But there is a Beta group of users, which I’m in this group, that sees a couple more things. There is a more social feed, and then there’s a direct message inbox.

And I’m somebody who has been aching for a Renaissance of social apps. I want to see new ideas in this space. And so this, to me, is one of the most interesting and exciting things about Artifact is what you can do socially when you have people interacting around the news stories that are most of interest to them.

Yeah. First of all, I’ll apologize and say this is like one of these bad habits I think we learned at Instagram slash Facebook, which is these feature gates, where some people have a feature, some people don’t. And you’re doing testing and all that stuff, but what I’ll say is our intention is to get it out to everyone as soon as it’s ready and we’re ready.

But I want to point out that when we launched Instagram, it was a filter app first for everyone. Most people didn’t even know that there was a network behind it. It was a filter app. My intention was always for it to be a social app, and I was bummed when everyone talked about it as a filter app.

But what I realized in there is that the best networks are first to utility, and then they piggyback on that utility, and they become a network. I think Facebook’s actually a great example of this. It may not be so clear, but it was clearly just a directory at the beginning. It was about finding people, communicating with them. And then it became the social network that it became once they added the news feed and all this other stuff.

So Instagram and Facebook are actually not that different in their evolution. So we decided, probably a couple of months ago, when we talked about launch, that we were just going to focus on the utility part of it first and be really good at that because starting a social network from scratch, I think, is the wrong order.

You see a lot of people launching things today that are social networks first and then utility is questionable. And it’s because the utility in the social network only makes sense when you’re at scale, when you have a lot of people. And it’s very hard to bootstrap that.

I’m very excited about Artifact being a place where you can discuss what you discover. So I don’t want to be an app that is discussion first and discovery through only discussion because I think that causes filter bubbles.

One of the things TikTok does really well, I think, is you be a nobody, post an awesome video, and go viral. That’s not possible on Twitter, really. That’s not possible on Instagram, or it wasn’t, really.

And in this place, I want a great post from an epidemiologist on Substack who nobody knows to become the de facto thing that everyone reads if it’s great. And then I want people to be able to discuss it and post about it and say what they think about it and debate and argue or whatever, but that comes second.

So it’s this way to connect around things that inspire you, but maybe backing up for a second, I think the best companies often do one slice of what existing companies do really well.

When Instagram launched, everyone was like, but I do photos on Twitter. I do photos on Facebook. Why would I use Instagram? And it’s like, actually, if you just focus on one job in particular, you can optimize for that job and create a really wonderful experience around it.

Yes, you can get news on Facebook. Yes, you can get news on Twitter. And yes, do I think the discussion on Twitter around news is way better than on Artifact right now? Yes, of course. But it doesn’t have to be that way forever. You can create a space for these things to blossom, but you have to do it in order and sequence matters.

I think there’s this feeling in the media business, at least among the folks that I talked to, that the media has been burned by people in Silicon Valley who come along, and they have some new idea about how to revolutionize the news industry, and they make all these promises and they say, well, we’ll share profits with you or we’ll pay you for exclusive things with your content or we’ll send a bunch of traffic to your website.

And then it just always ends up feeling like a bad deal or things go sour. I think there’s been a lot of failed attempts to sort of unite the media industry, which is producing all of the news and the technology industry, which wants to distribute it to consumers.

So have you found that in talking with publishers about including their stories on Artifact? Have you found that there’s some missing trust there? And how do you plan on working with publishers in a way that’s productive?

I think, unlike a lot of these companies that serve news as a side dish, we only have our relationship with publishers. If we take advantage of that, if we screw them, if we lose trust, we’re gone. I mean, gone. We have to have a great relationship with them. It’s not to say there won’t be challenges. And I’m sure if we’re successful, there are all sorts of questions about money.

But right now, we’re a product that’s trying what would happen — as a thought experiment, what would happen if you were a publisher first? What would happen if you had to have a great relationship in order to succeed? I’m not saying that’s all the reason of why these relationships have turned sour before, but I think it’s a big part.

The other part of this, I think, is working with publishers so far has been really fun. And it’s because — I won’t name names. There’s one large publisher, very well-known, trusted. And I got to the CEO of this place early on.

And I got an email yesterday, which was like, I just want to thank you because when I tap into our stories, it shows our login. It doesn’t remove all of our ads. I literally don’t understand why you guys are the first people to just have done that. Why was it so hard for everyone else?

Meanwhile, it was funny. I don’t know, Casey, if you saw this post this morning in social that was like, finally someone who agrees — it was about the group we’re posting about ads. And it’s like, Artifact will never be successful if it doesn’t strip out all the ads. And just maybe I’m a contrarian here, but I’m like, well, but these people have to win somehow. And by the way, Facebook’s full of ads. So is Instagram. No one’s saying those things won’t be successful if the ads are — if we just got rid of the ads on these platforms, they’d be so much more successful. No. It’s this fine balance.

What you don’t want is pop up over pop up over fixed footer over — and there are publishers who have done that, but maybe we can work with them to produce a better experience where you continue to make money, you continue to find subscribers, but Kevin, I don’t want to be that person who over promises the world and under delivers.

I just want to say, hey, we’re working on something cool. If you want to work with us, let’s do it. And I get why you wouldn’t trust a lot of us. I get it. So all you can do is take me on my word here that we want to do great stuff together, and let’s go.

Great. Yeah. I’m excited to try it. Casey did send me an invite code, so I’m going to sign up, although I mistakenly sent it to a different phone. I typed in my phone number wrong. So someone out there has an Artifact Beta link. I hope you enjoy the app.

You can do it again. It’s OK. It’s not a one-time use. You’ll be fine.

OK. Woo. Thank god. Well, I’m getting you additional users who may be random people, but I do want to ask one Instagram question. I know we’re here to talk about Artifact, but I have a question about Instagram that I’ve been dying to ask you, which is, what do you think of the app now?

I mean, when I log in to Instagram, it is almost unrecognizable from the app that I remember even from a few years ago. You’ve got reels. You’ve got shopping. It’s showing you stuff from people you don’t follow. It’s hard to find your friends sometimes. I assume you still use Instagram, this app that you built. What do you think of what it’s become?

I use Instagram every day. The thing I will say is this, these things are not static. They have to change with the times. They have to evolve. They have to move. And that doesn’t mean all the moves or the evolutions are good, but you got to go with the waves of evolution because you never know where it’ll navigate.

Now, what I’ll say is that’s a really non-answer to your question because you’re like, well, of course, but what do you actually think? Are the things about it that make you sad? Sure. I don’t want it to become this overcomplicated mess of features, but it’s funny because when I worked there and we added features people, said the same exact thing. And some of those things worked, and some of those things didn’t.

Instagram direct or messaging stuff worked really well, stories really well. IGTV, not so well, but I don’t know. Had it been slightly different, maybe it would have been a TikTok before TikTok, right?

Right.

So I’d look at it like — I assume with kids, how they are at three and how they are at seven and how they are 13 just changes dramatically. And sometimes you wish they were just that three-year-old, but you’ve just got to take it all together. And I think that Adam’s doing his best with what he has.

Adam Mosseri, who runs Instagram now.

Yeah. But man, the stuff’s hard. So without being in the chair, it’s really easy to be an armchair quarterback and say, oh, I would have done all these things differently. The one thing I know is that I’m happy to be playing in a world where I’m not — can I tell this quick story?

Yeah.

I had this meeting once with one of our executives who ran a vertical for us. And we were having this very serious conversation about Snapchat. And she goes, I just need to know. What’s our hot dog moment? And I was like, excuse you? And this person was like, what’s our hot dog moment? We need our hot dog moment.

And I literally thought at that moment, how long can I last in this job?

[LAUGHS]:

So I vaguely remember there was a Snapchat hot dog filter, right?

Yeah, and you could have your hot — it was like this AR thing where you have —

It was a dancing hot dog.

Dancing hot dog, right?

Yeah.

And I just remember thinking, is that the level of discourse and thought? We have to worry about whether or not we have a dancing hot dog in our app?

And it’s OK. I guess that was the right question to ask because everyone was talking about it, but I thought to myself — I was like, OK, not to be elitist, maybe hot dogs are interesting, but there was a moment where I thought to myself, I don’t know if I can last. I don’t know if I can do this forever. The fights I want to fight are different. And it’s not to say there won’t be similar questions about hot dogs in the future of Artifact. I don’t know.

But I wanted to work on a different product. And I’m thankful that I have some time to what I believe is focus on an area that if done well, can create meaningful value for people because they’re learning about the world, not just about, can we get an extra minute or two because you’re watching a hot dog dance? I don’t want to trivialize Instagram.

Right. In Artifact, the hot dog will — yeah, no, in Artifact, the hot dog will explain the news to you and get you caught up on areas of significant interest.

What an awesome Easter egg that would be, our little clippy is a dancing hot dog. It’s like, it looks like they’re trying to learn about something. Listen, that’s a hard — social media is just a really hard business.

But news is very easy.

[LAUGHS]:

So I’m excited. I’m not accusing you of being a glutton for punishment, but I will say, it’s an area that I think a lot of executives in tech are very excited about and then regret having ever dipped their toes into because it is such a thorny and complicated world.

But we also want people to try hard problems, right?

Totally, totally.

Can I clarify something really important here?

Yeah, go ahead.

Which is, I’m interested in news, but that is not why I started this company. Why I started this company was because I believe that machine learning can serve consumers better than anything that exists today.

And it just so happens that if you’re going to apply that thesis to the world, text we can understand really well. There’s a lot of it on the web. There’s a lot of it being produced very quickly, and people have an appetite for it.

My ambition is not to remain a news company forever. It’s actually to figure out how to take what we are building and apply it to all sorts of areas. So I get whenever you see emails, it’s like, wait, so just want to be clear. Are we’re talking a news app? No, that just happens to be where we’re starting. And I think it’s the right — there’s this book called “Crossing the Chasm.” They talk about beachheads and Jeff Moore, I want to say wrote it.

Where do you start? What’s your beachhead? And I think news, to me, is such an interesting place to start because it’s like all the concentric circles, they intersect in the right ways, but I agree. We may talk in a year, and I’ll be one of those tech execs who says, man, what did I get into?

But it’s OK. I think sometimes people talk about startups or companies as OK, you do one. And if your next one is not successful, you’re off to the elephant graveyard of Sand Hill Road, and you’re going to go invest. And there are a lot of these people try it once and say, oh, it’s too hard.

I think it’s more like making movies, where you make a great hit day one, it’s like “Star Wars” or whatever. And then you might make a couple of duds, but hopefully, if you just keep making movies, you can make great stuff.

And I just fundamentally believe machine learning is going to underpin all of this stuff. And maybe it’s not news. I think news is a great place to start, but we’re going to keep making movies. And then you can pull up this quote when I join something on Sand Hill Road, and you’re going to be like, what happened?

And I’m sure there will be a great article about it that gets served to someone via machine learning.

Sure, exactly.

If that happens.

Exactly.

But in the meantime, you can go to artifact.news. The app is available for Android and iOS, and you are letting people in off the wait list at what seems like a pretty good clip. Is that right?

Trying my best, trying my best.

Very cool. Kevin, thanks so much for coming on “Hard Fork.”

Yeah, thanks for stopping by.

Thank you. [MUSIC PLAYING]

“Hard Fork” is produced by Davis Land with help from a dancing hot dog. We’re edited by Paula Szuchman. This episode was fact checked by Caitlin Love. Today’s show was engineered by Alyssa Moxley. Original music by Dan Powell, Elisheba Ittoop, Marion Lozano, and Sofia Lanman.

Special thanks to Hanna Ingber, Nell Gallogly, Kate Lopresti, and Jeffrey Miranda. As always, you can email us at [email protected]. That’s all for this week. See you next time.

As written, it’s actually, “that’s all of this week.”

That’s all of this week. See next you time.

What if we had actually captured all of the week in those three segments? I mean, maybe we did.

[LAUGHS]: There’s nothing else of importance happening.

Yeah, look. Turn off your podcast app now. Go outside. That’s grass.

Yeah. Who needs AI-powered news apps after all? You’ve got “Hard Fork” bringing you all the news that matters.

[MUSIC PLAYING]