Meta 进入 MAGA 模式 | EP 117

I was just struck by how craven and cynical Mark Zuckerberg, in particular, was being about this. This was basically a laundry list of things that right-wing critics of social media platforms had been asking for for years, and Meta stood up and said, “We’re going to do all of it.”

I’m gay. You can now tell me that I have a mental illness. Kevin, you can go right onto Facebook and tell me that I’m mentally ill for being gay if you want. You can say that I don’t belong in the military. You can tell trans people they don’t belong in the military for other reasons.

Other reasons, and that’s important. Nothing to do with your sexuality.

No, I’m a terrible shot. There are some other changes. If you want to say offensive things about trans people, like that they can’t use the bathroom of their choice, or if you want to blame COVID-19 on Chinese people or some other ethnic group, you can just do that on Facebook and Instagram now. Mark Zuckerberg says, “Well, that’s sort of more in keeping with the mainstream discourse.” Those are the words he uses.

The standard on Facebook now is that it’s just going to feel like a middle-school playground, right? People could get hurt. People could die. I want to be very clear about that. This is not two pointy-headed intellectuals sitting in their podcast studio saying, “Oh no, Facebook isn’t a safe space anymore for the college students.” What I’m saying is that violence has been fed on Facebook before, and it will be fed on Facebook again. As a result of these changes, more people are going to be hurt.

I had kind of a disaster happen to me over this break, which was that I got robbed on Christmas.

Wait, was it the Grinch?

The citizens of Whoville are still looking for the suspect.

Who robbed you? How did you get robbed?

I wasn’t home, luckily, but someone broke into my house.

That is typically when the Grinch likes to strike.

Yeah, I got totally robbed.

What did they take?

We’re still sorting through it. We just got back, but it appears that the thief or thieves took some jewelry and some electronics. Weirdly—and this is the craziest part, and the tech angle here—they did not take the Apple Vision Pro.

Not even robbers? It makes sense, because robbers typically only want to take what is valuable, Kevin, and it’s not clear what they would actually do with a Vision Pro. Also, keep in mind, if you’re a robber, you’re out there moving through the world, breaking into homes. You can’t have that giant thing on your face. You need to maintain clear vision, so to speak.

Yes.

Let me ask you this: Even though all your items were stolen, did you look at your family and your dogs and think, “You know what? At the end of the day, I’ve got my family, and that’s all that really matters”?

I did, and I don’t know why you’re saying it with such sentimentality.

I was looking for a nice, sentimental ending, honestly.

That was sort of the moral of this robbery, much the same as the moral of How the Grinch Stole Christmas: The real Christmas, the real household items, are our families.

Exactly. If you get robbed again, maybe don’t worry about it.

Was it you?

I’m changing the subject. We’re moving on.

Okay.

Where were you on Christmas?

[Music]

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

I’m Casey Newton from Platformer, and this is Hard Fork. This week, Meta goes MAGA. We break down the company’s surrender to the right on speech issues, then why 2025 is shaping up to be a huge year in AI, and finally some HatGPT.

Well, Casey, I think we better talk about Meta.

We better do it, Kevin, because I never met a bigger story for this podcast.

Yes. The big news this week in the world of social media is that Meta is making a pretty calculated and transparent—“craven” is another word people have used—play to ingratiate itself with the incoming Trump administration by surrendering to the demands of right-wing speech critics and changing a bunch of things about the way its platforms work.

I think this is a very big story, not just because of what it represents about Meta, but because it is the biggest and most prominent example of a Silicon Valley tech company positioning itself for the second Trump term. I think it’s going to have very big implications for speech on the internet, for the rise of misinformation online, and potentially for the future of Meta itself.

Absolutely. We’ve talked about speech policies at Meta basically as long as we’ve been doing this podcast, but I think this set of changes that the company announced this week is easily the most important series of policy changes that it has made in the past 5 years.

Let’s run down what has actually been happening over at Meta. Over the past week, there have been 3 main things that people are pointing to as being part of this effort to curry favor with the incoming Trump administration.

The first was that last week, Meta’s global policy chief, Nick Clegg, a former British deputy prime minister who had served in that role for a number of years, stepped down and was replaced by Joel Kaplan.

Joel Kaplan is a longtime Republican operative going back to the George W. Bush administration who has been working at Meta in its policy division for a while now and has become the unofficial liaison between Mark Zuckerberg and the Washington right.

That’s right. Then this week, on Monday, Meta announced that it was appointing 3 new board members, including Dana White, the founder and CEO of UFC, the Ultimate Fighting Championship. Dana White is not known as a particular expert on social media governance, but he is definitely a close friend and ally of Donald Trump and someone who can presumably act as a liaison between Meta and the Trump administration.

They’re staffing that bench up with more Trump friends.

Then the big one came on Tuesday, when Meta announced that it was ending its fact-checking program and replacing it with an X-style Community Notes feature. The company also said it was redoing its rules to allow more speech and less censorship. It’s going to dial up the amount of “civic content”—that’s Meta’s term for political content and current-events content—in its feeds, and it said that it was moving its content-review operations from California to Texas to avoid the appearance of political bias.

There were some other details in there that we can talk about, including changes to the way that its automated content-moderation services will work. Basically, though, this was a laundry list of things that right-wing critics of social media platforms had been asking for for years, and Meta stood up and said, “We’re going to do all of it.”

Another way of putting it, Kevin, is that they accepted wholesale the Republican critique of Facebook’s speech policies and actually used the same words that Republicans would use. In a previous time, we only used the word “censorship” to apply to state action to prohibit speech. Some people would say it doesn’t actually apply to private companies policing online forums, but Mark Zuckerberg said, “No, effectively, you’re right. We do a bunch of censorship. We’re doing too much censorship, and we’re going to stop doing censorship.”

The reasons that Mark Zuckerberg gave, and that Joel Kaplan gave when he went on Fox & Friends to announce these changes—which was a very deliberate decision, and one that I probably don’t have to explain the meaning of to our listeners—were that Meta had been doing some soul-searching and had discovered that its former policies created too much censorship. They were going to return to the company’s roots as a platform for free expression.

I was really struck by the way that they completely backed down here. They accepted the critique, and they seemingly are terrified of what the Trump administration could mean for them and for Mark Zuckerberg personally if they do not comply in advance with everything that Republicans have said about them for years.

Keep in mind that none of these critiques are new. They were made throughout the first Trump administration, and Facebook stood up against them. They said, “We’re actually going to try to find a middle path here. We are going to try to do what we can to preserve free expression while also trying to make this a really safe and inclusive space for as many people as we can.”

In 2025, at the start of the year, Mark Zuckerberg came forward and said, “No, not anymore. We’re done with that. Everything that the Republicans have been saying about us is true, and we are going to lean into their version of what a social network should be.”

There’s been widespread debate about potential harms from online content. Governments and legacy media have pushed to censor more and more. A lot of this is clearly political, but there’s also a lot of legitimately bad stuff out there: drugs, terrorism, child exploitation. These are things that we take very seriously, and I want to make sure that we handle responsibly.

So we built a lot of complex systems to moderate content. But the problem with complex systems is that they make mistakes. Even if they accidentally censor just 1% of posts, that’s millions of people, and we’ve reached a point where it’s just too many mistakes and too much censorship.

The recent elections also feel like a cultural tipping point toward once again prioritizing speech. So we’re going to get back to our roots and focus on reducing mistakes, simplifying our policies, and restoring free expression on our platforms.

I was just struck by how craven and cynical Mark Zuckerberg, in particular, felt about this. He sounded like Elon Musk, to be totally honest. He used phrases like “legacy media” with this dripping disdain, which is a phrase that Elon Musk and his friends love to use in describing the mainstream media.

He also used the word “censorship,” which he had avoided studiously for years in describing the content-moderation work that every social network, including all of Meta’s social networks, does as a matter of business. It just sounded like a total capitulation, a total giving in to the demands of his most ardent right-wing critics.

So Zuckerberg talks about this in a Reel that he posted on Instagram. In addition to dragging the legacy media, Kevin, he also threw his own contractors under the bus. Let’s hear that clip.

Misinformation was a threat to democracy. We tried in good faith to address those concerns without becoming the arbiters of truth, but the fact-checkers have just been too politically biased and have destroyed more trust than they’ve created, especially in the United States.

He says that the fact-checkers had proven to be too biased. He gives no evidence for that, no examples. He just says that these fact-checkers, all of whom follow a very rigorous code for how they do their work, have been super biased. Who knows what that meant?

He also, as you pointed out, says that they’re going to move their moderation teams to Texas to avoid bias. First of all, I can tell you they have had moderators in Texas for many years, basically for as long as they’ve had moderators. They’ve also put moderators in red states for years. In 2019, I visited Facebook moderation sites in Arizona and Florida.

There’s absolutely nothing new about this, but he is throwing his moderators under the bus. The worst part about it, to me, is that he is suggesting that the moderators were the ones making decisions about policy, when in fact that person was Mark Zuckerberg. If Mark Zuckerberg wants to talk about the perception of bias around Facebook policy, he should reckon with the fact that he is the policymaker in chief over there.

What do you think the most impactful part of these changes is? For all of the talk about the end of the fact-checking program over at Meta, my sense is that the fact-checking program, for all the good people who worked very hard on it, really only ever touched a tiny fraction of the content shared on Meta’s platforms.

It was a pretty ragtag effort that never really had as much of an impact as I think the fact-checking community would have liked, in part because of the way that Meta restricted it. I don’t know that the average user of Facebook or Instagram is actually going to notice the fact that fact-checking has disappeared. What do you think the biggest impact on users will be?

Let me speak to the fact-checking first, because in some ways I agree with you. I rarely encountered one of these fact-checks on Facebook. On the other hand, I am someone who believes in harm reduction. Fact-checkers did look at millions of pieces of content that were getting, presumably, hundreds of millions or billions of views.

There were empirical studies that showed that, overall, people came to have fewer false beliefs if they saw those fact-checks. To the extent that people saw them, they were effective, and I think there was a case to continue doing them, particularly if you want to be a good steward of a network that you have built, that billions of people are using every day, and it’s important to you that they have a good experience on that platform and don’t come away from it stupider than when they started.

I don’t actually think that’s the most important thing that they announced, though. I’m going to point to something that Mark Zuckerberg said in his Reel.

We used to have filters that scanned for any policy violation. Now we’re going to focus those filters on tackling illegal and high-severity violations. For lower-severity violations, we’re going to rely on someone reporting an issue before we take action.

What does that mean? Whereas before Meta used automated systems to catch all sorts of things—not just illegal things, but also stuff that was annoying or hurtful, or that was a little bit bullying or harassment, like if I called you a name or a slur—Meta would catch that stuff in advance and maybe not show it to you or take some sort of disciplinary action against the person who sent it.

What Zuckerberg is saying here is, “We are not the content moderators anymore. You are, Facebook user, Instagram user. We’re now enlisting you in the fight. If you see a slur on our platform, go ahead and report that, and then maybe we’ll take a look.” I think this is a really big deal.

Yesterday, I talked to a bunch of people who either work at Meta or used to work there. One person told me they were extremely worried about what this meant because they had seen, in so many countries around the world where Meta has traditionally done much worse moderation than it does in the United States, that by not taking action against these lower-severity violations—stuff that was not obviously illegal—they had seen violence fomented again and again. They had seen harassment against women and abuse of LGBTQ people.

Zuckerberg said in his Reel, “Look, we are going to have more bad stuff on the platform,” but he didn’t take the second step and explain what that actually means. What it actually means is that people could get hurt. People could die.

I want to be very clear about that. This is not two pointy-headed intellectuals sitting in their podcast studio saying, “Oh no, Facebook isn’t a safe space anymore for the college students.” What I’m saying is that violence has been fed on Facebook before, and it will be fomented on Facebook again. As a result of these changes, more people are going to be hurt. That, to me, is the biggest consequence of these actions.

This reporting thing that you bring up is so interesting because, as we know, a lot of the worst stuff on Facebook happens in groups, in semi-private spaces with hundreds or thousands of members. Meta is essentially saying that it will be up to the members of those groups to report any violative content that they want moderated, rather than having these proactive scanners going around.

You might say, “What’s the big deal about that?” If you’re in a Stop the Steal group, a QAnon conspiracy group, or a group planning an insurrection at the Capitol, which members of that group are going to report each other for violating Facebook’s rules? I don’t think that’s a thing that’s going to happen.

I think what we’re going to end up with is a much more unmoderated mess over at Facebook, Instagram, and all the other Meta platforms.

When I was talking to employees this week, one of them pointed out what a strange step backward this is. For so many years, Mark Zuckerberg bragged about how automation was the future of content moderation. He boasted about the systems they were building that were getting better every single quarter at detecting hate speech and bullying and making this a better place for his community.

Now, instead of saying, “We’re going to lean into this even more and make these filters better,” he said, “We’re going to stop using them and go back to human beings who don’t even work for us or have any training or expertise.”

This is an abandonment of his technological project in favor of something that is obviously inferior. To me, that is one of the big twists here: Mark Zuckerberg walking away from the very good technology that he built.

What else in these changes caught your eye?

Some of our listeners, Kevin, may use Facebook or Instagram and just wonder what it’s going to be like now that these changes have been made. I thought it might be good to go through some of the offensive things that you could now say on Facebook and Instagram and not get in trouble.

For example, I’m gay. You can now tell me that I have a mental illness. Kevin, you can go right on Facebook and tell me that I’m mentally ill for being gay.

You can say that I don’t belong in the military.

You can tell trans people they don’t belong in the military.

For other reasons.

Other reasons, and that’s important.

Nothing to do with your sexuality.

No, I’m a terrible shot.

If you want to say offensive things about trans people, like that they can’t use the bathroom of their choice, or if you want to blame COVID-19 on Chinese people or some other ethnic group, you can just do that on Facebook and Instagram now.

Mark Zuckerberg says, “Well, that’s sort of more in keeping with the mainstream discourse.” Those are the words he uses: “in keeping with the mainstream discourse.”

I look at that and think, “The standard on Facebook now is that it’s just going to feel like a middle-school playground.” All this stuff is what I used to hear when I was 12 years old in Washington Middle School. Maybe not the trans-bathroom stuff—that was still yet to come—but everything else I heard in seventh grade. That is the new standard that Mark Zuckerberg has set for his properties.

He’s saying, “I would like the discourse on my platforms to more closely resemble the dialogue in a Borat movie.”

Which is satirical in the Borat case, but very serious here. It’s easy for me to joke about someone telling me I’m mentally ill for being gay. I can handle that. But if you’re 14 years old and queer, and people in your high school are calling you that on Instagram, we’ve seen over and over again that these kids harm themselves.

One of the things I find so crazy about this series of decisions, Kevin, is that 41 states and the District of Columbia are suing Meta over the terrible child-safety record it has on its platform. My understanding is that these changes apply to younger users just as they apply to everyone else.

The classifiers that once tried to find bullying, abuse, and harassment against young people are no longer going to be automatically enforced. It is going to be up to, I guess, the other kids in school to say, “Hey, it looks like my friend is being bullied over here on Instagram.” That seems like they’re opening up a huge amount of liability for themselves.

It’s not just right-wing culture warriors who have complained about excessive moderation on Meta platforms. People on the left complain that their pro-Palestinian speech is being targeted for takedowns.

Those are not phony complaints, by the way. It is absolutely true that Meta has over-enforced in some cases.

What’s so interesting, as I hear you explain the details of some of these changes and how they’re revising their rules, is that they all seem to be pointed in one direction. It’s like, let’s let people on the right mock people on the left in more ways.

Absolutely. I wrote in my newsletter that a younger and more capable version of Mark Zuckerberg truly did handle this differently. He would have said, “We’re over-enforcing in this way. Let’s improve the classifier. Let’s adopt a technological solution to this problem.”

What they said this week is, “We’re done trying to fix any of it. We’re just abandoning the project altogether.”

That’s a lot about what these changes are. I want to talk now about why they were made.

There’s an obvious explanation—the one that has been popular among the critics I’ve been reading and talking to over the past couple of days—which is the craven political-opportunism angle. This is Mark Zuckerberg’s attempt to ingratiate himself with the Trump administration. It’s all business, all strategy, all cynical, and probably all temporary until the next administration comes in.

What do you make of that explanation for why these changes were made?

I think there’s a lot of truth to it. I think another factor is that trying to be a good Democrat just didn’t really get Mark Zuckerberg anything.

After the 2016 U.S. presidential election and the huge backlash against Meta in particular that it created, Zuckerberg tried to say, “Whoa, whoa, whoa. I hear that you’re super mad. I’m going to try to fix this.” They went out and built all these fancy machine-learning classifiers to try to improve the service.

At the end of the day, I don’t think Democrats liked him even 1% better than they did before he did any of that. You have to remember that, at the end of the day, politics is transactional. People vote for people they think they can get things out of.

By the end of 2024, I think it was very clear to Mark Zuckerberg that he truly was not going to get one thing out of the Democrats. Then along comes Donald Trump, who has this interesting relationship with Elon Musk. Elon Musk used to be a liberal guy with a bunch of standard liberal positions, but then he changed his views for whatever reason, gave a bunch of money to Trump, and Trump said, “Hey, I like this guy. I’m going to give him every political advantage that he wants.”

Mark Zuckerberg is a pretty smart guy, and he thought, “Maybe I could do the same thing.”

I think the one thing we know about the values of Mark Zuckerberg and Meta is that they’re an extremely efficient organism at self-preservation. They will do anything to stay relevant and stay ahead. They will copy features. They will change the name of the damn company.

We know that Mark Zuckerberg’s own views on speech are very flexible. They tend to shift as the political winds shift. I also think there’s another potential “why” here, which is about Mark Zuckerberg personally and his own shifting political allegiances.

I’ve been talking recently with some people who know Mark Zuckerberg or who have worked with him in the past. What they’ve said to me is that this is a man who is following a very conventional former-Democrat-turned-Republican arc.

He is 40 years old and approaching middle age. He’s very into male-coded hobbies like mixed martial arts. He spends a lot of time talking with Joe Rogan, hanging out with Dana White, and immersing himself in this kind of manosphere outside of work.

He’s also been the target of a lot of criticism, especially from the left. One thing we know about successful men who get targeted by left-wing opprobrium is that they often respond by becoming disaffected former liberals who embrace the right because they feel they’re getting fairer treatment there.

I can’t prove this theory, but some people who know Mark Zuckerberg have suggested to me that he has actually become personally quite red-pilled or conservative over the last few years.

Obviously, he’s not Elon Musk. He doesn’t broadcast his political opinions on social media dozens of times a day. He has been more careful about signaling which team he’s on. But I offer this as a theory because I think we’re starting to see more evidence that his own views may have shifted quite a bit, independently of what’s good for Meta.

I think there was a version of all this that was less extreme. If Zuckerberg himself were truly liberal or progressive in his heart, we would not have seen these changes. So I do think the changes they announced this week offer some evidence for what you just said.

My colleagues Mike Isaac and Teddy Schleifer reported last year that Mark Zuckerberg has begun referring to himself as a classical liberal. If you’ve ever watched a right-wing YouTube video, that’s what every former liberal who has now become a Republican says. They call themselves classical liberals. I’ll just put that out there: It’s a code word.

Do you think we’re going to see an exodus of liberal and progressive users from Meta platforms the way we did from X after Elon Musk took it over?

It depends on how all of these changes play out, and we’re just not going to know for a while. My assumption is that Meta will continue to do a significantly better job at moderation than X does. It’s a much bigger company with more infrastructure in place, so I don’t think you’re going to get the overnight transformation you got with Elon Musk.

Facebook and Instagram are also structured very differently from X. Zuckerberg can’t really take over those platforms in terms of the actual posts you see in the feed in the same way Elon does.

On the other hand, if Facebook and Instagram truly come to feel like seventh-grade playgrounds at recess, and the discourse gets much rougher and coarser, I do think you’ll see people walking away from them.

While we almost only ever discuss content moderation in terms of its politics, the truth is that there’s a huge commercial demand for it. People do not want to spend time on networks that are full of violence, harassment, abuse, gore, and pornography. That is the main reason all of these companies build systems to remove or suppress those things.

The real question, I think, Casey, is how far Zuckerberg ultimately goes in this direction. Whatever the politics might be, the vast majority of his users just want a safe and friendly place to hang out online.

That’s where we are with Meta today and with some of the implications of these changes. Do you have any more predictions about where this will all head?

I have a fun one for you, Casey. Meta has told its partners in the fact-checking partnership that it has funded for the past several years that their contracts will end in March. In March, the fact-checks on Meta properties are going to end.

The Community Notes product that Meta is planning to build, which is essentially a volunteer content-moderation system, is going to take a little longer to build. That means you and I can look forward to a fact-free spring on Facebook.

Let’s go. We can truly say the craziest things, and not one person is going to be able to stop us. Let me just say, I’m cooking up some whoppers. The things I’m about to say on Facebook and Instagram—let’s just say you’re going to want to follow me.

Follow Casey over at Threads.

Start piling up the drafts now. The purge is coming, and you’re ready.

I’m ready for the purge.

[Music]

When we come back, everything you missed over the break in AI. There’s a lot.

Casey, we have more news from over the break about one of our favorite topics: AI. It was a huge couple of weeks for AI, Casey, during a time of year when normally the news cycle gets pretty slow.

I was wondering about that. Usually in December people are getting ready to go on holiday break, and the news kind of trails off. But not this year. The AI labs were trampling all over each other to get their big news out before the end of the year.

I think it was led by OpenAI, which announced its “12 Days of Shipmas,” where it tried to announce something—something big, something small—every day for 12 days. It wound up ending on something pretty important.

There’s a lot to catch up on today, and I want to take some time to dig into what happened and what we can expect for the first few months of the new year. But before we get into all that, Casey, you have something to tell us.

I do. Kevin, our listeners’ trust is of paramount importance to us, and I wanted to let folks know about something that happened in my life that I want to be upfront about.

At the end of 2023, I met a man who had many wonderful qualities. One of those qualities that I loved was that he worked for a company I had never heard of, which meant I could keep doing my job as normal. But as of this week, my wonderful boyfriend started a job at a company we talk about sometimes on the show. He is a software engineer at Anthropic.

Is his name Claude?

Many people have written to me asking if I fell in love with Claude. While I do find Claude to be very useful for some things, no. This was a human man that I am currently in love with. I’ve met him. He’s real. I can confirm that he’s wonderful.

You’re disclosing that you have this new—let’s call it an entanglement—because this is a company that you and I talk about and that you also cover in Platformer. We wanted our listeners to know that this is happening out in the world and in your life. Is there anything more you want to say about this?

People have some questions about this. I did not play any role in my boyfriend getting this job. Anthropic didn’t know about our relationship before this happened. Of course, we have since told them about this.

I do plan to continue writing and reporting about Anthropic because I think it’s a really important company, but whenever I do that, I’m going to remind you that this relationship exists.

A couple of other things I would say: My boyfriend and I do not have any financial entanglements, and we do not currently live together. I’m also going to commit to updating folks as that changes.

Basically, I’m going to try to do the same job that I always do and bring the same skeptical, critical eye that I bring to everything. But I’m also going to remind you that I have this relationship.

If you have questions about that, email the show at hardfork@nytimes.com. I’ll try to answer any respectful questions I can about this.

I’ll editorialize and add a little bit here to your disclosure, which I think is laudable. I’m glad you’re doing it, and I’m glad you did it in your newsletter and on the podcast.

I’ve known you for a long time. I’ve known how hard you have tried to avoid dating men who work in the technology industry. For more than 10 years, you would be on apps like Tinder and see that somebody cute worked at Google, Meta, Twitter, or one of the companies you cover, and you would always swipe left because you thought, “I don’t need that drama in my life. I don’t need that complication.”

Which is tough in San Francisco because everyone works in tech.

It’s a very small town, and the number of eligible bachelors out there who do not work at one of the companies you cover limits your dating pool considerably.

It really did. It sort of explains why I was mostly single for the last 10 years. I thought I had finally found something that got me out of that situation, but sometimes life has other plans for you, and you have to roll with the punches.

So here you are.

Here I am.

Thank you, Casey, for that disclosure. Transparency is very important. We’re obviously going to keep talking about developments in AI at Anthropic and elsewhere, but we’ll also include this disclosure in the way we do when we talk about OpenAI and the fact that The New York Times Company is suing OpenAI and Microsoft, alleging copyright violations.

When I disclosed this in my newsletter this week, one reader replied that they thought it was cute that I would now have a disclosure to go along with your disclosure that you do every week. We’re now one for one.

Let’s proceed to the real meat of this segment, which is the AI news. So many things happened.

Let’s start with OpenAI. We’ve already made the disclosure, so we don’t have to do that one again. This was a big month for OpenAI. On December 20, just before we headed out for the break, they announced a new model called o3. This was a successor to o1.

Funnily, they skipped o2 in the naming process because of a lawsuit threat from O2, the telecommunications company.

I’m not sure if it was a threat. They said they did it out of respect.

Presumably, there would have been some sort of legal problem. They skipped right over o2 to o3. This model is not yet available to users, but they gave a preview of it to some researchers and talked about how it had performed on some benchmark evaluations.

Casey, tell us about o3. What is o3?

O3 is a large language model, like the kind you would already find in ChatGPT, but it’s built in a different way. It’s known as a reasoning model.

The reasoning models are different in a couple of ways. The first is how they’re trained. They’re trained to be better at handling logical operations and structured data.

The second big difference is that, when you make a query—when you type into the little box whatever you want it to do—the reasoning model takes longer to go over it. It uses more computing power, takes multiple passes through the data, and really tries to bring true reasoning to what it’s looking at.

The result of taking more time, doing more passes, and being structured in a slightly different way is that it can perform much better on very complicated tasks. What OpenAI found with o3 was that it was able to get much further on some of the hardest benchmarks ever designed for LLMs than anything that had come before it.

We talked a little bit about this idea of test-time inference, or test-time compute, when we discussed o1, their previous reasoning model. This is a different step from the classic pretraining step of building a large language model.

Something happens when the user makes the query. Instead of just spitting out an answer right away, it goes through this secondary test-time step. Researchers were very excited about this when o1 came out. They thought, “Maybe we’re tapping out the limits of the pretraining step. Maybe there’s a new scaling law developing around test-time or inference compute.”

If we pour more resources into that step, perhaps the models will get better along a different axis. What people were excited about when o3 came out was that it looked like that had actually worked.

This stuff is not yet in the hands of everyday users, but OpenAI entered the o3 model in a fascinating public competition known as the ARC Prize. You know the ARC Prize, Kevin?

The basic idea is that they try to come up with problems that would be insanely difficult for an LLM to solve. One reason they’re difficult is that they’re original problems. These problems are not in the training data of any of these models.

One criticism of LLMs is essentially, “You already have all that data stored. You just did a quick search.” This prize says, “No, we’re not going to let you search. You’re going to have to show that you can reason your way through something really difficult.”

The ARC-AGI-1 public training set has been around since at least 2020. At that time, GPT-3, OpenAI’s previous model, got a score of 0%.

Just 4 or 5 years ago, we were at 0%. In 2024, GPT-4 got to 5%. With o3, it got to 75.7% in one evaluation where the limit was that you could spend only $10,000 on computing power.

In a second test, where they let OpenAI spend as much money as it wanted—which we think was more than $1 million—o3 hit 87.5%. Something that was essentially impossible through all of 2024, almost instantly, reached 87.5% of that benchmark.

That is essentially the only public data we have about how good this thing is, but it got people’s attention.

It got people’s attention. I also saw a lot of people paying attention to o3’s performance on something called Codeforces. This is a programming-competition benchmark and one way these AI companies try to assess how good their models are at coding.

OpenAI o3 received a rating on Codeforces of 2,727. That is roughly equivalent to the 179th-best human competitive coder on the planet. For context, Sam Altman, in presenting this result, mentioned that only 1 programmer at OpenAI has a rating higher than 3,000 on Codeforces.

Why does this matter? Think about some of the discussion happening at the end of 2024. You started to hear people say, “We are hitting a scaling wall.” The idea was that the techniques we used to build previous LLMs were running out of low-hanging fruit, and it would require some sort of conceptual breakthrough for them to continue improving.

O3 comes along and effectively does just that. What I think is important about these benchmarks, and why we want to spend some time going through them, is that there’s a lot of justified criticism about how much these things are being hyped. We know the companies love to hype their products and tell us how incredible they are.

But the benchmarks are something objective that you can use to measure performance. When a benchmark says there is now a model better than all but 179 people on Earth, it seems like we might be getting pretty close to superintelligence. What is superintelligence, if not a system that is better than every human at something?

I would add a caveat. These so-called reasoning models seem, from what we know about them so far, to be very good at the kinds of tasks for which you can design what are called reward functions.

Those are things that have a definite right answer. Either the code runs or it doesn’t. Math has a definite right and wrong answer. In domains where you can give the reinforcement-learning model a goal and an indicator of whether it is right or wrong in pursuing that goal, it tends to do very well.

If you asked it what the meaning of true love is, it would never know. It wouldn’t know the first thing about it, and I think that’s beautiful.

For the short term—the next year or 2—we’re going to have these early reasoning models that are very good, and potentially even superhuman, at some tasks. Those are the tasks that have definite right and wrong answers. For other things, like fiction writing, life coaching, or tasks that don’t necessarily have 1 right and 1 wrong answer, they may not advance much beyond what we see today.

Some people will use that as an excuse to say this doesn’t matter that much. I would point out that, at some point in your life, you’re probably going to see a surgeon who might not be a great painter. That doesn’t change the fact that the surgery you received was very valuable.

It’s important to think more in terms of what these things are capable of in the moment than what they are not capable of.

The other AI story we should talk about quickly is that Sam Altman wrote a new blog post on January 5 called “Reflections,” basically talking about his thoughts about the 2 years since ChatGPT was released.

The big headline from this blog post is that Sam Altman is claiming that OpenAI now knows how to build AGI—the artificial general intelligence that people have been speculating about for years. OpenAI has been hinting that it is within sight of that goal, and Altman believes it could happen very quickly. They’re already starting to look past AGI to ASI, artificial superintelligence.

What did you make of this blog post?

I spent basically a day trying to figure out exactly what Sam meant when he said they know how to build AGI. Another thing that happened this week, Kevin, is that Sam did an interview with Josh Wingrove at Bloomberg.

One of the things he told Josh was, quote, “I don’t have deep, precise answers there yet, but if you could hire an AI as a remote employee to be a great software engineer, I think a lot of people would say, ‘Okay, that’s AGI.’”

My interpretation, based on conversations I had this week, is that this is actually the destination everyone has in mind for 2025. This is where the race is going. You are going to see all the big AI labs race to release a virtual AI coworker.

If they can do that, and if the coworker is pretty good, they’re going to say, “This is actually what AGI is.” At the moment, you can hire a virtual entity to do some task or series of tasks in your company that you no longer need a person for. That is where this entire thing has been driving the whole time.

I agree, but it is not necessarily something we need to accept uncritically. Sam Altman is a person with his own goals and motives. OpenAI has its own reward functions, and we should perhaps apply some discount to what he says about his projections for AI because he has a vested stake in the outcome.

But we should also use this as a way of taking the temperature of what conversations are happening in the AI scene in San Francisco. People here are very sincere and genuine about the fact that they believe we are going to get AGI, or something like it, very soon—possibly this year.

When you look at the improvement in these models that we saw in December alone, I think you have to take them seriously.

Moving on from OpenAI, another thing that happened in December is that Google released Gemini 2.0, the new version of its flagship AI model. Casey, have you tried it yet? What do you make of it?

I have not tried it yet, Kevin, because it is not in the consumer-branded Gemini that I pay for, with the exception of a new feature called Deep Research. You can ask Gemini to go read the web and prepare a little report for you about something. I’ve used it only 1 time, and it seemed okay.

To be candid, I have not followed the 2.0 stuff as closely because it has not seemed as shocking or impressive as the OpenAI stuff. Have you?

I’ve played around a little with Gemini 2.0, mostly in a series of demos I got at Google before it came out. Some of what has been included is catching up with other models.

Google also released Gemini 2.0 Flash Thinking Mode, which was its first attempt at an inference-time-compute reasoning model similar to o1 and o3 from OpenAI.

I have not played around with Gemini Deep Research Mode yet, but I’ve heard people talking about how cool it is. People whose judgment I trust say this is basically Google announcing that it is on the same trajectory as OpenAI and the other companies that are its peers and rivals.

It is going to be scaling up very quickly in 2025, and we should look forward to more.

There was a post on X that went viral this week where someone asked Google, “Does corn get digested?” All of the image results were AI slop that appeared to be diagrams of corn and made no sense whatsoever.

It was extremely funny. Maybe it will be patched by the time this comes out, but if not, do an image search for “Does corn get digested?” and you’ll get a sense of where Google’s AI search skills are.

In conclusion, Google is cooking in the AI department, but not much of this has gotten into consumers’ hands yet. I think that will be the question for 2025: Is this stuff actually as good as Google says it is?

The third and final story we’re going to catch up on today from over the break is something out of a Chinese company called DeepSeek.

DeepSeek is a Chinese AI company. It is actually run by a Chinese hedge fund called High-Flyer. Right around Christmas, as my house was getting robbed, they released a new model called DeepSeek-V3 that ranks up there with some of the world’s leading chatbots and caught a lot of people’s attention.

I have not used this one yet, but there are a few things to know about it. One is that it’s really big. It has more than 671 billion parameters, which makes it significantly bigger than the largest model in Meta’s Llama series.

Up to this point, Llama has been the gold standard for open models. The largest one has 405 billion parameters.

The really important thing about DeepSeek is that it was apparently trained at a cost of $5.5 million. That means you now have an LLM about as good as the state of the art that was trained for a tiny fraction of what something like Llama or GPT was trained for.

I saw speculation from the great blogger Simon Willison that the export controls the United States is placing on chips are actually inspiring these Chinese developers to get much better at optimizing. You now have this state-of-the-art model for $5.5 million. This is a huge step toward the proliferation of LLMs everywhere.

Let me back up and go a little more slowly through what you just described, because I think it’s really important.

One of the big questions over the past 5 or so years has been about the Chinese AI industry and where it is relative to the leading frontier AI labs in the United States, whether we need to do more to slow it down, and whether we even can slow it down.

One view is that this stuff is common knowledge: As soon as someone invents a new way of doing AI, it spreads throughout the world, and there’s not much you can do to stop it.

In the United States, we passed something called the CHIPS Act, along with a set of controls that limited which AI chips could be exported to China. We put a lot of faith in the ability of these restrictions to constrain the Chinese AI industry. If China couldn’t get the latest chips from Nvidia and other companies, it wouldn’t be able to build models competitive with the state-of-the-art U.S. models. That was one way we were going to try to keep our national advantage.

What DeepSeek has shown, or at least hinted at, is the possibility that China is not that far behind. Whatever you think about this model—I have not tried it myself—according to its benchmarks, it is up there in many respects with the latest and greatest models from OpenAI, Google, and Anthropic.

By some measures, it is the highest-ranking open-source or open-weights model we have. It does not appear to have needed the latest and greatest hardware to be trained.

According to the report that DeepSeek put out, it trained V3 at an estimated cost of about $5.5 million. It did so not on the leading-edge Nvidia H100 or A100 chips that all the big AI labs use, but on a different version of Nvidia chips known as the H800, which is basically a less capable version of the state-of-the-art chips from Nvidia.

I think this boils down to the conclusion that regulating AI by limiting access to hardware is going to be much more complicated than we thought. One interpretation is that you can’t stop China from building state-of-the-art foundation models, and our regulatory regime is not going to be enough to keep the United States ahead of China.

What do you make of that?

The first thing I would say is that I get nervous when people frame the debate this way. A lot of the people who frame the AI story as a race between the United States and China are very hawkish, and they’re leading us toward a potential conflict that I would rather avoid.

It also presupposes that American companies have to race as fast as they can and build AGI as fast as they can, even if that means cutting corners on safety, because of this looming specter of China and everything that could happen.

I would say that we don’t necessarily have to do that. We can choose to move somewhat deliberately and with caution here.

Do I think this shows that it is going to be harder to prevent China from developing extremely high-end models, and that regulation is going to be more complicated? Yes, absolutely.

That is a small fraction of what happened in AI while we were gone, but probably the most important things. I think we covered most of what really mattered.

If there’s 1 thing we can be sure of in 2025, it’s that we’re going to be very busy talking about more AI changes and progress.

Somebody was telling me that if 2023 was the year that made everybody say, “Oh my gosh, AI is going so fast,” and 2024 was a year that felt very business as usual, 2025 could be a year when we go back to, “Oh my gosh, AI is going so fast.”

Maybe it’ll just feel like that all the time, forever.

Isn’t that a pleasant thought?

Happy New Year. AI vertigo forever.

When we come back, 2025’s first game of HatGPT.

[Music]

From time to time, we like to check in on some of the wilder headlines from the world of tech in a segment we call HatGPT.

We take headlines, put them into a hat, fish headlines out, discuss them for a bit, and when one or the other of us gets bored, we simply say, “Stop generating.”

We haven’t done a HatGPT in a while, and there’s been so much that I’m excited to see what’s in the hat.

Me too. Why don’t you get us started?

I’ll pick first.

This one is called “Meta Kills AI-Generated People Like ‘Proud Black Queer Mama.’” It’s from Futurism.

This was sparked by an interview given by a Meta executive in the Financial Times at the end of 2024, basically talking about plans to let users create a bunch of AI profiles and fake people and get them to share generated content on Meta platforms.

People then began discovering the existence of older AI-generated profiles that Meta had started up in 2023. Washington Post columnist Karen Attiah posted on Bluesky about one AI-generated profile in particular, described as a “proud Black queer mama of 2 and truth teller” named Liv.

Karen started chatting with this chatbot and then posted her chat on Bluesky. Meta summarily killed Liv and many of its other older AI personas.

This whole thing was so silly, and I think there’s been a lot of backlash against Facebook over it. This is truly a case where you wonder why they’re doing any of this.

The answer is probably that they saw Character.AI have some success by letting people chat with different kinds of characters. But Character.AI succeeded by letting you pretend you were talking to Luke Skywalker or Spider-Man—characters that were personally meaningful to you.

Meta just made up a bunch of essentially generic humans and said, “Go nuts.” It had them say generic things, and it felt incredibly creepy to people.

This is an idea that needs to be taken out back and dispensed with. Meta is not giving up on the idea of AI-generated personas, though. It has signaled that it intends to put more AI-generated personas inside all of its apps.

I’m fascinated to see what fresh horrors emerge.

Here’s what I hope: At some point, Meta will be able to detect when you’re harassing or abusing someone—which is now allowed under its new rules—and route you to an AI so that the AI can absorb all of your prejudice and bigotry.

That might be a nice solution.

I like that. An AI punching bag.

Stop generating.

I feel like normally, when it’s my turn to pick, I get to shake the hat, but for some reason this week you’ve decided you want to shake the hat.

I’m just going to shake the hat. It’s my right.

All right.

Here’s one: “Apple Agrees to Pay a $95 Million Settlement in a Siri Privacy Lawsuit.” This is from Chris Velazco at The Washington Post.

Apple has agreed to end a 5-year legal battle over user privacy related to its virtual assistant Siri, with a $95 million payout to affected customers, according to a preliminary settlement.

Apparently, Siri was a bit overzealous in listening for wake words like “Siri.” When it thought it was being called into action, it would start recording audio it wasn’t supposed to. A number of those clips somehow ended up in the hands of third-party contractors.

Back in 2019, The Guardian reported that Apple contractors regularly heard confidential medical information, drug deals, and recordings of couples having sex.

If a judge signs off on the settlement, anyone who qualifies can submit a claim for up to 5 Siri-enabled devices, with a maximum payout of $20 per device.

Would you be willing to let Apple listen to you have sex for $100?

I’d go for it.

No, I don’t think my price is that low.

Casey, I saw this making the rounds because people said, “Finally, they’re admitting that they listen to you through the microphone on your iPhone,” which has been a favorite conspiracy theory for years, including among critics of Meta.

There’s no proof that this is an omnipresent listening system that was listening when it shouldn’t have been. What this seems to be saying is that Siri obviously needs to listen ambiently in order to tell when a user says, “Hey, Siri.”

I’m sorry if we just woke up Siri on your iPhone and you’re no longer listening to this podcast because I said that.

It sounds like Siri was miscalibrated, so it was listening more than it needed to in order to hear the wake word, or recording more audio than it needed to.

I don’t care about the actual incident, Kevin. In the 14 years that Siri has existed, I think it has correctly understood me about 4 times. This is not a technology that ever knows what I’m talking about.

Siri could take an hour-long recording of me and have no idea what to do with it. What I care about is that this is going to fuel the most annoying conspiracy theory in tech, which is that all the tech companies are secretly listening to you.

We’re going to see a lot more conspiracies around this. It is unfortunate because, again, this is only Siri we’re talking about. It doesn’t know anything.

It’s not that serious.

Stop generating.

This one is from The Athletic: “Netflix’s WWE Investment and the Future of Live Events on the Platform: ‘We’re Learning as We Go.’”

Starting January 6, WWE’s popular weekly wrestling show Raw will stream exclusively on Netflix in the United States. This is part of a decade-long agreement worth a reported $5 billion.

Casey, as Hard Fork’s resident WWE fan and expert, why don’t you take this one?

Well, Kevin, I mean, did you watch?

No, I did not.

You missed something huge. Roman Reigns beat his cousin Solo Sikoa in a Tribal Combat match, winning back the Ula Fala and becoming the one Tribal Chief of World Wrestling Entertainment.

Is that true?

That is all true. It was a great match and a really fun show. WWE positioned this as a huge thing for them, and it is. It’s also huge for Netflix.

From WWE’s perspective, it can now be in something like 280 million homes around the globe. For Netflix, this is an opportunity to experiment with live programming, which it has been dipping its toes into.

There’s a lot of speculation about whether Netflix might soon go after more traditional sports. Maybe it wants a big football deal or a big baseball deal. I’m very interested to see how these things work together, and I’m very interested to see who Cody Rhodes will be fighting at WrestleMania this year.

I saw the Jake Paul–Mike Tyson fight that was on Netflix. On Christmas Day, Netflix also had some live football. Do you think this is hastening the death of cable TV, or was that already happening and this is just Netflix trying to pick up the pieces?

Absolutely. In addition to WWE, I watch another wrestling promotion, AEW. The reason I had my YouTube TV account, which cost me something like $80 a month, was so I could watch AEW programming, because it was only available on cable.

Guess what? AEW started streaming on Max, so I was able to cut the cord once again. Now I am fully streaming again.

As these live events with intense, weird fandoms move from traditional cable to streaming, it absolutely becomes a moment when more people cut the cord.

This is a little bit of a tangent, but I had an interesting moment over the break. We were stuck in a motel in Lake Tahoe, and the iPad we use to entertain our child had run out of battery. I turned on the hotel TV and tried to explain the concept of linear TV to my 2-year-old son.

It blew his mind.

I said, “On this screen, you can watch Bluey sometimes, but not all the time. You can’t pick a specific episode, and about twice an episode they’re going to interrupt it to try to sell you toys.”

He was so confused by the concept of linear TV that I thought this industry probably does not have a long time left.

Your child knows.

Yeah.

Stop generating.

This was a fun one. The YouTuber MegaLag posted a video on December 21 titled “Exposing the Honey Influencer Scam,” and ever since, YouTube has been overtaken by discussion of what Honey did.

In the world of YouTube creators, this was probably the big news story of the year. I don’t think I’ve heard much about it outside of YouTube because of the way that insular platform works, but this was essentially a massive scandal among major YouTubers over the holidays.

Maybe we should explain what happened for people who are not glued to YouTube 24/7.

We should. Honey is a company that was acquired by PayPal a while back. It is a browser extension. Before you check out online, before you make an online purchase, you click the Honey button and Honey scans for the best coupon.

Honey went to a bunch of YouTubers and signed deals with them, asking them to promote Honey. These coupon codes are a big part of the creator economy. We’ve talked on this show about affiliate links. A lot of the internet is built on companies that sell things giving a little kickback to people who talk about their products.

Before we say what the allegations against Honey are, we should set the scene for people who are not YouTube heads. Honey may have been the most prominent advertiser on major mainstream YouTube channels.

I would say Honey sponsorships propped up YouTubers and YouTube content creation in a similar way that online mattresses propped up the podcast industry for a couple of years.

Major YouTube influencers, including David Dobrik, Emma Chamberlain, the Paul brothers, and Marques Brownlee, had major deals with Honey to underwrite their channels. They were basically ubiquitous. It was hard to watch a lot of YouTube a couple of years ago without running into Honey ad after Honey ad.

What are the allegations that MegaLag publishes?

There are 2 things. One is that, hidden in plain sight on Honey’s website, Honey will go to online retailers and charge them money to keep their best codes out of the Honey database.

Let’s say you have an online store and a crazy 80%-off coupon. Honey will say, “Pay us some money, and we’ll make sure no Honey user ever sees that coupon code.”

Honey is straightforward about that, but it’s obviously a terrible user experience.

The way Honey works is that there are coupon-code sites where you can look up codes before you buy something and try to find a 10% or 20% discount. Honey goes out and scours the internet for these codes for you, then automatically applies them to your purchase in your browser for basically any e-commerce website that uses them.

That’s right. If that had been all Honey was doing, this wouldn’t have been a scandal. The second allegation from MegaLag was that when people saw products in influencer videos and went to buy them, those shopping carts would often have the creator’s affiliate link inserted.

The creator would then get a kickback, which is the whole point of creators working with companies that share affiliate links. The allegation is that Honey would go in at the end of this process and replace the creator’s affiliate link with Honey’s affiliate link.

Honey got to keep all of the affiliate revenue and cut the creators out of the process.

Let’s walk through this step by step. I’m watching a major YouTuber’s video. Let’s say I’m watching the Hard Fork channel, and we have an online mattress company in our videos. Every time you buy a mattress and enter the code “Hard Fork” at checkout, you get 10% off.

The allegation is that, in instances where a user went to buy a mattress through our affiliate link, if they used Honey in their browser, Honey would find that affiliate link and replace it with the Honey affiliate link. Instead of getting a kickback on that sale ourselves, the money would go to Honey.

That is exactly right. People are quite mad about this. There’s a channel called LegalEagle that is suing them, which I know nothing about, but I have to say it sounds exactly like what a YouTube channel named LegalEagle would do: sue one of its advertisers.

When The Verge asked PayPal about all of this, PayPal said, quote, “Honey follows industry rules and practices, including last-click attribution.”

I take that to mean that the industry rules and practices are horrible, and Honey is not doing anything to improve them.

This was really a case where creators took a look at the situation and said, “I don’t think so, Honey.” That’s a L’Accord reference.

I would say this is a case of people being naive about how the internet works. Honey was a very popular and profitable company—so profitable and popular that PayPal acquired it—and YouTubers thought they were providing coupon codes to people out of the goodness of their hearts.

Bless your heart if you thought that was what Honey was about.

YouTubers are telling Honey to mind its own beeswax.

With that, I’ll stop generating.

This is the last one: “Tech Entrepreneur Nearly Misses Flight After Getting Trapped in Robotaxi.” Passenger Mike Johns was reportedly riding in an autonomous Waymo car on the way to the Phoenix airport when the vehicle began driving around a parking lot repeatedly, circling 8 times as he was on the phone seeking help from the company.

Did you see this video?

I did. It was wild. He initially believed it was a prank, he told The Guardian. Then he got on the phone with a support person at Waymo while he was inside this car that was circling the parking lot and would not let him out. As a result, he almost missed his flight.

This is every Waymo support person’s fantasy: One day, someone picks a random Waymo and drives it around in circles in a parking lot with no explanation. Maybe they’re teaching their kid how to drive or something.

This would obviously be disconcerting, but if I made a list of the 10 worst things that ever happened to me in an Uber, driving around in a circle 8 times would not make the top 10.

I’ve almost missed my flight several times because Uber drivers thought they knew a better way to the airport.

We shouldn’t make light of this. People are placing their lives in Waymo’s hands when they get into one of these autonomous cars. I saw people saying, “This is why I would never trust a self-driving taxi.”

It’s worth taking these incidents seriously. At the same time, no one was hurt. This was clearly some little software glitch or some other issue with the map. I don’t think they ever got to the bottom of what happened.

Here’s another way of thinking about it: Maybe this was a Final Destination situation. If the Waymo had gotten immediately onto the freeway, there might have been a terrible accident. Something in the training said, “No, we need to stay in this parking lot. We’re going to drive around in 8 circles, reset the timeline, and ensure that Mike makes it safely to the airport.”

Something to think about.

Do you know how airport Wi-Fi sometimes makes you watch an ad before you can get free Wi-Fi? This is giving me an evil business idea: “You want to get out of your Waymo and make your flight? Time to click over to Honey and complete your purchase with Honey. If you want us to stop circling this parking lot…”

Someone out there is taking notes.

I’m so sorry.

Stop generating.

That is HatGPT. Casey, it’s so good to be back with you in the studio doing one of our favorite games.

Hats off to you, Kevin, and hats off to all of our listeners.

The worst thing happened: Without a weekly podcast to joke around on, I had to do bits for my family, and they’re much less appreciative than you.

I tried to do an Aqueduct rant to my wife, and she said, “You’ve got to save this for the podcast. This is not…”

I made some stupid pun, and my boyfriend just looked at me and said, “That was in the podcast voice.”

Busted.

I had to get all these out of my system.

He said, “I don’t think that’s how that works.”