我们的 Google I/O 实地探访 + 与 Sundar Pichai 面谈 + 系统更新

Casey, I couldn't help but notice that during much of the I/O keynote, you were on your phone—

Yeah.

—playing Balatro and doing emails on your laptop.

When you have your agent recording the entire thing and transcribing it for you, you can just glance up every few seconds and see if a news event is happening. It's actually really amazing, the future we're living in.

This is the promise of AGI.

Yeah.

Soon we will be able to play Balatro wherever we are.

You can pay attention to nothing and be fine. Welcome to the future.

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, it's our annual field trip to Google I/O. We have all the big news, plus our conversation with Google CEO Sundar Pichai. And then some other highlights from the week with our system update. I told it to install last night, but it didn't.

Did you delay it?

Yeah.

You're always doing that.

Well, Kevin, the keynote just wrapped up here at I/O 2026. What did you think?

Yes. This is like the Coachella of capitalism, the Warped Tour of the web. The Lollapalooza of links. I could keep going.

Please don't.

1. Google Unveils Its AI Strategy

They're doing a lot. We heard about everything from agentic search to these new video-editing models and products, all the way to Demis Hassabis at the end of the keynote declaring that we are in the foothills of the singularity.

Demis Hassabis

When we look back at this time, I think we'll realize that we were standing in the foothills of the singularity.

I love that Google's version of the Steve Jobs “one more thing” is like, “One more thing: the singularity.” A great way to end. So, what was your standout? What did they announce that caught your eye?

They have said that they just made the biggest change to Search in 25 years. I think time will tell if that is truly as significant a change as they're suggesting, because when we saw it in the demos, it just looked like the box grows if you type more into it. But they're also saying that by this summer, they're going to be generating custom user interfaces based on your query and a bunch of other stuff.

I think a through line through everything that they discussed here was “coming this summer” or “coming to a group of trusted testers.” With these kinds of productivity tools, until they're in your hands, you don't really know if anything has changed in your life or not.

There was a lot of focus on agentic stuff, like agentic coding. They have this new agentic search mode, which is basically the fancy upgrade to Google Alerts. You can have it tell you when there's a new home listed on Zillow or a new baseball score you might care about or something like that.

I was really interested that they seem to be betting big on cost and speed. Their new model that they talked about today was Gemini 3.5 Flash, which is the newest version, but they also say it's 4 times faster and much cheaper than the other leading frontier models.

I think Google's strategy of betting on scale and distribution and making their models as efficient to serve as possible is really the direction they're going in. They don't seem to care about having the absolute most cutting-edge models as long as they can serve them cheaply and quickly to billions of people.

I think they care about it. I think they just haven't built it yet. Fast and cheap is what you talk about when it isn't the best. And so that's what we're getting this year.

But I am interested in trying these new agents. I tried to use the new agent. I said, “Let me know if Kevin Roose writes a good story,” but it hasn't sent me anything, so I don't know if it's broken or if the AI isn't working.

Yeah, I'm writing good stories constantly—

Mm-hmm.

—so something must be going on.

My real curiosity is which of these agents are going to find true product-market fit. Which is the agent that is going to get 1 billion users? We saw them show off what is effectively their version of OpenClau, which definitely had a big moment earlier this year, but it's still not clear to me that that is an actual mainstream use case.

I have to say, of everything they showed off, I'm not dying to have Gemini OpenClau working overnight on my behalf, because I don't know what I would ask it to do.

I think that was actually one of my favorite parts, because I am currently running a Claude swarm on my open laptop—

Doing what? What is it doing?

And what they've said is that you can now do that through Gemini Spark—

Uh—

—this new agent thing, and it will run in a virtual machine on the cloud, so you don't have to keep your laptop cracked open like an insane person.

But, Kevin, please, what are you doing with your Claude swarm?

I can't tell you. I—

You can't even—because you're not doing anything.

I'm ushering in the singularity.

You're generating recipes.

I'm building a bioweapon—

Yeah.

—but I don't want to say that in this crowd.

They get sensitive about that sort of thing around here.

Yes. So what was the vibe of I/O this year?

Honestly, the only people that I have talked to are reporters and people who work at Google. I feel like we have not really had a chance yet to talk to developers and see what is on their minds. Clearly, they have a lot of new toys to play with, and I'm sure they're excited about that. But, as I said, until you get your hands on them, you don't really know if your life has changed.

I feel they have confidence in their place in the front of the pack. I don't think they believe that their models are the best in the world yet, but if they're cheaper and faster and they can do more stuff, they have this new omni model that can take in video as well as images and text.

I feel like they're doing the spray-and-pray approach to AI development, which is you just do everything, everywhere, all at once and hope that something hits. And it's worked for them.

I guess I will say one thing about the vibe here at I/O this morning, Kevin, which is this is the only recent gathering of a large number of people where mentions of AI did not produce a large chorus of boos.

Oops.

This was a crowd where people actually seemed like they wanted to hear about new AI developments. As we saw many, many AI-generated images and videos for almost 2 hours, there was nary a boo to be heard.

I agree. Not a lot of mention of the AI backlash or any of those things. I think they're just hoping that these new products are so useful to people that they'll forget about all their objections. And then, of course, comes Demis Hassabis to talk about the singularity, just to remind you that none of this is normal, that we are experiencing something very weird.

All right. I think those are some good early thoughts. We have a lot more to do at I/O, starting with hitting that lunch line in the press tent.

I saw some ube croissants on my way in. I'm still trying to figure out what that means, but I'm going to go try one out.

Perfect. All right, Kevin. We are back in the studio the day after I/O, and we have a few hours before we head back down to interview Sundar Pichai. But before we do that, you have spent a little bit of time playing around with this new Gemini 3.5 Flash model. So what have you been doing with it?

2. Gemini Flash Gets A Reality Check

I did try to use Antigravity, their coding assistant, to put the model through some of my own evaluations, and it seems like 2 things are true.

One is that it's very fast, and you don't hit your token limits very quickly at all, which is a nice contrast to some of the other coding models that I've used. The second thing is that it just doesn't seem like that much of an improvement over what's out there now.

There have been some people on my feeds talking about how they are feeling disappointed by the new Gemini model. There are some indications that it's maybe not as good as Gemini 3.1 Pro at some coding evaluations, but it is much faster.

I think the other thing that people are reacting to is the cost. The cost of Gemini 3.5 Flash, while lower than other frontier models, is significantly more expensive than the last version of Gemini Flash, so people are upset about that. But in general, my sense is that it's a good model.

It’s a pretty good model. It’s not blowing me away, but it is quite fast, and if you are a company that’s burning billions of tokens a day, you might be interested in that. But for the average person, I would say nothing about it has made me want to switch to it as my daily driver yet.

Yeah, I have to say, this class of models has always just felt like it’s not for me, right? I am somebody who is always willing to wait a little bit longer and spend a little bit more to get the best answer. I would rather do that than have instant answers from a very cheap model.

You like to surround yourself with awe-inspiring intelligence.

Yes.

Right?

That’s exactly right.

Which is why we’re hosting this podcast together.

That’s exactly right. And so, look, I think for a very wide variety of use cases, a model like this is going to be perfectly fine. If you’re the average Gemini user and you’re just trying to pass 6th grade for free, I think this model is going to be a lot of help.

But the more vocal AI power-user crowd, which are the people that you’re going to find on social media putting these models through their paces on day 1, does seem to be a little disappointed here. And I do think it’s notable that a model that was sold to us yesterday largely on the strength of its low price does not actually seem to be as cheap as maybe we were told.

Yeah, and I think part of this is just that the cycle of I/O and developer conferences is maybe out of step with how these models are actually being built. I’ve been hearing some whispers from folks at Google that they are working on something that is more powerful, but that it just may not have been ready in time for I/O this year.

I think that’s one tricky thing about doing these conferences once a year: Every deadline has to line up at exactly the same time, or else you end up looking like you’re behind.

Yeah. I mean, I think the real test for Google, or at least the next real test for Google, will come next month when it puts out Gemini 3.5 Pro, which I think will be in a similar class to the frontier models that we have today. That’s when I’m going to be putting the model through all of its paces and trying to figure out what part of my daily workflows this kind of thing might slot into.

Yeah. Casey, what is your verdict on I/O? Did it meet your expectations?

It felt like a base hit to me. I think that if they ship everything that they showed off and it works basically as advertised, there is some good stuff in there. I don’t think there was a single product that made me think, “If I can’t use this today, I will be in extreme physical distress.”

But overall, I thought it was a solid set of announcements. What did you think?

Yeah, I think that basically meets my expectations. I talked to 1 reporter who was walking around at I/O and who said it was mid, sort of like—it’s good. It’s good. It’s a better position than they were in a couple of years ago at I/O, when they were basically scrambling to get Bard out the door and everything looked like it was going sideways.

So they found their footing, but it is still not clear whether they are producing a model that will truly be state of the art. All right, so those are some more of our thoughts and early hands-on impressions of Gemini 3.5 Flash.

Now let’s head down to Google to talk to Sundar. And because I imagine we might want to ask Sundar a question or two about AI, let’s make our disclosures. I work at The New York Times, which is suing OpenAI, Microsoft, and Perplexity.

And my fiancée works at Anthropic.

We should also disclose that we both use Google.

I’m a longtime Google user.

Yeah.

I was part of the Gmail beta in 2004.

Wow, you’re old.

So old.

Sundar Pichai, welcome back to Hard Fork.

Thanks for having me. Great to be here.

Thank you.

3. Sundar Pichai Assesses The Race

So the last time we had you on the show was in 2023. Bard, RIP, had just come out, and I think at the time, the perception was that Google was catching up in AI. How are you feeling about your position in the race these days?

Well, that brings back memories, so it feels like eons ago. Those 3 years feel like a long time ago now. But I think it’s staggering to see how much progress the technology is making and how much progress we have made as a company. I think it’s a very dynamic moment in the industry.

I think our models are at the frontier in some areas, and there are areas where we are behind the frontier, so it’s a combination. If you look at overall capabilities, including text, multimodality, voice or audio, reasoning in general, and overall intelligence, I think we are very capable.

When it comes to agentic coding with tool use, instruction following, and long-horizon tasks, I think we are a bit behind at this moment. But we are hard at work, and the space is so dynamic. All of the leading labs have their own pretraining cycles, so you have these cadences, and they may not exactly match up.

I think the moment is intense enough that if you’re slightly off, 3 months ago people were saying, “We are ahead, and no one could catch up with us,” and now the conversation flips. But that’s part of the territory of being at the frontier. I think we are the only large company that is actually at that frontier, right?

One way to think about it is that, in this moment, there are a couple of startups that have made extraordinary progress, and we have been deeply working on this for a long time. I think we took a big step forward with Gemini 3.5 Flash. It does address some of the areas in which we have been behind, and getting it out in the real world and iterating with that data coming back is going to really help us.

I think coding was an area where getting access to the data flows was important. I think we maybe didn’t quite have the surface, like Cloud Code, as an example, or what Anthropic maybe had with Cursor, too. And so getting Antigravity with 2.0—we’ve been using it internally at Google for a while. I shared the token usage at Google I/O. I’ve never seen anything like it internally. We’re doubling every week, and people are really putting the models to work.

So that is helping us hill-climb quite a bit. But the frontier is very dynamic, and I’m very, very optimistic and confident we’ll push through there.

It sounds like if there’s any place where you feel maybe not quite at the very lead, where you actually want to be, it is coding. Is that right? Is that where you’re putting the pressure?

Look, I think coding ends up being very foundational in everything we do, so I think it’s an important frontier to be on. There are areas in coding where we have been very good. We’ve been very, very good at creating single-shot web front ends and everything.

But in terms of this long-running task, where serious developers are working on complicated code bases, I think we are making progress. It’s just that there is a gap to the frontier where others are. But we are working. We are well aware of it and making progress there.

Gemini 3.5 Flash has been out for a day. I do think it typically takes a few days to really put these models through their paces.

Yeah.

We have seen some complaints, though, about pricing and model quality. I’m curious what you’ve made of the reception so far.

I’m looking forward to being done with my interviews and so on so that I can spend more time with the teams.

Yeah, wrap it up, Sundar.

No.

No.

Look, I’m going to meet the teams right after this.

I think it’ll take a day or two to settle in. It’s a new model in a new area where we’ve made some progress. There could be some regressions, but we will be able to address them quickly through our post-training, I think. There are some artifacts and behaviors we are seeing that I think are easy to address, so we will.

Given that it was a day after us putting out a lot of things, I think we had tightened usage limits to avoid outages, but you will see us make progress on usage limits very soon. That is rightfully a source of frustration when you encounter them. I feel the same, but those are areas we will address pretty soon and make progress.

It seems like one thing that some of the AI companies are succeeding at is focus. Anthropic and OpenAI have this sort of relentless focus on coding. OpenAI was criticized last year for spreading their bets too thin, trying to do too many things all at once. They’ve now tightened that. Do you feel like Google is appropriately focused on coding, or are all the other bets you’re making taking away resources and time and focus from the main push?

I think all of us saw an inflection point in coding, and I think we are all responding to it. I think we all have pretty serious strikes around this area, so I don’t see it as an issue. We are a large company, and we have scale, so we will be able to focus on multiple things at the same time.

I don’t see any fundamental issues. We are making progress. We’re going to make progress. I think we are in a moment in time in this field where 30 to 60 days look like 5 years. That’s all it is.

Another thing that got a lot of attention this week was the changes that y’all made to the search bar and the sort of front door of Google—the biggest change in 25 years. I think a lot of people expect that at some point the normal Google, sort of classic search interface, goes away, the 10 blue links maybe go away, and you just have this AI Mode as the default. But you haven’t done that yet. There’s a lot of integration, but you still can get the 10 blue links if you want them. Do you think that goes away at any point, that you sort of rip the Band-Aid off and just go full AI Mode?

I think it’s important to bring users along the journey, as well as making sure the product is working for their expectations. So I try not to get ahead of that. I think it is very clear, as we have evolved through these changes, that people are responding positively. We can see it in the long-term metrics of the product in such a clear way, right? And so I think we understand that.

But people want Search to be fast. I do think through Search people are looking to connect with what’s out there on the web, so that’s important to us. So it’s all of that. I think you’re seeing us evolve the product, and I think you’ll continue to see it be methodical, but we didn’t have AI Mode a year ago. But now a lot of people are experiencing it. I think we have made it more seamless to go there than before, and so it’s a continuum. But I do see that sources and links will always be there as part of it.

He was telling me on the ride down that he feels like he basically has not done a traditional Google search in the past year, that he’s sort of fully doing these AI searches. When you hear that, are you like, “Cool,” like, “This is the kind of user that I want right now”? Or does it send you a little chill because the traditional search ads business is a pretty good one for you?

If anything, in AI mode, in an agentic mode, these things are going to do a lot more for you than what we were able to do for users 10 years ago. I think the economic value is always a function of the total value you’re giving users. I think all of us would say, over time, the value we are providing users increases. There’s more competition. There are more choices. So I feel comfortable that, between a combination of subscription and ads, the right models will continue to be there. Adam Smith’s rules don’t change in this new world.

Right. Let’s talk about public perception. The New York Times and CNN did a poll this week, found that only about 16% of people say that AI is mostly good, while about 35% say it’s mostly bad. What do you make of the AI backlash that we’re seeing right now, and how much leverage do you think Google has to change that perception?

4. Google Confronts AI Anxiety

I’ve always viewed AI as the most profound technology humanity will ever work on. It’s progressing at an extraordinary pace. Humans aren’t evolved to process that much change, and I think people, rightfully so, are anxious about what future this technology will bring. I understand it. It feels natural to me with such a profound technology shift.

We’ve had far simpler technology shifts where there has been anxiety around those shifts. This is of a scale unlike anything we’ve seen before. I think we as an industry have to do a lot more to continue driving and showing the benefits that are possible with the technology, so that is something in our control. We have more work to do to make sure that, when we are scaling up the infrastructure investments, et cetera, we ask ourselves what we can do to make some of that work better.

But I think people’s concerns are a bit more fundamental around the shift than all of that. I think a natural part of this is that people are anxious about their economic future in this world, and you have a lot of conversation where people are saying jobs are going to radically change, some of them will go away, et cetera. I happen to think the outlook is better than some of those dire predictions. But, as a society, if you’re hearing that, it’s natural. I would be surprised if people weren’t more anxious about it.

I do think it’s important because the change is happening so fast. In democracies, you need citizens to be engaged, be aware that this is happening, and make their preferences known. That’s what causes action in society. So I think there’s something healthy about this dialogue, too, and given the pace at which the technology’s moving forward, it seems right to me—both the concerns and the fact that we need to take it seriously.

You’re giving the commencement speech at Stanford next month. I’m sure you’ve noticed or heard that a bunch of commencement speakers have been booed recently by college students who are worried about AI. What are you planning to tell the graduates about AI, and do you have your boo strategy in place?

Anytime we have driven technology progress, I think it helps drive progress in the world, and in some ways, these graduates are actually both going to be a big part of that, driving that progress and also dealing with the impact of that technology. So I think we have to be very mindful of that.

I’ve always been extraordinarily optimistic about the next generation. I think we always have this view in the world: We are anxious, and we worry about the next generation. But I think the next generation rises to the challenge and builds a better world. So I view it as no different from those moments, and my goal would be to share my experiences and share that with them. That’s what I’m looking to do.

You can just pretend they’re saying, “Google.” It’s close enough.

I’d be curious to hear a little bit more about your case that, sure, jobs may change a lot, but, as an entry-level graduate, your economic future is still bright. What is that case in your mind?

On a basic level, I do think there is a new level of capability all of us are going to have to be able to do things. I wasn’t there when spreadsheets rolled out to people. I wouldn’t know how you did financial analysis before that. How did people do it, right?

I’ll say it: I didn’t do it. I had no idea how to do it.

Okay.

Yeah.

But spreadsheets changed that. And so there’s an aspect of this: I think it’s just going to change the starting point for many, many people. Even coding—I think if you fast-forward the progress we are seeing there, so many more people are going to be able to code in the world, right? And I’ve heard you two might be examples of that—

Yeah.

—and in that journey. But I think you’re just at the leading edge of what is going to happen more and more. So I think those are the new serendipitous ways this will all work out that we underestimate.

I think people are going to be more productive. They will have more time for leisure. All of that will simultaneously be true. There are so many areas where, today, people’s work involves a lot. Doctors have high burnout rates, and it’s because they train, and their calling is to spend time with patients, taking care of patients.

But most doctors would tell you, if you actually watch their time, the percentage of time they spend with patients is less, right? So I think AI will actually help them do more of that, right? And I think those are examples of it.

The radiologist analogy has been fascinating. It’s been running for a decade now. I look at myself and I say, “Well, I’ve gotten a lot more scans in my life than my dad ever did.” Each of the scans has 10 times the amount of information that his scans had because they were constrained by printing film, whereas we’re digital.

And I think that number is going to be 10 times in 10 years. So with that projection going forward, you’re actually going to need AI to keep up with that demand coming. I think the impact of all this will be nonlinear.

I don’t want in any way to minimize the fact that every technology shift brings disruption with it. There will be disruption, and we as a society need to be super serious about it and engage. Some of those conversations are rightful, in thinking through that.

But I do think there are many positive dimensions to it that maybe are not being talked about. I also think there are overly deterministic, dire scenarios that I don’t quite agree with.

Let’s talk about agents, because I feel like agents tie into this question of what’s going to make us more productive in the future and how they’ll change our jobs. Later this summer, you’re releasing Spark, which seems intended to be an agent for the regular person. Could you walk us through something this agent is doing for you personally?

I’ve used it a lot more in my professional context because it was mainly available in my corporate account. In that context, I’ve definitely used it to prepare for any meeting. I wish I had brought the prompt and the output. As a test case, I used it for Hard Fork.

Honestly, if you email it to us, we would flash it on the screen.

Yeah.

It had some things about the two of you, so I don’t think I can project it.

No, that’s what we want. We want that.

Just kidding.

No, we want that. We want that.

I’m not sure I’ll allow that. No, I’m just kidding—partially kidding.

Yeah.

You should see Casey’s browser history.

Yeah. Partially kidding.

Yeah.

But I’ve had it in my personal account more recently. Here’s a simple task I did: I asked it to look ahead at my meetings and color-code them by category so I could keep track of how I’m spending my time.

It’s extraordinary to watch. It came back with suggestions for 2 color-coding schemes, and I just had to choose one. It’s actually sci-fi. It just changes the color in the calendar: personal meetings, health-related meetings, time I’m spending at work, et cetera.

That’s an example of a personal query I just did to see what’s happening. With agents, I think you have to give people a sense of control. I think about this as what allowed us to get someone to sit in the back seat of a self-driving car. We did it in steps, and I think there’s an element of that with agents. If something unexpected happens, people will back off from this.

Part of it is earning their trust and giving them a sense of control and transparency. More importantly, from a security standpoint, these systems can be hacked, and so we want to make sure we’re not ahead of the frontier in the wrong way.

Speaking of meetings and your calendar, we hear that you’re headed to the White House for some kind of AI executive order signing. What should the government be doing right now to regulate AI?

Do you like this idea of a kind of prerelease strategy where the government gets to see models before they’re released and sign off on them? Is that a good idea? Is that potentially dangerous if it gives them the ability to censor or jawbone companies into releasing different kinds of models? What’s your take on that?

Look, we’ll have to wait and see the details of the full executive order, but they’ve engaged with the industry in a very robust way. I think the approach really balances innovation and oversight. We’ll have to wait for the details to come out.

There are a few areas that are coming up. We will need more cross-industry and government coordination, so it makes sense to me. Cyber is a great example of that. I think we all have to work together. It makes complete sense to me that if you found an exploit that could impact a governmental agency, the government needs to be prepared for it.

There is validity to that, but of course, we need to do it at a moment with this important technology when it’s important as a country to be at the frontier, too. We shouldn’t do it in a way that overly slows things down, and maybe that balance has to shift as we reach more advanced levels of technology. But to me, this seems like a prudent approach.

In general, part of what we are doing—building Synthy, open-sourcing it, sharing it with others, and more importantly, building a consortium together—is an example from a different area. These things only work together if we can come together as an industry. I’m glad they’re approaching it in that way.

Another sort of safety-related question: All of the big labs are racing toward what you call recursive self-improvement, building AI systems that improve themselves rapidly. Do you think that can be done safely? Do you feel like you have a line of sight to it right now?

These models are getting better at coding and agentic workflows. You can see in Antigravity today that, in over 12 hours, it can build a simple operating system from scratch. Those are genuinely multiple thousands of hours for somebody to do.

You’re seeing some of that in the work today. In our products, we all have agents and subagents, and the orchestration of those agents is building things together. It’s a continuum, and we are all definitely making some progress.

But in the way people describe recursive self-improvement, I don’t think we’re there yet. That would represent a next level of acceleration, and I think it would have a lot of implications, but we aren’t quite there yet.

Is there a plan for, “Uh-oh—great news, we just hit recursive self-improvement”? Do we break glass, or what happens then?

It’s a great example of why, if you’re approaching moments like that, all responsible labs would be consulting with others. It shouldn’t be an internal conversation at that point. It has to be a much broader conversation than that, and I think we all have to avoid race conditions at those stages of AGI.

Right now, all the labs are racing to get more compute. There seems to be bottomless demand for compute. They’re hoarding it wherever they can, striking deals, building their own data centers.

Google is still selling access to TPUs to rivals and other companies in the race. Why aren’t you just keeping that for yourselves and your own models?

I think each one is not a constraint on the other. As long as we can make enough chips, it’s not a constraint. The right way to think about it is that we have Google DeepMind and our first-party services. You can think about that as a company business with cash flows, and you’re planning for that.

Then you have Google Cloud, which is a business with revenue cash flows, and you’re doing long-term planning for that. If we didn’t have Cloud and weren’t providing them, we wouldn’t be planning those chips anyway. That’s the simplest way to look at it.

Obviously, it’s a bit more complex than that, but there are a lot of advantages to providing TPUs to others. The fact that researchers at Anthropic are using TPUs, in addition to us using them, allows us to make the best hardware for the next generation.

By the way, we use NVIDIA’s chips too, and their next-generation chips are incredible. We use those, and we work together too. When you’re running platforms and you have a platform side of the business—and I’ve always worked on many platforms in my life, be it Chrome, Android, or Google Cloud—why would you ever open-source something, or why do you provide this technology? All of that makes sense on its own merits.

I do think there are advantages. As I mentioned, it allows us to stay at the frontier. Economies of scale help in various spaces, and so it makes a lot of sense that way.

Yeah.

The last time we had you on, we asked you about AGI and your feelings about the term. At the time, you responded that it didn't really matter whether you've reached AGI or not because the systems are going to be very, very capable, and Google's strategy should be the same. I noticed that you did not say AGI in your keynote. Demis did, but you did not. What's your relationship with the term AGI today, and with the idea that all of this progress is building toward something singular and world-changing?

We are—you know, there is inevitable progress toward AGI that's happening. I've long understood it, and otherwise I wouldn't have pivoted the company 10 years ago to put that technology at the heart and center of the company. All I meant by that statement was that even in the scenario where AGI is going to take 10 years, the technology three years from now will be so much more powerful than what we have today that I don't want people to think that because AGI is 10 years out, they don't need to act or prepare. That is all my statement means in those contexts.

Are you AGI-pilled?

Well, I absolutely—I’m sure that the technology is making foundational progress toward AGI. I'm less able to predict with certainty whether it's in the three-to-five-year timeframe or the five-to-10-year timeframe. The rate of progress over the last 1 to 2 years has made me feel it's on the closer side than not.

In your role running one of the largest companies in the world, which has a responsibility to society, the language I choose to use around it might be different from other people's. But I think, as a company, 10 years ago, on the I/O stage, I announced TPUs and AI-first data centers. Yes, we clearly understood where this technology was headed.

Maybe as a last question, one of the more memorable phrases from the keynote this year came from Demis, when he said that we're in the foothills of the singularity. Can you tell us concretely what that means from Google's perspective, and should people be excited about that, afraid, or both?

Look, I've had many conversations with Demis, obviously, on this topic. I think in this context, he's defining singularity as the advent of AGI. I think he's talking in that context, and if you think—if I remember correctly, I think he had articulated it as happening by 2030 or so. If you believe that, it makes sense to you. That's what you're articulating, and it's as simple as that. I think for him, that's how he defines singularity.

I think Demis, myself, and many others all feel it's important to articulate that if that's what you believe. We are all at the frontier building this technology, and hopefully people are listening. I think it's important that, as a society, we internalize that and get ready for it.

Sundar Pichai, thanks so much for coming.

Thank you, sir.

Thanks.

Great to talk to you.

Appreciate it. Take care.

5. System Update Tracks AI Fallout

Well, Casey, other than I/O, there has been a lot happening this week on some stories that we have covered in the past. So it is time for our segment that we call System Update.

All right, update number one: The Elon Musk–OpenAI trial is finally over after weeks of testimony. We got a verdict on Monday.

And what was that verdict, Kevin?

After less than 2 hours of deliberation, the jury came back and unanimously rejected Elon Musk's claims, not on the merits, but on the grounds that he had waited too long to file the lawsuit. They said it fell outside of the 3-year statute of limitations, and they basically said, “We're not going to even debate the merits because it's too late. You're too late, bro.”

When is Elon Musk going to catch a break in this crazy world?

Yes. I think the trial ended up being quite important, but not for reasons related to the trial. It was all the evidence and the gossipy juice that came out during the trial.

It felt like a stimulus package for the tech media—

Yes.

—where we just had a steady drip of new lore, and it kept us all very entertained. I don't know that we learned very much that truly affected public sentiment one way or the other, but I definitely learned things from this trial that I'll never forget.

Yep, me too.

Yeah.

And Elon Musk is obviously saying he is not done fighting OpenAI.

We haven't seen the last of him.

His lawyer said that he would appeal the decision, and on X, he said, quote, “Creating a precedent to loot charities is incredibly destructive to charitable giving in America.” Charitable giving is obviously something that Elon Musk famously cares a lot about.

Also, one of the great revelations from this trial was that Elon Musk himself had been scheming, prior to 2020, to turn OpenAI into a for-profit company and even absorb it into Tesla. So effectively, he had planned to do all of the things he would later complain that OpenAI actually did.

Yeah. Do you agree with Musk that this is essentially a technicality and that there are still some substantive issues that need to go to trial here?

Well, every law, in some sense, is a technicality, Kevin. That's how the legal system works. He put this one to the test, and he lost.

Yeah. Are there any lawsuits you've been waiting to file that you should get in under the statute of limitations?

Oh, my backlog of lawsuits is so big. I have to finish this podcast and get to it.

All right, update number two: Meta is continuing to reshuffle and shrink its workforce. This Wednesday, the company officially laid off 10 percent of its workers, roughly 8,000 people. We talked about these rumors on the show a couple of months ago in terms of their relationship to AI and the fact that they were seemingly cutting workers to make space for more AI in the organization.

We knew that these layoffs were coming, but we also got a new update to the story this week: Meta announced that, on top of these layoffs, it is also reassigning 7,000 workers to new initiatives related to building new AI tools and apps. Janelle Gale, Meta's head of HR, said in an internal memo that the restructuring “will make us more productive and make the work more rewarding.”

You know, there truly is no company among the big labs that has done as many reorganizations of its AI teams as Meta has done. You could almost set your watch to how often these things have been happening.

Yes, this is a company that specializes in corporate reorganizations—

Yes, they monetize their social media—

Yeah.

—but the product is reorganizations.

Yes.

But look, I have talked to some people who have actually been impressed with the scale and scope of what Meta is doing here. If you want to know what it looks like for a big tech company to move heaven and earth to try to catch up in an area where it is behind, this is what that looks like.

So far, I don't think the efforts have borne much fruit, but they have just reassigned 7,000 people to work on initiatives related to AI. Of course, that's in addition to all the other hiring that they did last year. I'm far from saying that Meta is back in this race, but of all the reorganizations they've done, I do think this is one of the most intriguing.

Yeah, and I think we should expect a steady drip of stories from Meta over the coming months about how bad morale is. It does seem to be quite bad over there at the moment. That always happens when you do these kinds of major shake-ups and reorganizations, and they're also doing all this employee-surveillance stuff that workers are understandably upset about.

There was some leaked audio published this week from a meeting in which Zuckerberg basically said, “Look, we've got to train our systems to be good at coding and using computers, and the way that we do that is by training them on your computers and your coding because we have a bunch of really smart people here.” That was his way of explaining why they're surveilling their employees.

Yeah, but I have to say, the whiplash between the Meta of today and the Meta that we were covering 10 years ago is pretty extreme. The thought just keeps occurring to me: Meta used to be a fun place to work. They have a fake Main Street at their corporate headquarters where you used to be able to walk into a Mexican restaurant where everything was free, and you could drink a margarita at lunch. Now, that restaurant may still be there, but I think the Margaritas at Lunch program is on the decline because these people have to go create training data for large language models now.

But I have to say to the employees, if you're thinking, “Man, this really sucks,” my heart goes out to you. Everyone should have a job that they like. On the other hand, I think it's been clear for a long time that if you work at Meta, your job is to help Mark Zuckerberg win the race to build AGI before anyone else, and that is the entire point of the company now.

Yeah.

There's not a secondary objective.

Totally. All right, next update. This one is a system popedate because this one has to do with our new pope.

Did he go on a date?

No, they're not allowed to do that.

Oh, okay.

But he is set to release his first encyclical on the topic of AI this coming Monday. Pope Leo XIV is expected to release his document, sort of circulate it among the church, which is meant to guide the clergy. The surprising part of this is that Chris Olah, one of the co-founders of Anthropic, is going to be there at the Vatican to help him launch this thing.

I didn't realize that we did launch events for encyclicals.

Yes.

It's like the pope finishes up with Mass and says, “But I do have one more thing.”

It's their I/O.

Yes, this is Vatican I/O. I mean, I guess we'll be very curious what the pope has to say. You know, Kevin, as you know, this pope seems unusually interested in tech and has had a relatively lot to say about it during his short tenure.

Yes, and I also am very excited about this because I think this means that Pope Leo XIV is the most likely pope to come on Hard Fork.

Absolutely.

And if you are in the Vatican listening to us right now, Pope Leo, please come drop your encyclical on our audience.

Your Holiness, we would simply love to create content with you.

All right, update four. The literary world has been rocked by 2 new scandals related to AI use by writers. The first is that an award-winning short story has been criticized by people who think that it was generated by AI.

Oh, boy.

This story was a regional winner of the Commonwealth Foundation's short story prize. It's called “The Serpent in the Grove” by Jameer Nazir of Trinidad, and it was published online by the British literary magazine Granta. But some readers thought, “Hey, this thing seems a lot like AI.”

Hmm.

It had some of the clichés of AI writing, and people claim to have run this through Pangram, the AI detection tool, and found that the story appeared to be significantly or entirely generated by AI.

Yeah. I'm not sure that the backlash here is warranted because this story has gotten more attention than any previous story written in Granta. I think it's possible that more people have read this story than have read a Granta story in years.

We have figured out how to rescue fiction, and it is just to create controversies around AI.

Yeah.

The Commonwealth Foundation responded to these accusations, saying in a statement that, quote, “Place our confidence in the integrity of our contributors and the caliber and experience of the judges and chair of the judging panel, and stand by the assurances given by our authors as part of our process.”

And I mean, look, we don't know.

We don't know.

We don't know.

But this is, I predict, a thing that is going to happen to every major literary prize: Every submission will just have to be run through an AI detector.

Also, we're all now just reading so much AI-generated text that I feel like inevitably we are all just going to start writing more like it. Even when you are writing something by hand, I think there is just a risk that over time it is going to look more like slop.

It's not a trend; it's a transformation. There's another writing scandal that came to light this week, this one in the nonfiction world. This was according to my colleague Ben Mullin at The Times, who wrote that a buzzy new book called “The Future of Truth,” which is about truth in the age of AI, appears to have included numerous made-up or misattributed quotes concocted by AI. The author, Steven Rosenbaum, acknowledged the errors on Monday. He said that he took full responsibility for the errors, that he used AI tools like ChatGPT and Claude during the research, writing, and editing process, and that he was working to correct future editions.

This story implicated our dear friend Kara Swisher, Kevin. I don't know if you saw this, but Kara was one of the people whose fabricated quotes appeared in this book. It's so funny because if you read the quote attributed to her, it reads just like pure ChatGPT.

Yes.

Do you want to hear it?

Yes.

“The most sophisticated AI language model is like a mirror. It reflects our own morality back at us, polished and articulate, but ultimately empty behind the surface. It's not bound by Asimov's laws or any ethical framework. It's bound by the patterns in its training data and the objective set by its creators.” If you've ever talked to Kara Swisher—

She has been brain-swapped.

A true pod-person situation. So, yeah, I mean, truly nothing could be funnier about a book about AI and truth just having a bunch of hallucinated slop quotes in it.

Yep.

You have often said that anyone who puts a rule into a workplace, a school, or a publishing house that says, “Hey, no AI use,” all you're doing is telling everyone that they have to lie about their AI use.

Yes.

This seems like we're just hitting some kind of inflection point where it's becoming obvious that even people who absolutely should not be doing this are still getting lazy and not doing the work.

Yeah, and I would say it's not just laziness. There are a lot of things that AI tools can help you with on books. I've been using AI, not for writing my book—it's all human-generated—but in collecting endnotes and indexing and various things that are not the main book itself. I have had to go through with a fine-tooth comb to make sure that all of the sources that my AI tools wanted to cite were actually real. Most of them were, but most is not enough, especially if you are writing a book about truth in the age of AI.

And so I have had to do some cleanup on that before sending it off to the printer. So I understand how this kind of thing could happen, but you really have to go through if you're doing this and make sure that you are not passing off someone's words that they didn't say.

Well, let this be a lesson to all of the writers among our listeners.

Because Kara Swisher will get your ass.

She will absolutely get your ass.

All right, before we go, we have one more update. This one is very exciting. It is about Hard Fork Live.

Yay.

That is our upcoming live event in San Francisco on June 10th, and we are finally ready to share our guests for the event. Casey, who do we have?

Well, leading off our lineup this year at Hard Fork Live, Kevin, will be Microsoft CEO Satya Nadella.

I'm hitting every sound effect on the board right now. There are only two.

This is somebody who we have been trying to talk to ever since we started Hard Fork, somebody who is at the center of everything that is happening in AI, and also just a really fun person to talk to. So thrilled to have Satya as a marquee guest for us at Hard Fork Live.

Yes, we have a bunch more exciting guests. Figma CEO Dylan Field will be there. We've got Cindy Cohn, executive director of the Electronic Frontier Foundation. Plus we've got a bunch of other great guests, including friend of the pod and fellow tech podcaster Dwarkesh Patel, AI as Normal Technology co-author Sayash Kapoor, and AI 2027 author Daniel Cocytello.

And you might be wondering, Kevin, will there be any fun surprises this year?

Will there be any fun surprises this year?

There will absolutely be some fun surprises. So now—

For those—

Here's ordinarily where we would tell you you could get tickets, but you can't. They're all sold out.

Yes.

But the good news is we will be, of course, bringing you all of those conversations here on the podcast in the weeks to come. So fear not, Hard Fork Live will be part of your life one way or another.

Can't avoid us.

And that was our system update.

What, what year is it? Are we snapping again?

We're snapping.