Google’s Strike Team for Coding Models, Anthropic’s Powerful CFO, Polymarket’s Raise

0:13Hi, everyone. Welcome to the Information's TITV. My name is Rocket Drew, in for Akash Basritcha. It is Monday, April 20th. First up on our show today, the information published exclusive reporting about a new strike team at Google. The team is focused on AI to automate coding and eventually AI research itself. Our Google reporter will join to share what she knows. The information also exclusively reported that Polymarket is raising millions of dollars at a$15 billion valuation. Our crypto reporter will come on the show to share more. And I'll be joined by the CEO of Redwood Research to talk about the conference they ran over the weekend focused on one of the hottest areas of AI safety and security.

0:58We'll wrap up the show looking at an in-depth profile of the information published about Anthropics CFO. So it was just a matter of time. Google is trying to get back into the AI coding game. The information's Google recorder, Aaron Wu, wrote the exclusive story on Google's new strike team. Welcome on, Aaron.

1:17Erin Woo:Hey, thanks for having me. Absolutely. So break it down for us. What is this new team trying to achieve and why is Google creating it just now? Yeah, so this is a new strike team that is focused especially on making Google's own internal models better at coding. And so they, like everyone else in the AI industry, have been trying to get better at coding for quite some time. Like everyone's trying to catch up to Anthropic basically. But this team in particular is focused on Google wanting its AI models to be better at writing code for Google. And so that's like a new shift in focus for them. I see.

1:54Google engineers using their own models to be better at writing the code that they spend their days writing. Okay. And do we know who's involved with this team? Yeah.

2:02Erin Woo:So it's being led by Sebastian Bourgeon, who previously ran pre-training. But the really exciting or the really interesting thing about this team is that Sergey Brin is Google's billionaire co-founder, is personally involved. Corey Kavakcholo, the Google DeepMind CTO, who's also very powerful, personally involved. So even though there are a lot of strike teams, like Google DeepMind is a place that likes to be like, okay, this is really important. We're putting a ton of resources into it. So there are a lot of strike teams, but the personal involvement from leadership from Sergey and Horei speaks to the importance of this team in particular.

2:37That makes sense that it would be a high priority. At least that was my read. Was that your sense as well, that this was sort of a matter of time? You saw this coming?

2:44Erin Woo:Yeah. So it makes sense that it would be a matter of time, but this is definitely a really high priority for Google DeepMind. All of the Google DeepMind researchers I'm talking to, when I'm like, oh, what's going on at DeepMind these days? It's all about coding. it's about the CodeStrike team. JOSHUA SHARFSTEIN - Hmm, okay. And you said everyone's kind of trying to catch up to Anthropic. I guess OpenAI has already made some progress in that regard. They've been cooking with Codex. My sense is Google's kind of far behind right now, but is there anything, any advantages they have that would help them catch up potentially?

3:15Erin Woo:ANN OUTLAW - I mean, look, one of the things we've been writing about for more than a year is both the advantage and disadvantage of Google having so much money, so much resources. is like, unlike OpenAI and Anthropic, they're not like pushing towards an IPO. They're attached to Google's wildly profitable search advertising business. And so they've got more flexibility, and they've got more resources. And so this is both like a good and a bad thing in that they can continue pursuing all of these other areas like multimodal, which OpenAI, for example, is walking back from. But at the same time, it means that they haven't always been as focused, and there's a lot of different things pulling them in different directions.

3:52Erin Woo:So I mean, I think what we can say for Google is that they are very good fast followers, right? Google hasn't really been in the lead that much in this AI race. And they would like me to point out that they did invent and transfer in my paper. That is true. But in the most recent generative AI boom, they've been doing a lot of trying to catch up to open AI, trying to catch up to Anthropica. And I think in their favor, they are good at doing this. Their models, like Gemini 3, a very good model, like they figured out thinking. And so I think like when Google is putting a lot of sustained effort into something, when like another lab has been like, hey, this is good, you should go in this direction, like Google to its credit has been able to do that.

4:34Got it. Well, then why start with internal use? Why start with their own engineers rather than beefing up their coding products like anti-gravity for external engineers to use? Yeah.

4:45Erin Woo:And so it is very similar in some ways, like the internal coding tool that they're using is like jet ski. So that's the internal version of anti-gravity. But one of the things that Sergey Brand is really focused on is this idea of AI takeoff. And so AI that improves itself and getting better at coding models is the way to do that. But I was able to report on an internal memo that Sergey sent to Google DeepMind engineers. And in this, he's saying things like we have to pivot aggressively to work on agents. We should be forcing every Gemini engineer to be dogfooding our own internal models to be testing our own internal coding tools.

5:22Erin Woo:And if there's any kind of friction to doing that, that should be treated as the highest priority blocker for AGI. And so that's really where a lot of this push comes from. So they see a path to eventually having AI automate the jobs of DeepMind's own researchers and engineers. Yes. Fascinating. Well, thank you very much, Aaron. That is Aaron Wu, our Google reporter here at The Information.

5:46Polymarket is in talks with investors to raise millions of dollars in fresh funding. Our crypto reporter, Yuqi Yang, and Deputy Bureau Chief of Venture Capital, Katie Roof, broke the story. Yuqi joins me now to share what she knows. So welcome on, Yuqi. Hey, Rocket. Hi, how are you doing? Good, how are you? Good, I'm doing good. Can you tell us about the new fundraising? What do we know? So my colleague and I broke the news late Sunday night that Polymarket is currently in talks with investors to raise about$400 million at a$15 billion valuation. And this is the latest update on their ongoing fundraising effort.

6:26We know that last month, ICE, which is the parent company of the New York Stock Exchange, already invested$600 million into this round. So is the new funding related to the fundraising that Polymarket announced last month? Yeah, so this is related. This is the same round. And after raising$600 million from the New York Stock Exchange parent company, Polymarket, is getting more interest from investors. and they are now raising an additional about$400 million or more. And we know that ICE has been a big investor in Polymarket. Last year, ICE, which is the parent of New York Stock Exchange, announced that they plan to invest up to$2 billion in total into Polymarket.

7:15And they already deployed$1 billion last year. Wow. Okay. Well, prediction markets are definitely a hot area right now. How is Polymarket stacking up against its competitor, CalShea, these days? Yeah, that's a great question. I think one thing that surprised a lot of people about our story is the valuation differences between Polymarket and CalShea. So Polymarket is currently raising at about$15 billion, and that's much lower than the$22 billion valuation that CalShea was getting last month. And for a long time, the two companies were kind of competing head on and they were kind of getting the similar valuation in the past year or so.

8:01But then this is the first time that we're seeing a divergence in their valuation. So why is that? Why is it that CalXI has taken a lead? So one possible reason is the differences in the revenue that they're generating. Polymarket for a long time has been free to use for users and they have only recently started to charge fees for transactions. Whereas CalShea always charges fee and their annualized revenue has reached$1.5 billion last month. So that's a huge leg up from CalShea. I see. Okay. And I guess both companies face some of the same issues that all prediction markets face. For example, there's been a lot of noise lately about the risks from insider trading on these platforms.

8:53What's the latest we're hearing there? Yes, so insider trading is a big concern for investors and also for users. We know that there's been a series of well-placed, well-timed bets on prediction markets such as polymarket right before the U.S. capture of the Venezuelan leader earlier this year. There's also been well-timed bets related to the Iran war that's been going on. So people are concerned that They're these are insiders that are trading and profiting from their information, just given their role in the administration. And they're placing these bets on prediction markets. And both Polymark and CalShare are taking steps to address insider trading.

9:43They're adding analytics tools that hopefully will help them monitor and capture these insiders. but yes insider trading is one of the the concern that people have for prediction markets right i guess there's a legal question there about whether someone is breaking any laws by insider trading but also a pragmatic question about if i'm trading against someone who has a lot of knowledge why would i even do that in the first place maybe you're going to scare off your users who think that they're trading against very sophisticated counterparties yes that's a great point and i And I think this is especially a concern for institutional firms like hedge funds that are looking to potentially participate on prediction markets.

10:29Usually their edge is to build these highly sophisticated quantitative models that will allow them to make money on prediction markets. But I've heard people talking about how even the best modeling cannot give you a better edge than someone who just has insider information. Fascinating. And Polymarket has had a run-in with regulators before. Is that right? What's the history there? In 2022, Polymarket reached a settlement with the CFTC to exit the U.S. market because they didn't get regulatory approval to offer their product to U.S. users. But last year, Polymarket acquired a licensed entity that allowed them to start a brand new U.S.

11:21platform. So now they're able to accept US users on this new platform that's still at early stage and still relatively small comparing to Kaushin. Got it. Well, that's big news for Polymarket. So thanks very much for coming on, Yuchi, and breaking that down for us. Thanks. Up next, we're talking about AI control, which has become one of the hottest areas in AI safety and security over the past year. Over the weekend, researchers from major AI labs convened in Berkeley for a conference focused just on this topic. Joining us to talk about the conference is Buck Slegeris, CEO of Redwood Research, which co-hosted the event and pioneered the field of AI control.

12:01Thanks for being with us after such a busy weekend, Buck. It's great to be here. Definitely. So let's start with what exactly is AI control and what sets it apart from other directions in AI safety and security? Yeah, so AI control is the approach to AI security where you try to work under the assumption that your AI systems are trying to make your security measures fail. So instead of asking whether, we try to analyze our techniques by asking if they would still work if the models involved were trying to break those techniques. And this leads to a more adversarial set of assumptions and techniques that have to rely less on the goodwill of the AIs than other techniques.

12:38Got it. It strikes me as a little bit of a pessimistic approach almost. Like we're assuming that other techniques have failed and we're saying in the worst case scenario, what is our last line of defense in terms of security? Is that fair to say? Yeah, that's right. And this is a very normal thing to do in computer security or in security in general. So when you are trying to prevent insider threat from rogue employees at your company, it's very normal to do this partially through trying to screen people on the way in. But then you also just try to work under the assumption that some of your employees probably have been compromised.

13:08and you try to, you know, mitigate against fraud and mitigate against other security issues under that conservative assumption. So I'm just advocating and working on something sort of similar for powerful AI models. I guess sometimes I hear security people talk about defense in depth, I guess meaning that there should be multiple layers of defenses in that way. That's right. This is a classic choice for a way to, you know, have one more layer in your Swiss cheese model, which is the other thing the security people always talk about. Sure, like layers of Swiss cheese with holes at each layer. Yeah, each one has holes, but the light still can't get all the way through because there's always some piece of cheese.

13:43That's great. Well, one of your reflections from your opening talk at the conference was that over the past year, AI companies have really built out their control teams. They started staffing up. Was there something that changed over the past year to get companies to prioritize this area more? Yeah, so it's a little complicated. So the main thing which happened over the last year is that agents are now extremely important commercial products. A year ago, basically no one was using AI agents very much. And nowadays, they're obviously a huge deal. And this has meant that there's been a lot of interest in trying to prevent AI agents from just kind of accidentally doing really bad stuff.

14:20There's been all these stories on Reddit of the agents that delete people's production databases and then lie about it. And so there's been a lot of interest in how companies can deploy these systems in ways that avoid those failure modes. This is also true inside AI companies. So people I know who work inside AI companies, obviously AI companies use AI agents a lot. And a lot of the time those agents have access to really sensitive IP, really sensitive code, really sensitive model weights and so on. And they're pretty worried that their models will just like go crazy one day. Just not in a way where they were lying in wait trying to cause damage, but just in a way where they just go crazy sometimes because AIs do that.

14:56And then just like send a bunch of secrets to some random person on the internet, maybe some journalist or whatever. And obviously, the AI companies really don't want their agents to do that. So they're just very obviously directly interested in trying to prevent their AIs from doing crazy bad stuff like that. And this has meant that there's been a lot of interest in developing a lot of security measures and control measures to help with those threats. And then those target, to some extent, the kind of longer-term threats that I'm worried about, where the AIs are really carefully and proactively trying to subvert security measures.

15:28I see. So even if a company is tackling a relatively modest version of the problem, like we don't want the AI to share some of our secrets on Reddit, the kind of interventions that they make, the kind of protocols they set up might be useful for a broad range of threats. That's correct. And it seems like even the government has taken an interest in this. Someone pointed out to me that the AI action plan from the Trump administration even mentions control as an area of priority. So it's definitely picking up some interest. Do we have a sense of what distinguishes each of the AI companies in terms of their approaches to this topic so far?

16:02Yeah. So GDM, Google DeepMind has not said publicly very much about what they're doing. And probably in OpenAI have both said a lot of things publicly about what they're doing in terms of their agent controls. And those are really interestingly different. And this is mostly a result of the way that their models are quite different. So a big difference between ChatGPT and Claude is that the models behind ChatGPT, the GPT models, are much more inclined to vocalize their reasoning in their chain of thought. And so it's a lot easier to see how those AIs are thinking just by reading that if you're the AI company.

16:39Whereas Claude is a lot more inclined to just kind of like shoot from the hip and take actions without really talking them through to itself beforehand. And so the open AI control systems are much more based on reading what the model is doing, or how it talks about its actions before it takes them. Whereas the anthropic control systems have to be much more based on looking at the actual actions that the AI takes and trying to detect whether the actions are bad, rather than detecting whether the reasoning that led to the actions was bad. So that's probably the single most interesting difference.

17:13Yeah, that's a big one. And then even without knowing much about what's going on at DeepMind, I guess we can make an informed guess that probably Google has a lot of security people. I imagine they have like pretty large security teams and maybe that's a resource for them as well. Totally. So Google has the best security team in the world, probably. And they also have, unlike, they're probably the best tech company in the world at working with employees under the assumption that some of those employees are compromised. So in 2012, Gmail was compromised by Chinese state-affiliated hackers. And Google then went through and basically reinvented the field of computer security to make it so they were more robust to some of their engineers being compromised.

17:54And that means they have a lot of infrastructure that is very well set up for this kind of problem. So if GDM, if Google DeepMind ends up building powerful AIs, that will probably be quite helpful for them in implementing good control measures. And then in terms of Anthropik's approach to cybersecurity, there's been a lot of interest lately in their new model Mythos, which they trained and then said is so good at cyber attacks that they shouldn't release it publicly. I think a lot of people saw that decision as a very security-minded one, very prudent. But some of your colleagues have argued the opposite, that withholding the model could be worse for cybersecurity, could set a bad precedent.

18:29That's a little counterintuitive. Can you break down what the argument is for us? Yeah. So long term, a lot of the ways that AI can go really badly, I think, are made much worse if AI companies have access to AI systems that are extremely powerful and allow them to run rings around the rest of the world. So, for example, if AI companies are able to make use of AI systems to manipulate politics, manipulate regulators, do AI research and development much faster than their competitors could possibly do, I think this just makes it really hard for the rest of the world to pose any checks and balances on them.

19:03which I think is very concerning. And so I think basically it would be a very bad precedent if AI companies were systematically not deploying their best models to anyone in the outside world for a long time. I guess another point we could make about this is that Anthropic is currently not robust to high-effort security threats. They're not robust to high-effort state attacks aiming to steal their model weights. And they're definitely not robust to people trying to steal. So Mythos found this really long list of vulnerabilities in a variety of software, and there's no way that Anthropic is able to protect that list of secrets.

19:38So I think it's reasonably plausible that external hackers have already stolen the list of security vulnerabilities that Anthropic found using Mythos. I kind of have the attitude that if you aren't going to be able to secure your model or dangerous vulnerabilities found by your model, I kind of think you should not train the model in the first place. Sure, those are two points there. Got it. I guess if you get hacked and the model gets stolen, that's one way for the model to get out there. But you would probably say, well, that means it just ends up with the bad guys and not with the good guys or something like that.

20:10Yeah, definitely. Probably when this model is actually deployed, it'll be deployed using it with some cyber safeguards so you aren't able to use it for hacking. And obviously, if the model weights are out there, then we have all the costs associated with everyone doing hacking. We also get the benefits of the outside world having access to the model. And so it's possibly better than no one having access to the model outside Anthropic. But yeah, I would prefer the version where Anthropic deploys it publicly, but with safeguards to prevent people from doing cybercrime with it. Yeah, yeah, it makes sense.

20:39And then one of the announcements that came out of the conference over the weekend was that Redwood just released a new project called Linux Arena. Tell us about that. So one of the difficulties you have when you're researching control is you need basically benchmarks where you can test how well your AIs can sneakily accomplish one thing while supposedly being told to do another. So these are tasks that you could have given an AI agent or sorry, settings in which you can give an AI agent a task and measure whether it can sneakily accomplish some side objective. And it's always a problem in AI to have good evaluations, good benchmarks, good environments to get your agents to act in.

21:19It's always tricky to make those. And in our case, it's even harder because of the fact that we need to have these subtle side objectives to get the AIs to do. So Linux Arena is just a project we did over the last seven months to build a big set of these environments using human contractors so that we can more carefully see how good agents are at sneakily accomplishing side objectives while doing programming. Great. Well, thanks very much, Buck. That is Buck Slegeris, CEO of Redwood Research. Ahead of Anthropik's expected IPO, the information published an in-depth profile of the company's CFO, Krishna Rao, a former Airbnb finance executive who has largely stayed out of the spotlight despite his critical role in Anthropik's corporate moves.

22:01Sri Mupiti and Valita Pau join me now to talk about the report they wrote with their colleague Corey Weinberg. So Sri, let's start here. Why did you decide to write this profile? Why this guy? Why now? Totally great question. Krishna Rao, I don't believe has been on any tech podcast to date, which is very rare for a tech executive, especially at the scale of Anthropic. I think also we found out through our reporting that the last profile that Krishna has actually been in was something from 2005 in the Harvard Crimson where he had gone to college. And so I think this was an opportune time to actually write a profile on Krishna, just given that there is so much excitement about Anthropic, and especially in Krishna's position where he's in charge of the compute deals, the fundraising, as well as even a potential IPO if the company chooses to go public, Krishna is really at the seat of it all.

22:51And so I thought that it would be a great time for readers to understand who he is. Yeah. Wow. No tech podcast. He should come on our show. I know, exactly. Well, great. Let's hear a little bit more about those recent initiatives that he's been spearheading. Maybe over to you to start, Valida. So he assembled a team for acquisitions, is that right? What does that look like? So Anthropic has now make like some acquisitions and have a team focus on that. After a year, Krishna joined in 2025. He recruited Andrew Zloto from SoftBank, who used to work with him at Airbnb and delegated some tasks to him.

23:26And a year or so, they poached more former bankers and investors to work on that. Around the fall of 1025, we reported that they started telling bankers, hey, we are ready to do some deals. And unlike OpenAI, which bought tech podcast TBPN and struck some multi-billion dollar acquisitions, Anthropic has told bankers they prefer to write smaller checks of under$500 million and want companies with promising technologies or talented researchers. An example of this is this biotech startup called Coefficient Bio. They bought for around$400 million, the information first reported. And that deal could help Anthropic to go after the healthcare market, like using AI for drug discovery.