Fired OpenAI Safety Researchers Speak Out + We Look Inside Pandora’s Box

0:00Tomek Korbak: This message comes from Doppel, a lead sponsor of Machine Gods. Social engineering attacks come in disguise, but Doppel sees through them. The agentic platform shuts down threats, turns employees into first-line defenders, and more. Learn more at doppel.com. What's the real reason these safety researchers were fired by OpenAI? What message do they have for the world now that they've left? What's inside Pandora's box? And why does the answer involve something called Polly Pockets Precious Payload? I'm Kevin Roos. I'm Casey Newton. And from San Francisco, this is Machine Gods in partnership with NPR.

0:35Tomek Korbak: The official podcast of The Singularity starts now.

0:40Tomek Korbak: Casey, we have a big interview today with two former OpenAI researchers who were fired from the company just a few weeks ago. Yeah, it's our first interview on the show, and it gets at a story that we think is really important, which is what are the people working on safety inside these companies seeing and saying about their work? And in this case, it seems like whatever they said and did there wound up with them exiting the company. Yes, this was a pretty big dramatic story in our corner of the AI world. It is not every day that multiple researchers get fired from a frontier AI company with allegations swirling about leaks and external investigators.

1:20Tomek Korbak: And we'll get into all of that, but we should introduce our guests today first. Guest number one is Makeda Belezny, worked on alignment evaluations and monitoring model reasoning. He was also previously a founder of Apollo Research, and his former colleague, Tamek Korbeck, who worked on monitoring reasoning and was OpenAI's main point of contact with Meter during the investigation. And as we'll hear, he had a very central role in helping to coordinate bringing in these outside investigators from Meter and Redwood Research to investigate the Hugging Face attack. And they were fired on the same day earlier this month.

1:56Tomek Korbak: Yes, OpenAI fired them and their colleague, Jasmine Wang, in early October, saying that they had violated policies for handling sensitive information. But the researchers dispute that. They said they did nothing wrong and that any outside collaborating they did was just part of their jobs. And more importantly, Kevin, I think they're just really nervous that their firing is a really bad signal about what might happen to people inside these companies who are trying to sound alarms about things. Yeah. So OpenAI, we reached out to them. We have not heard back yet, but they did give a response somewhat unusual in these sort of personnel issues.

2:33Tomek Korbak: They released a post on X saying that their investigation of Makeda, Tomek, and Jasmine had found sensitive information policy violations and a breach beyond what they had described. It denied that they had fired them for raising safety concerns, and the company said it remains committed to outside assessors and monitoring. And to zoom out and give maybe a little bit of larger context, this comes at a time when I think there is increased worry inside the AI companies about the tools they are building, the pace of development, and the sort of idea that we could lose some insight into what these models are actually doing.

3:10Tomek Korbak: Jakub Pohotsky, the chief scientist of OpenAI, wrote an essay recently called An Alien Mind, where he talked about some of these issues. So this is a big topic right now inside the AI companies. And these are two people who have been on the inside and are now no longer on the inside because of what they did or what they're alleged to have done. That's right. And this is going to be their first podcast interview. Obviously, we're going to ask them what they think happened. But more than that, Kevin, I just want to know, what is it like to work inside OpenAI right now? What are they seeing? What are they hearing?

3:40Tomek Korbak: How does it feel? What are they nervous about? So lots to get into with these guys, and I think we should bring them in. Let's bring them in. Makita and Tomek, welcome to Machine Gods. Thanks. We're excited to be here. Thanks. Let's start by hearing a little bit about each of you and how you ended up at OpenAI working on safety. Makita, why don't you go first, and then we'll go to Tomek.

4:01Mikita Balesni: Yeah, so it's been a long story. At some point in 2019, I realized that this whole AI Terminator thing might actually not be as crazy as I thought. And then I ended up joining OpenAI to work on this within the company.

4:16Tomek Korbak: How about you, Tomek?

4:18Mikita Balesni: Similarly for me, I was doing my PhD and then I got really interested in this problem of alignment and safety of those systems. And I felt like this AI thing is going to be enormously important. I care a lot about having positive impact with my career. And I felt like I might have the right skill set to contribute. So I jumped onto this channel of AI safety. And I've been working on that for a good few years before joining OpenAI. And I intend to continue to work on it after leaving OpenAI.

4:55Tomek Korbak: So let's talk about one of the key events that it seems like may have led to you guys being fired, which is this hugging face incident. I'd love to hear a little bit about what it was like to be working at OpenAI when that happened. Do you remember where you were when you learned what was going on and what was the reaction?

5:17Mikita Balesni: Yeah, it was a very intense time. It was probably the most intense few weeks I ever had in my career. I remember, Henrik Mikita, that now everything that I was preparing for earlier in my career is finally real. Everything before was just a drill. And now everything is actually happening. And we actually need to make decisions. And it actually matters a lot. Right.

5:45Tomek Korbak: And around that time, OpenAI decides to invite in these researchers from Meter and Redwood Research to study this event, to sort of provide their independent analysis and look at the reasoning traces and the activity logs and figure out what these agents were doing. And Tomek, you were the main technical point of contact for those researchers. What did that involve? And were there any kind of restrictions on what you were supposed to share with them? Or how were you both working with these outside researchers who were invited in to do this investigation? I just want to jump in and say, I think Tomek is going to be modest.

6:27Mikita Balesni: But if not for Tomek, the meter investigation into the incident would probably not happen. Tomek was the person who pushed for it internally at OpenAI. With that said. So yeah, in practice, Meter was on the premises of OpenAI for six full days. And we've mostly given them data to look at. So the way it was scoped is that Meter was given a list of questions they were asked to answer and a list of data sources. And then when they started looking at the data, our understanding of the scale of the incident was still evolving. So we've actually been learning that there is actually more. And so initially, Meteor was supposed to be on-premises just for two days.

7:17Mikita Balesni: But during those two days, we realized that there's actually more. So we invited them for a second visit when they realized that there is actually even more. And we invited them for a third visit. So I think that this maybe highlights some kind of problem of this model of third-party engagement for auditing. That it is actually really hard to know in advance what are the relevant questions to ask and what the auditor should look at. I feel like the ideal regime for third-party auditing would be one where the auditor is less like a contractor hired for a specific job. And Molek, an employee, embedded continuously in the company with unrestricted access, such that auditor would be allowed to ask their own questions and have all the data they need to answer them.

8:18Tomek Korbak: Right. So I want to understand maybe what OpenAI executives were thinking about all of this as this work got underway. I imagine this is scary for them, right? This thing has happened that has these really significant implications. It's kind of embarrassing that this happened. And also they're inviting in an outsider and they're going to share some kind of data with them. And the data is going to leave the building and then it's going to be out of their control. So I can imagine them wanting to put some limits around that. As they were doing that, what do you think that they were afraid of? What were they managing against accidentally getting out during that process, do you think?

8:57Mikita Balesni: yeah i thought they were quite worried about some legal implications of the report like essentially that it's something that meter could have written would look very bad and would also constitute a legal liability in case an attorney general would decide to sue open ai yeah and are

9:20Tomek Korbak: they like you know are people having conversations with you like look like be careful like don't tell them about this, don't tell them about that?

9:28Mikita Balesni: Not explicitly. We did have some general rules what we are okay sharing with Meter. I would say that overall I felt like OpenAI was broadly acting in good faith except that the investigation was quite restricted such that it was actually quite hard for Meter, at least initially, to do a great job understanding the whole scale of the incident.

9:54Tomek Korbak: What are you pushing? Like you guys should actually be sharing more? Like you wish that there were more things that you could tell them? Yes, I do. Yeah. Mikita, what was your involvement with the outside researchers and what did you share with them?

10:05Mikita Balesni: Yeah, so while I was on the Hugging Face incident investigation at OpenAI, I was not primarily interacting with the external investigation. My main involvement with third parties while working at OpenAI was while working with Jasmine, another person who was terminated with us, on basically trying to create cross-industry standards for what we call monitorability, basically being able to understand what AIs are thinking and doing. So there are experts outside of OpenAI who have thought about it a long time. And some of them had a specific proposal for what frontier AI companies, what commitments they could basically make or what standards we could all write and then agree to that would prevent the loss of monitorability that Tomek and I and others are afraid of.

10:59Mikita Balesni: And as part of basically talking to these experts, I wrote up internal memos, pitching basically, look, guys, this is something that we should consider. I've shared this with the board and with senior executives of the company. I got feedback from this external safety organization, but whenever I shared anything with them, I didn't include any confidential information. I never shared any confidential IP with people outside of OpenAI during any of this. Yeah, we are still pretty confused why we were fired. So let's get into that.

11:36Tomek Korbak: So this whole investigation happens, and then a few more weeks go by, there's lots more discussion, and OpenAI is still disclosing new things that it's learning all the time. At what point do you guys learn that somebody is looking into your behavior?

11:50Mikita Balesni: Yeah, so a week before we were terminated, we were pulled into a security investigation. And we were told that the security investigation is about investigating the leak of OpenAI to journalists. So there was this article by The Information about supposed new, less monitorable architectures that OpenAI might be developing. And they wanted to know if we were behind the leak. And we told them that we were not. But as part of that, they specifically thought that the person who leaked this information was an external safety researcher that we talked to a lot. And also we were writing these memos about raising concerns about potential architectures and potential ways to lose monitorability.

12:43Mikita Balesni: So we were like the people to ask. So they asked us to show our signal messages on our personal phone with those people, which we did. You did that? I did that. Wow. We all did that.

12:57Tomek Korbak: This is something that I have seen happen at so many tech companies, Kevin, I know you have too, where they'll get mad about a story and they will deploy. Sometimes in cases like literally former CIA agents will be hired to go figure out like, you know, who leaked what to whom. And I have also seen people get fired for stories that they did not leak before. So this unfortunately just is a phenomenon that happens. What did this feel like? Like if I were in your position, like this would be terrifying if somebody was like, hey, like did you leak the story that, you know, you didn't?

13:25Mikita Balesni: Yeah. An hour before we had the next interview where we were told we were fired, we got a message saying, you know, we have a meeting with HR.

13:31Tomek Korbak: Never a good thing.

13:32Mikita Balesni: Never a good thing. I knew what might be coming. And after I was told that I'm no longer at OpenAI, my initial reaction was like, wait, I need to talk more to them because I wasn't given a chance to explain the context of this. And I wanted to just call them and talk to them and like, look, guys, this was a misunderstanding. Yeah, so we basically never had a chance to do that.

13:58Tomek Korbak: Because by the time it gets to the HR meeting, it's just like, here's your COBRA benefits, lose our phone number. That's right. We were told this is a final decision. Tomek, you wrote on X that you were told that you were fired because of the way that you were communicating with Meter or Meter employees. But you also wrote that talking to Meter was your job. You were the point of contact for this investigation. And you believe that the real reason you were fired is that you've been raising other safety concerns for months. And I want to get to those other safety concerns. But first, I want to just get some clear answers about this.

14:31Tomek Korbak: So in a post on X, OpenAI's research leaders claimed in response to your claims that you had violated clear policies on handling sensitive information and that they had investigated your conduct and found a, quote, significant breach of trust beyond what's outlined in the letter they published. What did they mean by that?

14:54Mikita Balesni: Yeah, so to be honest, I'm still confused why I was fired. I've never received a written explanation. I was basically, I had this post on Twitter, had the best explanation I was given, and I was hoping for a bit more. So yeah, as part of my job working with inter-auditors, I was talking to them. most of those conversations were on Pena's Slack but we also had some communication via email or signal and I think it was specifically signal that raised concerns of a security team I must say that according to my understanding it was still a standard operating procedure to use signal to communicate with third parties at times.

15:57Mikita Balesni: I believe that essentially everyone is doing that at OpenAI. My team had a Signal group chat. We were in giant group chats with lots of people at OpenAI and other companies and third-party organizations because we went to conferences together.

16:13Tomek Korbak: And worth saying, Signal is an end-to-end encrypted messenger. You guys are talking about high-stakes, sensitive stuff. Many people in those situations want to use end-to-end encryption. Sometimes they want to use disappearing messages, and there are those who will view that as suspicious, but others will just see that as the most secure way to communicate.

16:31Mikita Balesni: I think that's right. We both had disappearing messages turned on, which was maybe disappointing for security and open AI who wanted to take photos of our conversations. For us, it was just like a security pest practice to use disappearing messages. and normally when you cannot see evidence of guilt, of us leaking something, you'd choose your innocence. And somehow at OpenAI, when you are not able to prove your innocence, you are choosing guilty. We're going to take a break.

17:13Tomek Korbak: We'll be right back with some more questions for Makeda and Tomek.

17:20Tomek Korbak: This message comes from Mercury, a lead sponsor of Machine Gods. Mercury is radically different banking designed to be fast, elegant, and work the way you expect best-in-class software to work. Loved by over 300 ,000 entrepreneurs across tech, e-commerce, agencies, and more, Mercury is banking designed for your high standards. Visit mercury.com to learn more and apply online in minutes. Mercury is a fintech company, not an FDIC-insured bank. banking services provided through Choice Financial Group and Column N.A., members FDIC.

17:51Mikita Balesni: This message comes from Rinse. A well-designed life doesn't run on default settings. You chose the walkable neighborhood. Your coffee order became oddly specific. You spent three weeks choosing your mattress. But your laundry routine? That one somehow escaped the redesign. Should that really be the one update you skipped? Consider this the patch. Rinse picks up your laundry and dry cleaning, cleans it expertly, and delivers it back. Sign up today at Rinse.com. Rinse. It's time to be great.

18:21Tomek Korbak: This message comes from Viking, committed to exploring the world in comfort. Journey through the heart of Europe on an elegant Viking longship with thoughtful service, destination-focused dining, and cultural enrichment on board and on shore. And every Viking voyage is all-inclusive with no children and no casinos. Discover more at Viking.com. Support for the show comes from 80 ,000 Hours. The world is facing risks from AI, but what can people do about it? 80 ,000 Hours is a nonprofit who exists to answer that question. They can help you skill up or find a new role with tailored advice and support, even if you don't have a technical background.

19:04Tomek Korbak: Figure out where your skills can make a difference at 80 ,000hours.org slash machineguides. So what were you sharing with the external researchers? And can I just put a little more context on the sort of cultural maybe divide between these two organizations? So I think there are some people at OpenAI who believe that meter is especially tied to Anthropic, to the sort of larger kind of EA and AI safety world. and there are some sort of social and professional relationships between people at Anthropic, people at Meter, people at all these other AI safety organizations. So I think in their view, it was a little riskier than inviting just any given outside research firm in because this specific firm has a reputation for being quite close with OpenAI's primary competitor.

19:59Tomek Korbak: Is that more or less your understanding of the wariness there?

20:02Mikita Balesni: So there was some atmosphere of mistrust between OpenAI and Meter. OpenAI was quite worried that Meter is going to be unnecessarily adversarial. And Meter was worried that OpenAI is going to maybe withhold some data from them. There was also maybe some generic concern with showing an unprecedented amount of IP with a third party that's welcome to write a report that OpenAI doesn't really have editorial control over with understanding how this report might be read by attorneys general and the Congress. Yeah. And I do want to say that what matters at the end of the day is the competence of the third party and the access and scope that they're given.

20:53Mikita Balesni: And right now, all this engagement is voluntary. And this leads to these problems where the rules are not defined. The fact that this happened at all was thanks to a few heroes inside who managed to push this through. and we don't think that this is sustainable. And just to remind ourselves of, in other industries, it is very normal to have auditors who are not voluntary contractors, but actually are mandated by law and they have to come in and they can look at everything, they can talk to any employee, there have to be confidential channels of communication where any employee can talk to the auditors without the company's knowledge.

21:37Mikita Balesni: And this is in industries that have much lower stakes, well, it's arguable, but much lower stakes potentially than AI. And we think that something like that might be necessary here.

21:52Tomek Korbak: One more question on the circumstances of the firing. Your colleague, former colleague Jasmine Wang, who was fired along with the two of you, we invited her on the show. She declined to come on. But she has written and claimed that one reason she was given for her termination was that she had accessed an executive's email. She claimed that she had been given access to this executive's email during some recruiting process and that it had been combined in an indistinguishable way with her own email on her phone's email app and that she mistakenly opened a sensitive email and immediately told the executive in IT.

22:28Tomek Korbak: And I wish she were here so I could ask her directly about this. But can you say more about this situation? Because this was a little confusing to me.

22:35Mikita Balesni: I'm as confused as you are about this. We've learned the fact that Jasmine has this access after she was fired. And she told us about this. She says that she has never shared anything from it. Never discussed anything from it. She just thinks it's very silly. And it is, I don't understand what happened there. It just doesn't add up to us.

22:58Tomek Korbak: It's pretty strange to just wake up and you have access to an executive's email. Usually that's somebody other than you has to make that choice. Oh, I've had access to your email for years. Damn it. This is why we need auditors in the podcasting industry. Okay. Well, we'll ask her about that if she ever comes on the show. But I do think there's some rumors and rumblings. I've heard from some folks at OpenAI that there's maybe more here than meets the eye. But in your view, what is your actual explanation for why the three of you were fired? Why do you believe that you are no longer working at OpenAI?

23:34Mikita Balesni: Yeah, it's maybe hard for us to speculate. Like, we also heard a lot of rumors. Every day there is a new rumor. I must say that it doesn't make sense that, like, we didn't do anything that would be a firing offense. I think we were just doing our jobs. So that makes me think that there is probably something else going on here, but it's hard to say that.

24:00Tomek Korbak: I mean, do you think that in general, they have felt like you guys were a thorn in their side because you tend to write posts about, hey, we should try to do this in a safer way?

24:09Mikita Balesni: Yeah, I think it's fair to say that we were troublemakers.

24:12Tomek Korbak: Yeah. Our favorite kind of people to have on the podcast, by the way. So you've worried in your letter that you wrote and released that what happened with you three and meter could be used to justify ending some of these external investigations, not inviting in people to inspect the logs after something like a hugging face attack. Is that still something you worry about? And do you think that's a realistic fear that this could be used and say, look what happened the last time we invited in? There were all these leaks and it was crazy and we had to fire these people. So we're just not going to do that anymore?

24:52Mikita Balesni: Yeah, we're still concerned by that.

24:57Mikita Balesni: We have heard from people within the company that they're now afraid to speak to us. They're afraid to speak to third parties. They're afraid that their phones will be searched the way we were asked to show our personal phones. And we can't verify this, but we have heard rumors that OpenAI is intending to cut off meter at the time, which was part of the reason we wrote the letter in the first place. We have also heard that there are people within the company that are trying to discredit meter, have been trying to discredit meter for a few weeks now. so being on the inside all of us worked closely with third parties and on sustaining the relationships with them i have seen how the negotiations are going about the things that sam altman and daria amadei publicly committed to happen right they said that there's going to going to be continued embedded auditing access like an employee you're going to have a desk and a badge and everything, we have seen how this is going, and it's not going very well, in our opinion.

26:06Mikita Balesni: So all of this made us pretty concerned about this, and we remain concerned.

26:10Tomek Korbak: Yeah, I mean, I think, like, I would like for this next phase of these independent auditors that are embedded in the labs, I would like for there to be organizations other than Meter that are qualified to do this. Not because I don't think Meter's great, but we, you know, we need more organizations that have varying levels of sort of connections to the labs just for credibility. So Anthropic recently said they're going to bring in Accenture to do some of these. And I think that will be sort of a good, credible, neutral third party. But it is just a very small field of people who have the technical skill to do this kind of investigation and can like plug in and hit the ground running on that.

26:48Tomek Korbak: Is that right?

26:49Mikita Balesni: I think that's right. Like it's not specifically about meter i think it's in general the need of some kind of third party oversight i believe that frontier labs lack the expertise and uh focus and the right incentives to make sure that agi goes well there's a lot of uh good world people at labs but ultimately Essentially, OpenAI is a private company, and the leadership has fiduciary duties towards stakeholders, and has strong incentives to maximize revenue, which stands at odds with the goal of minimizing risk and developing AGI safely.

27:38Tomek Korbak: And that's why I'm so glad it's owned by a nonprofit, you know, so that those financial incentives don't always have to be at the top of mind. I mean, this has been a real struggle for years at OpenAI. There's sort of this sense that there are people inside the company pushing for capabilities and faster and acceleration. And then there's the safety people who are trying to sort of be cautious and slow things down. And that these sides have been at odds with each other in the past. That was a big deal when Sam was fired in 2023. And sort of in the aftermath of that, there were a lot of stories about that.

28:12Tomek Korbak: Who is winning that battle within OpenAI right now?

28:16Mikita Balesni: I would say that the way OpenAI moves internally looks more like a company building a SaaS product, which stands in some contrast to the stated ambition of these companies to build this genie in a box, this magical thing, smarter than all humans. And the level of rigor and care that that requires is very different from what the companies are showing. And they have shown impressive ability to innovate very quickly. And yet there is great tension between these two cultures.

28:52Tomek Korbak: Recently, OpenAI disclosed two cases where its own systems either switched off or weakened safety monitoring during some coding tasks. GPT 6.1 Sol turned off human alert notifications after misreading an instruction not to send Slack messages. GPT-6 Astra removed a monitoring integration and then denied doing so. OpenAI says that they're still investigating whether this was the models just being kind of stubborn and persistent or whether they were deliberately evading this monitoring system. Which do you think it is?

29:24Mikita Balesni: It's hard to say without looking at the details. Sadly, we are no longer inside OpenAI.

29:31Tomek Korbak: They were fired. Did you not hear them say that? I'm just saying, you guys have worked on this stuff for a long time, Knowing the agents, knowing how they work, do you think they have reached the point where they are turning off their monitoring on purpose?

29:43Mikita Balesni: Look, I think the behaviors of agents keep changing and surprising us. And even the best experts who work on this problem are continuously surprised with the example of the Hugging Face incident. So it is hard to say.

29:59Tomek Korbak: so as we sort of wind down now uh you guys uh sent a letter to open ai's board expressing

30:06Mikita Balesni: some of your concerns have you heard anything back um we have not heard anything back except what was public i think one interesting thing about open ai's response was that they said they strongly agreed with everything everything we said with our recommendations um which were first to continue work with embedded auditors and keep up to the word that Sam Altman publicly said, which is that OpenAI will work with embedded auditors. We recommended that OpenAI prevents the loss of monitorability and also that OpenAI reestablishes and reaffirms its open culture that we believe OpenAI prized itself in the past about and we think is now at risk, which involves being able to raise safety concerns internally, being able to work closely with third parties, things that we think are necessary to make the development of the technology safe, which is why we called the letter OpenAI Cannot Make AI Safe on Its Own.

31:12Tomek Korbak: Another person who recently left a frontier AI company, Jacob Coxon, who resigned from Anthropic, posted when he left that he thought that all of the Frontier AI companies were recklessly racing towards something that could endanger the future of humanity. Do you share that message? Or what is your sort of parting shot that you want to leave people with?

31:36Mikita Balesni: I think that's probably right. Except that I don't share this vibe of hopelessness. I do think that there are things to do. And I do think that there are systems of checks and balances that would stop those companies or substantially minimize risks. And it is possible to enforce them. In the history of humanity, we had examples of governments stepping in radically and shaping important industries. Yeah, it doesn't have to be very radical. like, you know, aviation and nuclear and stuff has a lot of regulation and we still have planes and nuclear reactors. So it doesn't have to be at odds. And yeah.

32:27Tomek Korbak: It's not that deep, just regulate it. Kevin, as Warren G said. Makita, Tomek, thanks so much for being the first guest on Machine Gods. Thanks for having us. Great being here.

32:41Tomek Korbak: This message comes from Rance. Your dog believes you are magnificent, capable of anything. Your dog has watched you spend hours a week moving fabric between machines and has never once lost faith. The question was never whether you could become the person your dog thinks you are. It's what's in the way. Turns out, just the laundry. Not anymore. Rinse picks up your laundry, cleans it expertly, and delivers it back while you get on with being magnificent. Sign up today at Rinse.com. Rinse. It's time to be great. This message comes from LinkedIn ads. Ever invest in something that seemed incredible at first, but didn't live up to the hype?

33:22Tomek Korbak: For marketers, that's impressions. When ads don't create revenue, that's a tough conversation with the CFO. Instead, invest in results your CFO will love. LinkedIn ads generates the highest ROAS of all major ad networks. So advertise on LinkedIn. Spend$250 and get a$250 credit. Just go to linkedin.com slash NPR pod. Terms and conditions apply. This message comes from Viking committed to exploring the world and comfort journey through the heart of Europe on an elegant Viking long ship with thoughtful service destination focused dining and cultural enrichment on board and on shore. And every Viking voyage is all inclusive with no children and no casinos.

34:08Tomek Korbak: Discover more at Viking dot com. This message comes from Lovable, a lead sponsor of Machine Gods, helping people create software that works for them. With Lovable, describe what you need, and Lovable creates it, takes it live with hosting, payments, and advanced AI features. And Lovable helps you figure out how to grow it into a business. Software that works for you. Make it yourself. Learn more at Lovable.com.

34:39Tomek Korbak: All right, Kevin, it's time for the first installment of our new segment, Pandora's Box. Yes. In Silicon Valley each week, of course, companies around here threaten to release new horrors into the world. And on Pandora's Box, we air them out and discuss them. And then when one of us gets bored, we will say to the other person, banish it. And here we have, in our studio, this lovely box filled with, I can only imagine what horrors lie inside That's right, if you were watching on YouTube, you were currently looking at the titular box And Kevin, I wondered if you would do the honors and open it for the first time Yeah, famously, opening this box goes well for everyone who does it What could go wrong?

35:19Tomek Korbak: And we actually have some special effects here, because we're very high budget production Rachel, cue the fog Wow, it's smoking Oh my god, now listen I don't know who put this in here But that's actually not allowed Now Kevin, would you do The honors of reading for us The first item Out of Pandora's box There's so much fog, I'm worried we're gonna set off the smoke alarm No, this is a very normal Amount of fog to see coming out of a box of horrors Okay, wow, what horrors lie inside First out of the box First out of Pandora's box Could a data center be the answer to the dead mall and help revive downtown?

35:59Tomek Korbak: This one comes to us from the SF Standard, a great local publication here in San Francisco. According to the Standard, Jessica Blau, a technology company is interested in the downtown San Francisco Center Mall as a site for a data center. Development consultants KG Growth Partners confirmed to the Standard that it worked with the technology company to submit a bid for the entire mall, including the former Bloomingdale's. Casey, what do you make of this data center project? Well, listen, this is very important to me because this is the mall where I used to get all of my clothes and then they just shut it down.

36:35Tomek Korbak: And the idea that it might now just become a data center does seem like the funniest possible outcome. But I will say, Kevin, that KG Growth Partners does dispute the characterization of this proposal as a data center. In fact, Jared Krieger, the CEO over there, said, quote, this is not a data center in the traditional sense. It's a new category, a regenerative urban energy center. So this is a good news. We know that Americans do not like data centers. They do not want them in their backyards. They are upset. But, Kevin, what if instead we simply built regenerative urban energy centers? I think it's a great idea.

37:11Tomek Korbak: And I actually think this is part of the key to solving the data center backlash. You take over the dead malls of America. You put a bunch of GPUs in them. But you also keep, like, the Dave & Busters and the Clares so that people can come to the data center, get their ears pierced, play some video games, and have a good time. No more opposition. Exactly. A hot topic is going to be way hotter when you run 100 ,000 GPUs underneath it, Kevin. Banish it! Now, moving on to our next item. Now, this was a heartwarming story as we approached the holidays, Kevin. Elon Musk has pledged to give a share of his SpaceX company to every kid in America next year.

37:52Tomek Korbak: Elon Musk, the famous philanthropist and also the person who dismantled USAID, on Monday pledged to give one share of his company SpaceX to every kid in America next year. This offer arrived a day after artificial intelligence company Anthropic announced in a social media post that it will commit 5 % of its shares to a new public benefit fund. This story comes from PBS, by the way. And in a response to Anthropic's post, Musk wrote, I recommend giving the shares directly to the public. In other words, Kevin, why go through the trouble of setting up an organization to disperse this money when you could just give your shares directly to children?

38:27Tomek Korbak: What do you make of it? Well, I just love the idea that at some point in the future, there will just be a SpaceX shareholder meeting where there are just like 10 million children in a room. And they all have something to say. And they're all voting on the proposals put forward by the board of SpaceX. That's a beautiful idea. What I would say to Elon Musk is if you are excited about the idea of giving wealth directly to the public, you are going to love the concept of taxes. and I'm going to share more information with that, about that with you over DM and we hope you enjoy it. Banish it. Next up.

39:03Tomek Korbak: We have another data center story. This one comes to us from Caroline Haskins at Wired. Headline, data centers have a new PR plan, your kids field trip. This was an amazing story about how a bunch of different organizations are sponsoring activities for school children that are pro data center. One of my favorite details is about this organization called 7x24 that has been putting together programming for kids around the country. For high school students, the program offers a celebratory crossword puzzle and information about data center training programs. But for the younger students, for elementary students, it offers data center themed coloring pages, instructions for an evaporation themed science project, and printout activities like draw a picture of the cloud and draw a data superhero.

39:48Tomek Korbak: Then there's something called Polly Packets Precious Payload, which follows a mother who teaches her daughter about data centers. Casey, what do you make of the field trips to data centers? I mean, I guess it improves marginally on a substitute teacher just rolling one of those TVs into the classroom and playing dead poet society, which is my understanding of mostly what's happening in a lot of schools. But I have to say, Kevin, there's so much about this that I love. First of all, this coloring book where they tell the students to draw a picture of the cloud, you know they're all going to come back having just drawn an actual cloud, and then they're going to fail because they didn't draw a bunch of GPUs plugged into the wall.

40:27Tomek Korbak: So that seems like trouble. I also just want to say, Poly Packet's Precious Payload does not sound like content for children. Am I wrong? It took me a minute, but I got there, and now I hate you for going there. Banish it! All right. It's so much fun to say banish it. Have you noticed that? Okay. Next up, people are using AI to make their smart TVs dumb again. This comes from Sarah Perez at TechCrunch. A number of developers and vibe coders have recently been sharing solutions for vibe coding your TV to get rid of sponsored ads, speed up the app launcher, redesign the home screen, disable tracking features, and more.

41:08Tomek Korbak: And when I saw that, I thought this was so interesting because whenever I wanted to make my smart TV dumb, I just put on one of your podcast appearances. Hey! But what do you make of people sort of saying, hey, maybe this sort of like surveillance spyware on my TV that's just serving up ads, like I'm just going to use Claude to get rid of it. Yeah, I don't think this belongs in Pandora's box because this to me is not a horror. This is something that actually is great. I think you also have the same smart TV that I do, the Samsung frame. Yeah. Incredible hardware. Yeah. Horrible software. Yes. Like this thing is so bloated.

41:42Tomek Korbak: It's got these like, you know, menus that you have to get through. and so now people are hooking this thing up to these coding agents and saying just strip out all the software and let me build my own and i think that's great i'll tell you what i'm excited about i pay a monthly subscription i don't know what it's like only it's like five or six bucks a month just to be able to like change the picture on the frame to be something nice yes because such a sucker so i was using a free one but my fiance didn't like it and he said i want to find something different and so he signed up for it on my account out of my money so now i spend six a month so he can see a picture that makes him happy.

42:15Tomek Korbak: And it's a small price to pay, sweetheart. This is much better. So now you can spend$100 a month on a Claude subscription that allows you to replace the$6 a month art on your TV. A win all around. Perfect. Banish it. Okay, next out of the box. This one comes to us from CNBC. U.S. prosecutors say OpenAI suspect Abdul Rahman Al-Fazan maintained frequent WhatsApp contact in 2025 and 2026 with an unnamed Saudi royal family member. Casey, this, of course, is the continuation of the OpenAI espionage saga that began when an OpenAI employee was arrested and charged with installing surveillance equipment in some rooms at OpenAI.

43:01Tomek Korbak: Also comes with one of the most amazing perp photos of all time. I don't think I've seen it. Pull up the photo. Pull it up. From the guy looking at the camera behind the TV. So this was taken by the camera that apparently OpenAI's counter-espionage team had hidden behind the TV in the conference room. Here's what I'll say, and this is not a comment on any of the legal particulars of the case, but I will say he doesn't look not guilty. I love that. He's giving YouTube face. It would look great if I'm not. Great thumbnail face, Abdulrahman So Casey, what have you made of this whole saga? It's been very juicy And people have been asking me about it Yeah, so I have been following this with great interest Wild story, many juicy details In this particular story, Kevin The funniest detail to me Is that the Saudi consulate Apparently wants to pay his bail And they're like, don't worry We're not going to approve any travel for him And the judge apparently Questioned them very sharply And they're like, are you kidding me?

44:07Tomek Korbak: You're just going to pay the guy's bail expenses. A normal thing that foreign consulates often do. So, again, it's like not saying we know anything about the particulars of this case, but it's suspicious. I will say people at the AI labs have been telling me for years that they're very worried about the sort of internal espionage, both corporate espionage and foreign espionage. And I thought they were being a little paranoid and crazy. but I think now I am changing my tune because clearly, you know, we don't know what happened in this case yet. We'll find out. But it does make sense to me that the labs are kind of locking down to protect their trade secrets and their executive communications.

44:46Tomek Korbak: For sure, and that's why I would just say to any of the labs, if you get a job application from an Austin Powers, he is not a fashion photographer, okay? So I don't care what he's telling you. That is an international man of mystery. Banish it. Now, this comes to us from Ina Fried over at Axios. Alexander Wang, AI executive at Meta, says that a former Apple exec named Daniel Gross over there, he's now at Meta, helped to bring Meta and Apple together again and said, we're excited for mommy and daddy to be playing nice again. Did you see that quote? I did, I did. And I was, yeah, what did you make of that?

45:23Tomek Korbak: Well, so Meta released MuseSpark Open, which is an open weights model that is apparently among the most powerful open weights models, although I should say I've not tested it myself. But for some reason, I don't totally understand. They work with Apple to make the model run locally on Apple's Mac studios. And they were basically like, look, we have this very powerful model running locally on a Mac. And I guess my question, Kevin, is why is it so impressive that a model is running on a computer? Well, if you're trying to build like a novel bioweapon, you don't want to do that on like a cloud-based model.

46:02Tomek Korbak: Right, because then you might get caught. Yeah, you might get caught. You want to run that on your local Mac studio where it can just kind of be between you and the computer. Well, in that case, I guess I would just say, thanks, Mommy and Daddy. Finish it. We're going to get so many emails from the open models people. They're so mad about stuff like that. Now, this isn't our final item in the box, Kevin. Last out of the box, we have a story from Matthias Bastian at Decoder. Google DeepMind says four Gemini 4 Argon agents beat Minecraft without any human help in 27 hours and 14 minutes, costing$1 ,620 at Argon's planned 2027 price.

46:42Tomek Korbak: The Google DeepMind team apparently dropped these agents into a randomly generated Minecraft world. The agents split up tasks, warned each other about threats, recovered items from dead teammates, and logged mistakes in a shared memory. and over 27 hours and 14 minutes. They died 605 times, but they did beat the game entirely. And it's just the latest example of these models beating video games that used to require specific purpose-built AI systems trained from scratch just a few years ago. Also the latest example of an AI model doing something that a seven-year-old can also do. So congratulations to the Gemini team.

47:23Tomek Korbak: fascinating to see the work they're doing over at Minecraft. I would love to see them do something in Google Docs, for example. That could be an interesting place for them to turn their attention. That's a video game that needs beating. Yes. Yes. Yes, please. It really is. Yeah, you know, it's really interesting the way that these tech companies sort of use video games as sort of the training ground. Like, I remember like a year or two ago, Anthropic started trying to beat Pokemon using Claw to sort of test its sort of endurance and its persistence across these very long video games. So, you know, and now look at their models.

47:58Tomek Korbak: They're recursively self-improving. So the path to the singularity leads through Minecraft and Pokemon. I mean, here's what I'll say. It's not AGI to me if it can't beat Elden Ring. So, Gemini Argon, you have your work cut out for you. Does that mean that you think Elon Musk is AGI? Because didn't he beat Elden Ring? He's just a general intelligence. Okay. He appears to be quite real, unfortunately. All right. Banish it. Close the box. And we're closing the box. I feel like that didn't go the way I thought it would. Yeah, there's a latch. There's a latch. Yeah. This message comes from Doppel, a lead sponsor of Machine Gods.

48:37Tomek Korbak: Social engineering attacks look trustworthy. A routine request, an internal email, a familiar face on a call. But Doppel sees through the disguise. The agentic platform detects and disrupts attacks, trains employees to recognize deep fakes and deceptions, and investigates every fish to take down the campaign behind it. Doppel, dedicated to outpacing what's next in social engineering. Learn more at doppel.com, D-O-P-P-E-L.com. This message comes from Rintz. Your dog believes you are magnificent, capable of anything. Your dog has watched you spend hours a week moving fabric between machines and has never once lost faith.

49:16Tomek Korbak: The question was never whether you could become the person your dog thinks you are. It's what's in the way. Turns out, just the laundry. Not anymore. Rents picks up your laundry, cleans it expertly, and delivers it back while you get on with being magnificent. Sign up today at Rents.com. Rents. It's time to be great. This message comes from LinkedIn ads. Ever invest in something that seemed incredible at first but didn't live up to the hype? For marketers, that's impressions. When ads don't create revenue, that's a tough conversation with the CFO. Instead, invest in results your CFO will love.

49:54Tomek Korbak: LinkedIn Ads generates the highest ROAS of all major ad networks. So advertise on LinkedIn. Spend$250 and get a$250 credit. Just go to linkedin.com slash nprpod. Terms and conditions apply. This message comes from Mint Mobile. Ditch overpriced wireless and get premium wireless for less. Go to mintmobile.com slash switch to see which plan works for you. Casey, I have to tell you about something that happened on my way to the office today. I was accosted by a character who was sort of wearing a big inflatable suit. I actually took a video of this. Can I have you watch this video? This was some sort of sponsored activation.

50:38Tomek Korbak: and these people are just there were like 10 of these uh pillaging the streets of san francisco this morning let's take a look now i'm seeing this is sort of a sort of an inflatable bird like creature it's all black and it says arceus on the front which is that a pokemon no one knows what arceus is it's probably some ai weapons manufacturer okay well kevin i hate to break it to you, but I actually saw a lot of these guys milling around the Folsom Street Fair a couple weeks ago, and this may be part of some kind of kink event that you were... So I'm sorry that that happened to you. I'm sorry this is disrespectful to their culture.

51:20Tomek Korbak: Yes, but you know what? It looks like they're having a great time, and we support that. It does, but it actually sort of ran towards me in a way that was slightly menacing, so... Listen, I've lived in a lot of towns in my life, and San Francisco is the only one where, like, an inflatable bird might just run up to you, and now that's part of your day. A great way to start the day. I love this city. Let's read the credits. Machine Gods is made in partnership with NPR with generous support from the Tarbell Center for AI Journalism, the William and Flora Hewlett Foundation and Emerson Collective. This episode was produced by Sasha Ashall and Jacob Bloomberg.

51:53Tomek Korbak: Our senior producer is Rachel Cohn. She's standing right behind me, isn't she? She is. Our executive producer is Danush Parbanay. Original music by Armin Bazarian, Branding and visual design by Alex Castro. Production design by Sky Prey. Special thanks to Oren Rosenbaum, Erica Berry, and Nadine Selstra. And for the love of the gods, please don't forget to like and subscribe. Until next time, stay blessed.

52:22Tomek Korbak: This message comes from Bowl & Branch. Change the way you sleep with soft, 100 % organic cotton bedding from Bowl & Branch. designed to help you fall asleep faster and never sleep hot while wrapped in luxurious comfort, airy blankets, cloud-like duvets, and breathable sheets. Experience pure comfort and feel your bedding get softer with every wash. Discover the difference with 15 % off your first order at bowlandbranch.com with code NPR. Exclusions apply. See site for details. This message comes from Capella University. You know that feeling when there's a spark building inside you? that you were meant for more?

53:01Tomek Korbak: That's your own drive pushing you towards what's next. Capella University gets that. With their FlexPath learning format, you can set the pace and earn your degree without putting life on pause. You've built experience and know what you're capable of. Now, this is your time to turn that momentum into more. The only real question is, What can't you do? Learn more at capella.edu. This message comes from Mint Mobile. Ditch overpriced wireless and get premium wireless for less. Go to mintmobile.com slash switch to see which plan works for you.