BRSL Weekly Brief
Your weekly brief on current events from the Berkeley Risk and Security Lab.
BRSL Weekly Brief
AI News Roundup: Google/Pentagon deal, U.S. data centers, and Mythos
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
This week on the BRSL Weekly Brief, Senior Research Scholar Sarah Shoker discusses three recent AI headlines: the Google/Pentagon agreement for use of Gemini in military contexts, the future of data centers in the U.S., and Anthropic's decision to limit access to its Mythos model.
Hi. Welcome back to the Berkeley Risk and Security Labs podcast, the BRSL Weekly Brief, where we bring you the latest information on current events from our lab experts. I'm the lab's communications manager, Vivian Bossieux-Skinner, and I'm here with Senior Research Scholar, Dr. Sarah Shoker, today we're going to do an AI news roundup and kind of do a deep dive into a few topics that you might have heard about in the news recently related to the AI sphere. So we'll start with the Pentagon/Google agreement. Google reached a deal recently with the Pentagon to allow use of its Gemini models for military purposes, and that's kind of sparked a lot of employee backlash at Google. What do you know about the details of this partnership between the Pentagon and Google, and how similar or how different
Sarah Shoker:Yeah. Thanks so much Vivian for having me on. So it is to the one that anthropic ended up negotiating last month? I think in terms of the partnership with the Pentagon, Google's agreement looks very similar to what OpenAI signed and also X AI, in that they accepted the all lawful uses terminology that was at the center of the controversy that we saw in February. So Google spokespersons have stated that they've accepted the industry norm of the two red lines that were originally identified by Anthropic and just as a reminder those two red lines, if I can summarize very broadly, is no fully lethal autonomous weapon systems without appropriate human oversight, and also no mass surveillance on American persons. So it seems like the agreement is very similar to other agreements that have already been signed in the industry. I think there are certain questions about which models are being offered on the classified network, and whether those models will be so called rails free models, or if they will have the same type of safety mitigations that are, you know, that are viewed and interacted with by their consumers.
Vivian Bossieux-Skinner:Okay, interesting. So it sounds like they kind of negotiated something similar, but didn't really go through the same open backlash, because I feel like that was a big part of the Anthropic and OpenAI negotiations was this big societal backlash against how things were going. And it seems like Google kind of avoided a lot of that until the last moment.
Sarah Shoker:Yeah, it seems that Google was really able to stay out of the limelight in a way that anthropic and OpenAI were not, to best of our knowledge, as members of the public, there were around 100 Google employees who sent a letter to Jeff Dean, who is a chief scientist at GDM, beseeching him to reject access of Google's models to the Pentagon's classified networks. And then, most recently, last week, there was a public letter, I believe, signed by hundreds of Google employees beseeching Sundar Pichai to not allow access to the Pentagon's classified networks. And as a reminder, Google already has enterprise has an enterprise contract with the Pentagon, but that particular contract had not previously allowed access to the Pentagon's classified networks. This would be an update to what Google had previously agreed to Interesting.
Vivian Bossieux-Skinner:Okay, and then so you were mentioning the open letter. Can you say a little bit more about what the employees were actually asking for, and what this means for the industry in general.
Sarah Shoker:Yeah. So they were essentially asking for a for the two red lines that had been previously identified. But the Google employees also noted that their concerns went beyond those two red lines that we previously discussed. They didn't really specify what those additional concerns were, but I think you know in itself that that I think you know, is worth noting, because as a result of the conversation surrounding these two red lines that open AI and anthropic had agreed to, I think there was this public backlash and, you know, and a perception that, you know, the bar was actually quite low, like these two red lines, these companies were not actually asking for a lot, and in some ways, they were acting as a red herring to some more pressing concerns around the integration of large language models into decision support systems, as we've seen with its usage in the war in Iran, so that there was so there was that request, but really it centered on, really, it centered on access to classified networks and a perception that they would not be able to adequately monitor usage and. If these large language models were used on the Pentagon's classified networks.So I think, broadly speaking, what this says about employee employer relations in you know, in industry is quite telling. Listeners may remember that in 2018 there was an employee walkout from Google employees specifically related to the project Maven contract. And in fact, that walkout, that protest was was quite successful in that Google did not renew the project Maven contract. And then, of course, we know, you know, has from historically that that project was then taken up by Palantir. So I think the difference between then and now really illustrates the changing relationship between employers and employees in Silicon Valley. I mean, I'm certainly not the first to note this. There has been some reporting on this particular subject from the New York Times looking at how employees have essentially adopted a more conservative approach towards expressing themselves in in Silicon Valley, companies and that employers as well are more reluctant to adopt employee feedback. So I think that's potentially one reason why we see this discrepancy or difference, between what is happening today with Google and the Pentagon and what we saw in 2018 Interesting.
Vivian Bossieux-Skinner:Okay, so we wanted to shift a little bit towards mythos there. So there's been a lot of conversation recently about anthropic decision to limit access to their new model mythos because of its high performance in cybersecurity. So how do you think this decision will affect other industry players with their own choice to deploy increasingly capable models like this?
Sarah Shoker:Yeah. I mean, I think anthropic has done something quite clever here. So historically, they've favored enterprise customers, whereas OpenAI has, you know, has typically favored, you know, the average consumer. Open AI is shifting, certainly, and trying to attract more enterprise customers, but anthropic that has been, I think, core to their strategy from the very beginning, and now that they've essentially had a very limited release of mythos, right? So I think a consortium of over 40 companies have access to this model. What they've essentially done is they've forced other industry players to respond to what anthropic has already done. So we know that other models are eventually going to reach similar capabilities. That has just been the trajectory of AI development, and I think now the ball is in the court of companies like OpenAI and Google. Are they now? Are they going to release publicly models with similar capabilities to mythos, or are they going to adopt a more limited, a limited strategy, and in doing so, are they going to be able to frame themselves strategically as responsible actors. What are the comms going to actually look like when increasingly capable models are released, when we already see what an alternative deployment strategy looks like, as demonstrated by anthropic. So I think strategically, or you know strategically, what anthropic has done is actually very interesting, and I'd be curious to know how other players are going to respond in the near future.
Vivian Bossieux-Skinner:Yeah. And kind of taking that a little further, what is the impact of creating and then restricting access to models like this across different countries or different like companies or society, kind of like generally, what's the impact of that?
Sarah Shoker:Yeah, so, I mean, I think one of the arguments for limited a limited release, is that you're able to obviously pick and choose the enterprise customers who have access, and you're able to empower defenders, rather than, you know offenders or you know malicious activity broadly. The question is, with the help of mythos, whether defenders are going to have a sufficient Head Start open source typically lags six months to a year. And certainly there are other companies in the U.S. who, you know, who may choose to you know, who may choose a different deployment strategy. So the field of cyber security generally is going to, I think, change pretty, pretty dramatically, assuming that the mythos capabilities are what anthropic says it is. I just think we're going to encounter a very different internet and different vulnerabilities or bugs that may not have necessarily been worth the time of, you know, human cyber security operators might now have to be prioritized in a way that they were able to, you know, neglect or deprioritize before. So I think on the business side, or from a business operations side, the everyday workload of cyber security specialists may end up changing. But I think too, you know, I'm openly speculating here, but I'm a little bit concerned about the, you know, the role of open source in this particular, in this, you know, in this particular dance, because I don't think we're going to live in a world, a permanent world where only a specific set of enterprise or industry actors have access to these models, by any means. And certainly I think that's, that's the world you know, a number of security specialists are envisioning too.
Vivian Bossieux-Skinner:Oh, that's, that's fascinating, and I'm curious to hear more about this when we talk with your colleague in a few weeks on the podcast. So stay tuned for that.
Sarah Shoker:Yeah, surprise guest!
Vivian Bossieux-Skinner:..and we'll go more in depth on this topic in just a couple of weeks. So subscribe for that. So now we wanted to talk a little bit about in this news roundup that we're doing. Let's shift to data centers. So their data centers have been in the news a lot lately, you know, even over the last few years, but more recently, because Bernie Sanders and AOC have called for federal data center moratorium. But you know, that's a big part of the AI industry, and a lot of companies are trying to build bigger data centers, more data centers, and, yeah, these moratoriums are kind of meant to push back on that. So what are data center moratoriums meant to accomplish? And does this signal something about American politics?
Sarah Shoker:Yeah. So the moratoriums, I mean, they're not only happening at the federal level, right? So at the local level, there have been over 100 data center moratoriums. Maine just became the first state to issue a statewide moratorium. So in in some sense, you know, Bernie Sanders and AOC are laggards in the data more data center moratorium discussion. Or, you know, dare I say, data center moratorium race. So what they're meant to accomplish? Well, it depends, right? So at the local level, there have been a variety of concerns expressed, and that ranges from, you know, the typical, you know, they're loud, they are not employing as many people as, you know, as data center operators state or certainly, the jobs that are being offered by data center construction are for a limited term. Because certainly, you don't actually need all that many employees in the long term to staff a data center. You know, at the federal level, the conversation really does seem to be about risks and trying to give enough time for safety to catch up with AI model capabilities. And that has been the language of Bernie Sanders and AOC. So I think you know, depending on where, you know, at what level of government you're looking at the, you know, the conversation around moratoriums looks a little bit different. And you know, there's, of course, in the environmental impact that people at the local level have been, have been wondering about, including energy usage utility bills rising. And then, of course, you know the ever you know, the permanently controversial question about water consumption, or water usage, by by these data centers too. So what I'm hoping people get from this is that there is, in fact, a range of concerns that moratoriums are meant to address. And as a result of, you know, as a result of of this big range of concerns, it's probably no surprise that it has united people from, you know, all, all, all sides of the, you know, the political spectrum, and so especially at the local level, it's, it's not, you know, it's not easily or it's not clear that the conversation is between blue and red. I think, in fact, what has happened is that the data center moratorium conversation has really made for some very interesting bedfellows, and has at the local level, especially altered some typical or traditional voting patterns. So I think there's been, you know, an attempt at the federal level by some actors to try and articulate data center moratoriums as a as a typical, you know, left wing concern. I'm not sure if that will catch with, you know, with the U.S. public at the moment, though, I think data center moratoriums and you know, who supports them, and whether it's connected to a particular, you know, vote, voter behavior, I think that's still all up for grabs and up for questioning. You know, I haven't done the survey on this myself, but anecdotally, it seems like questions around utility bills and environmental impact is not distinctly, you know, Democrat versus Republican, but you know 28 is coming up, and we'll see if that holds. Yeah, it's a fascinating time, for sure, yeah.
Vivian Bossieux-Skinner:And how much would kind of need to shift in how AI companies do business, or how AI works, I guess. If there were a data center moratorium, and kind of, what would that look like for the industry?
Sarah Shoker:Well, I think it's not clear to me whether companies would go elsewhere for their data center needs. There was a, shall we say, relatively viral, or what widely read, widely read piece recently authored, I think, the Jacobin arguing that these companies would just go to other countries for their data center requirements. You know that might be true for certain compute needs related to inference. I'm not sure about model training. On the consumer side, you may end up seeing model some uneven experiences with with your model. So sometimes model performance will degrade, and that can be the result of insufficient compute in those moments. But you know, by and large, the argument is that if you restrict data center access, especially at the federal level, this is sort of the argument, you know, given by Bernie Sanders, is that you will get model providers to their training, and the deployment of ever increasingly capable models will stall, or certainly you'll, you'll add friction to that process. I think that's probably right as a theory. I mean, whether it's desirable is another question entirely, of course, but I think it would probably create very like significant problems to you know, to you know, to their commercial activities. So I don't necessarily know how they would get around that, other than, you know, the same mechanisms that they have always used, which is, you know, in some cases, we've like, actually, just recently, we saw OpenAI get rid of Sora, in part because the computational requirements for enabling Sora were just, you know, or just through the roof, and just did not make sense commercially. So we may end up seeing in the language of one of OpenAI CEOs, a reduction in side quests and just a lot more, you know, strategic focus on what models are actually deployed publicly and where the research bets should be placed in the organization itself.
Vivian Bossieux-Skinner:Interesting. That's fascinating. And maybe, you know, maybe we can expand on that too in a future podcast as well, because I think it's probably going to be in the news a lot more these days, and especially, like you mentioned, with the election coming up in just a couple years. So, yeah, this has been a really fascinating conversation. Thank you for joining me on this podcast. Is there anything else that you know we talked about that you want to say anything more about, or anything like that?
Sarah Shoker:I mean, you know, like comment, subscribe.
Vivian Bossieux-Skinner:Yes, very important. Make sure you subscribe to the podcast, you can find it on any of the podcast platforms, and we try to post every week or every other week, and kind of give you some you know, updates on the news and what you need to know from our experts. So thank you again for joining me, Sarah, and we'll talk to you very soon with more updates on mythos.