Trump Lifts Claude Fable 5 Ban, Anthropic Preps Agent in Microsoft Teams, Elon Musk’s Inner Circle

0:13Welcome, everyone, to The Information's TITV. My name is Akash Pasricha. It is Wednesday, July 1st. Today on the show, the White House has lifted export restrictions on Anthropics Fable 5. We'll find out how the administration came to that decision. We'll dive into Anthropics' quiet takeover of workplace apps, which is now expanding into Microsoft Teams. We're then taking a closer look at Elon Musk's inner circle at Tesla and what it tells us about a possible merger down the road with SpaceX. And to close out the show, we've got Bradley Tusk coming on to talk about the current state of AI regulation.

0:49It's going to be a great show, so let's get right on into it. Fable is back. The White House has eased its export restrictions on Anthropic after its latest model was effectively put in purgatory after a brief release. For more on that story, I want to bring on our AI and politics reporter, Leo Schwartz. Leo, welcome back to the show. It's great to have you here. Good to be here. Fable's back. Fable's back. It was, what was it, two weeks? Week and a half? Just over, just over two weeks. Okay, just over two weeks. What changed? Why did the government give? Did anything change? I think that's a good question.

1:28I don't know how much we actually learned over these very wild last two weeks. After a lot of reporting last night, Anthropic finally put out a lengthy blog, basically explaining what did change. And if you remember, this all stemmed because of an Amazon report where the company found a vulnerability or a jailbreak in Fable that basically surfaced different security risks the model could do. Anthropic, when this was surfaced by the White House, said this is present in every model. It's not really an issue. It doesn't surface the mythos capabilities that are inherent in our model. The White House obviously disagreed, put on these export controls.

2:07And what we can glean from the blog that Anthropic put out yesterday is really not that much changed. Anthropic did put on what's called a classifier that would basically block those type of requests that were surfaced in the Amazon report. But as it said over and over again in the report, we still don't really think this was a big deal, but we're fine putting on this extra layer protection, which all that might mean for users is a bit more annoyance when they try and do basic debugging or other tasks and they get rerouted to a different model. As long as Andy Jassy is happy. I mean, you know, he if you if you recall, we reported that it was it was really it was Amazon, as you said, that that kind of started this whole, you know, journey.

2:52Talk about this classifier thing. I mean, how hard a concession was that for Anthropic to give? it seems like it was a pretty minor change uh what anthropic said in the blog post is this will just it adds to the defense system that we already had it will just basically block even more requests and the way they framed it is you'll probably notice this for more basic tasks that you're trying to do but for the most part we we don't really think it was an issue in the first place and what's interesting is they said in this blog post that they'd actually already worked with casey which is an ai safety institute under the commerce department before this to work out this type of defense system casey had already approved it and they approved it now so if you read between the lines what anthropic is basically saying is not much has changed but we did this fix to to appease the government and sorry casey what what is that casey is a small center within the commerce department it's the the center for ai standards and innovation it's staffed with technical experts who can basically look at different standards within models and make sure that they're safe.

3:58It was started under the Biden administration. It was renamed under Trump to take safety out of the name and significantly lowered in its amount of staff. But it worked closely with Anthropic in this last two weeks of negotiation to basically make sure that everything was above board. Now, in the background here, we obviously have OpenAI, who is working with the government as it seeks to release its own model. How much of a fire do you think that lit under Anthropic to maybe develop this classifier system quickly? Do you think there was a little bit of a rivalry there in that, well, if OpenAI gets their model out there faster, then adoption could suffer?

4:41What's interesting is that the roles are almost reversed now. So Fable is out there and OpenAI's 5.6 is now in this limited release. So Anthropic does have a bit of a lead there. But I don't know if it necessarily changed the situation from Anthropic's standpoint. I think when the government implemented these export controls, Anthropic was very incentivized to work on a solution. Obviously, they knew that OpenAI had been talking about releasing a model with similar capabilities. It does seem like, though, that the government, understanding that OpenAI was right around the corner, did push it to move a little faster with Anthropic and make sure that there was a bit of a level playing field.

5:22Now, you wrote last night that industry figures hope a forthcoming executive order will replace ad hoc decisions with a predictable evaluation process. How likely is that? I think this is the magic bullet that every company is hoping is going to come right around the corner in August. If you remember, we reported about the AI framework that's getting implemented by the early June executive order that gave 60 days for this framework to be created. As we've reported, there hasn't really been that much progress on it. There's only been one session that the government has held with industry to talk about what that might look like, although we know that there's been some smaller communications and back channeling with companies.

6:06So we're now just a month away from when that's supposed to be implemented. And theoretically, once this is there, this will be a framework to make sure that companies know what has to be done for a model to be released. Although, if you remember, it is technically voluntary. I don't know how voluntary it actually is. whether this will be a catch-all to make sure that this type of situation never happens again really remains to be seen, especially as models get more and more advanced in a way that I think is a little faster than the government can keep up with. Right. In other words, the answer to the question, well, what happens to future anthropic model releases, that really depends on what the framework that comes out probably in about a month actually says because there's no way of knowing without knowing that framework then whether or not models will get stuck in this limbo or purgatory phase upon release.

7:02Yeah, so the framework will, you know, have certain benchmarks that if models reach them, then they'll be covered and they have to submit the models to the government for some period of time before they launch. But what's also interesting from this blog that Anthropic put out is they're developing a separate framework. And this is a framework for assessing with private industry, how to look at security vulnerabilities and models and how to approach them. So I think what we're seeing is just a lot of different types of solutions being put forth, some by the government, some more of a self-certified system that private industry will probably do.

7:40And the looming question over all of these is, will they be able to keep up with the very fast moving field of frontier models? Right. Great. Leo, well, I want to thank you for coming on. That is Leo Schwartz, our AI and politics reporter here at The Information. We will have more coverage on The Fable Story towards the end of the episode. We've got Bradley Tusk coming on. He is an expert on all things regulation. But in the meantime, The Information has more exclusive reporting that I want to get through first. Anthropic is prepping an agent for Microsoft Teams, which could be in competition with Microsoft's own co-pilot AI tool.

8:17I want to bring on our Microsoft reporter, Aaron Holmes. Aaron, welcome back to the show. It's great to have you here. It's a pleasure as always. Tell me about this new Anthropic agent that's gonna come about in Teams. Yeah, so essentially this is a product that Anthropic, you know, originally debuted for Slack last week and has recently told Microsoft, according to my reporting, that it's going to release a similar agent for Teams. And basically the idea is that you can now use Teams to talk to Claude and tag it in your Microsoft Teams channels and interact with it in the way that you would interact with a coworker.

8:55And, you know, as you noted, it's interesting because Microsoft has its own co-pilot feature, which directly competes with Claude and which they would love more of their customers to use instead of Claude. But, you know, they're taking a conciliatory approach. And so far they are, you know, open to Anthropic or other AI agent makers having these plugins directly in Teams in front of their customers. And who would bear the cost then of the compute and, you know, what's needed under the Anthropic model? Somebody has to be paying for those tokens, right? Yeah, so that would just be the customer themselves.

9:33Basically, if you already have a Claude account, you would just connect it to your team's account and keep paying for both as you usually would and, you know, pay based on how much you use Claude to Anthropic. Got it. Why is Microsoft allowing this? I mean, it has Copilot. What's the strategy here? Yeah, so this is basically, you know, emblematic of this platform company approach, as Microsoft calls it. The idea is, you know, they could, in theory, kind of wall off their products and make it a place where you can only use Copilot. But their bet is essentially that, you know, if customers already want to use a tool like Claude, they might as well be using it inside a Microsoft app like Teams.

10:14And I think the strategy there is essentially about sort of owning user behavior. And, you know, if Teams becomes the place that you're talking to all of your AI agents, which is kind of how Microsoft CEO Satya Nadella has pitched this vision for the product, that probably means that, you know, you as a company are more likely to keep renewing your team's license if your employees are using it that way. Obviously, there are trade-offs to that since Anthropic is a competitor and in many ways is looking to shave away customers from Microsoft. But that's sort of the calculus on Microsoft's end. Now, we had Laura Bratton on the show yesterday, and we talked about this exact same story almost to a T that was playing out with Slack and Claude Tag, which is the Anthropic agent that has been installed in Slack.

11:04and she was talking about Salesforce employees that were a little bit confused as to why Salesforce was promoting that tool. I mean, you've sort of outlined the same approach here. Do you get the sense that Microsoft employees are confused in the same way or is everyone pretty bought into this vision of, you know, we're selling teams, we want teams to be the best product, we're on board with whatever it takes? You know, I think there's definitely a healthy debate going on inside Microsoft as well. I think it was even more pronounced earlier this year when Anthropic came out with these new features that plug into Word and Excel and PowerPoint and let Claude automate work there because that's even more of a direct competitor to what, you know, Copilot has been pitched to customers as something that it can do.

11:51But at the same time, I think that there's, you know, an acknowledgement inside Microsoft that right now, and especially earlier this year, you know, Claude was sort of the front runner in the enterprise AI market. And rather than try to just convince all of their customers to not use Cloud at all, I think Microsoft is happy to kind of use Cloud to further entrench its software. Meaning, you know, if you're going to be using Cloud to automate sending messages or automate, you know, creating financial models in Excel, at least you're still using Teams and Excel in some form. And I think for now, Microsoft is happy to, you know, maintain that sort of stature within customers' environments.

12:31Well, and the other thing I'm thinking about is if I'm a top AI researcher, and maybe let's talk about the tier below the top AI researchers, because, you know, the best of the best are getting wooed with all these hundreds of million dollars pay packages. But, you know, if I'm on a research team, right, and I care about doing the work that is going to really impact the end customer, I mean, I sort of wonder if the more that Microsoft relies on Anthropic, right, and I have the choice of working in Anthropic or Microsoft, all things equal. I mean, this is kind of more of a reason to go work for Anthropic, right?

13:06Because, I mean, Microsoft basically saying that, hey, we've sort of acknowledged that, you know, our models are not going to be as great. You know, we think Anthropic can do the job anyway. I mean, I don't know. I just wonder if this has implications from like a talent perspective. If Microsoft is basically saying, well, you know, if you work at Anthropic, we'll get to you anyway. Yeah, I don't think Microsoft has been able to compete in the talent market for AI researchers with Anthropic for years, if ever, both because of how much Anthropic is willing to pay researchers and just, you know, its stature is basically the leading AI lab right now.

13:45But it is definitely interesting. And I think you're right that Microsoft has made this sort of pitch that it's not going to train the frontier cutting edge models. It's going to focus more on these smaller models that could be cheaper. but according to Microsoft, in some ways, more useful to customers. And they're really leaning into that pitch heavily at the moment. So I think it'll be interesting to see if customers agree with it, as well as, to your point, the talent that they need to train these models. Right. Okay, so two more questions here. Does OpenAI have an equivalent here to Anthropic integrating into Teams, given that OpenAI's models power a lot of Microsoft's own products, right?

14:27Yeah, they currently do not have any sort of connector for Teams and chat GPT. I'm not sure if that's something that is coming, but I think, you know, OpenAI, you know, is basically happy to sort of corner the chatbot market for enterprises as much as they can. But yeah, unsure if that's coming yet. And then, you know, you've reported extensively on the corporate data wars story and companies having access to data sitting within Slack, Teams, all these types of interfaces. Does this tell us anything about maybe the corporate data wars easing at all? You know, it's interesting. I mean, Microsoft has basically said that they will let any company build a connector to their apps as long as they, you know, comport with Microsoft's policies.

15:18At the same time, we have seen them block some features. For example, Databricks built a feature that would have let people connect agents to Microsoft's, you know, Power BI software, which competes directly with Databricks. And Microsoft actually blocked that feature, as my colleague Kevin reported. But, you know, Microsoft said, oh, that wasn't a competitive move. That was just due to, you know, our worries about the accuracy and reliability of this feature. So, you know, I think it is so far a little bit ad hoc, and it'll probably depend, you know, exactly what Microsoft wants its customers to be doing in a given piece of software.

15:56But no, I would say that the data wars are far from over. Great. Well, Aaron, I want to thank you for coming on. That is Aaron Holmes, our Microsoft reporter, here at The Information. Tesla and SpaceX are becoming increasingly intertwined as the two companies collaborate on TerraFab, the giant$55 billion chip manufacturing project that Elon Musk has embarked on. The information just updated its Tesla org chart to reflect the senior most executives at the company. Org charts of the information are accessible to our pro subscribers. members, I want to bring on Grace Kay, our Elon Musk reporter, who updated the org chart, wrote the story on it, and I want her to walk us through what she is paying attention to.

16:38Grace, welcome back. Welcome to the show, I should say, first time. It's great to have you here. Yeah, thank you for having me. Okay, so remind us here before we get into the org chart, I mean, what is the ambition here with TerraFab and where do things stand right now? Yeah, I mean, I would say it's one of Elon Musk's most ambitious projects. So it's a semiconductor manufacturing facility spread across three of his most important companies. It's a collaboration between SpaceX, Tesla, and XAI. And basically, he is trying to bring this vertically integrated approach and create these chips that could power full self-driving, Optimus, and data centers.

17:17So basically spanning across his companies. But right now, it's very early in the process. So he's still bringing on talent across the chip industry, bringing them in from Tesla and SpaceX to work together and getting permits. Right. So they haven't actually made any. They're still sort of in the planning phase. Nothing's really been constructed yet. Very early. He had just announced it in March, but it's something he's really kind of hyped up on Twitter. Something he's, you know, building up in the company right now. Okay. And both Tesla and SpaceX have both contributed a significant amount of funding to this effort?

17:55Yes. So they both jointly agreed to fund$55 billion initially. And that can grow over time, I think, to about$101 billion over the course of several years. So it's a significant investment. Okay. So now let's get to the org chart that you published. you documented the, I think it was 70 people approximately at Tesla. Across the whole company, one of the themes that came of the org chart, which you pointed out in the story, is that a number of these people are working on the TerraFab effort. What did we learn about who a Tesla is directly involved in that program? Yeah, so we learned that there's a lot of SpaceX people in Tesla working on this endeavor.

18:39So John Fetterspiel, who was a senior level product engineering person on Starlink at SpaceX, is the leader of the project and is reporting directly to Elon Musk. They've brought in a number of people from Starship, Starlink, Dragon engineering and managers to work on the project. And they are also listed internally at Tesla. They've also been recruiting a bunch of people from Intel. We found four key hires over the past few months, including Gary Yang, who is like a 17-year veteran at Intel and is considered a pretty significant hire for the team. And when you say bringing them in from SpaceX, I mean, I just want to make sure I understand this.

19:22Is Tesla, is this Tesla hiring people from SpaceX, or is this just people from SpaceX that are sort of like on secondment over at Tesla or something like that. Yeah, it's pretty confusing because it's kind of a unique deal. And it's a very Elon Musk thing. So they're working at both companies. So they're brought in as SpaceX partners is what they're identified in in the org chart. They have their own roles at SpaceX, but then they're also in internal Tesla systems because they're working together. So they're still employed by SpaceX, but they're also working with Tesla. In other words, I think their LinkedIn like literally has to say, I'm just working at the Musk Corporation because there's really no other way to describe it.

20:07So you talked about some of the people here, which are interesting. I wondered, as you looked at the Tesla org chart as a whole, other than the number of people who are intimately involved in the TerraFab effort, were there other themes that you drew from the org chart or other questions that you came up with in terms of how the company is structured? Yeah, there's a few interesting things. So one thing I noticed is this year, only one of Elon Musk's direct reports has left. And that's quite a contrast to last year, where he lost several direct reports. And generally, he's known for having a lot of turnover.

20:44So he lost Raj, who was leading Tesla sales and AI efforts at the company. And Raj's responsibilities were spread across three new direct reports to Elon Musk. And they're also kind of emphasizing this focus on AI in the company. So there's multiple new roles around like bringing AI to the factory, AI infrastructure at the company. And then, of course, there's, you know, the humanoid robotics program under Konstantinos, Laskaris, and Ashok's autopilot team. So there's a lot of different people reporting to Elon Musk working specifically on AI. And last question for you, Grace. I mean, what did this tell us about a possible merger between Tesla and SpaceX?

21:32Does it make it more likely based on the way the organization is structured? It's hard to say. It's definitely a rumor inside the company and something that investors are talking a lot about. It's something I'm very curious about. I do think it shows us that they are collaborating more than ever and getting closer and closer. Whether that means there's going to be a merger or not, I mean, it's too soon to say. Great. Well, Grace, I want to thank you for coming on. That is Grace Kay, our Elon Musk reporter here at The Information. Our next guest is an expert on all things regulation, and more importantly, in an era where tech policy is just about as dynamic as it has ever been.

22:10Bradley Tusk, founder and CEO of Tusk Ventures, is one of few people that can help us make sense of it. Bradley, welcome to the show. It's great to have you back. Yeah, thanks for having me back. Okay, so Fable is back. It was just over two weeks? we're back in business what uh what were your reflections on on this whole saga you know i mean

22:29Bradley Tusk:i think that the broader lesson to take from all of this is this is all a white space right and if you think about it it's not totally unusual in the sense that uber was a white space crypto was a white space autonomous vehicles were the white space by definition the people who are watching this show are involved in exploring and creating technology often that hasn't been done before. Otherwise, why would you do it? And when it comes to regulation, frequently, if it's a new thing, there is no existing framework to work in because if the bureaucrats could have thought of the ideas and built them at the companies, they would do that instead of being bureaucrats.

23:07Bradley Tusk:And so then whether it's AI or anything else, the question is, how do we take this new thing? And AI is obviously evolving faster than almost anything else I've ever seen and come up with the right regulatory framework for it. And I think there are a few different competing factors that it's helpful to understand when you're thinking about AI regulation, whether it's something like Fable or a chatbot, right, or anything else, which is one, as much as everyone like us likes AI, is excited about it, the public does not. It is wildly unpopular. And that is true in red states, in blue states, at the state level, federal, local.

23:47Bradley Tusk:People are very scared about losing their jobs. They're very scared about catastrophic harm caused by frontier models. They're scared about chatbots doing mental health, telling their teenagers to kill themselves. And it's not to say that all their fears are valid, but it doesn't matter if they're valid or not. If they feel that way, then the elected officials who represent them are going to want to reflect those anxieties in law, in policy, in regulation, because if you're a politician, your goal is to win the next election, stay in office. And the way you do that is making your voters happy, right?

24:19Bradley Tusk:So the first thing is just a mind shift, which is to understand the way that we see this and the way that they see this are radically different. And where founders, in my experience, fail the most around regulation is when they think, well, I'm so much smarter than they are, and I know so much better than they do why can't they just listen to me and do what i say and the answer is because they don't care about you right they care about their politics their election their career and if you can't explain what you need in terms that work for them politically it doesn't matter how smart you are it doesn't matter that you went to stanford or you and y carminator or ac16 invested in your seed round like that's all totally irrelevant to a governor state senator or whatever else so that's yeah go ahead what what do you think this tells us about future model releases though i mean we've got this we had the ai executive order we have a 60 day 60 day uh period i guess waiting period or 30 days out maybe from from some kind of a framework do you think that models are going to get stuck in this purgatory phase uh from now on you know yeah probably but the reality is and And maybe I'm not being pro-innovation enough when I say this, but if you are releasing technology that has this level of power and this ability to shape things, to expect there to be no sort of regulatory review is totally unreasonable, right?

25:42Bradley Tusk:If anything, Trump's order was probably way too milquetoast and way too lax, right? Right. It was voluntary. It was like, do what you want. Solitary 30-day review. You know, if you're talking about systems that can hack through any corporation's, you know, infrastructure, if you're talking about systems that in theory could, you know, launch drone strikes without any sort of human, you know, authorization or whatever else, the idea that government would want and need to have a say in all that is not unreasonable at all. So I think that if you're doing a product release, whether you're anthropic or anyone else, you have to build that into account.

26:19Bradley Tusk:And the other thing is, you know, this is not a beg for forgiveness situation, right? This is some combination of ask for permission or at least be proactive. And the more that, you know, there's a quote from Pericles that just because you don't take an interest in politics doesn't mean politics won't take an interest in you. And the more that you put your head in the sand and either think, oh, they won't notice or we'll deal with that problem when it comes or they're so stupid that when they realize how smart I am, they'll just do what I want. That's how you lose. Right. And you need to be proactive and thoughtful around who politically has jurisdiction over this.

26:53Bradley Tusk:What will their concerns be? How do I alleviate them? And the more that you wait to the last minute, the more you're going to be frustrated because then they are going to come in and stop you from your release or whatever else you want to do. So you got to be smart and proactive about this, just like you are with your engineering, your fundraising, your go-to-market and everything else. Right. So if you were to sort of take apart the differing political playbooks of Anthropic and OpenAI, I mean, OpenAI, we just had our AI and politics reporter on the show. OpenAI took a more proactive approach with 5.6, right?

27:26They went to the government. They said, hey, we want to do a phased out release. Anthropic sort of did the same thing, although then they came out with Fable and then, you know, they landed themselves in this whole debacle. but at the end of the day, Fable is out sooner. I mean, the net result is Fable is out and 5.6 is not. So, I mean, unpack the differing playbooks here of these two companies. Which one do you think is doing it better? What critiques do you have for both sides? What do you think?