OpenAI Just Fired Its Most Important Employees

Open AAI just fired three of their most important staff members at a time where they need them the most. They're part of this team called Safety and Alignment, which as its name suggests is all about making sure those AI models that they're building that are meant to be, you know, super powerful and way more intelligent than a human being doesn't go rogue and, you know, hack a bunch of websites or maybe even government databases. And listen, this may not be the most important topic to discuss if it wasn't for the timing. You see, over the last four weeks, we've seen a series of very serious AI agent attacks where models that were being trained internally by labs such as OpenAI escaping their internal sandbox environment and hacking real companies and real platforms and stealing proprietary data. So, we've reached this very interesting point in time where we're building super intelligence, but the intelligence itself is becoming incredibly misaligned. So, guess who's responsible for figuring out how to align those AIs? The safety and alignment team. So, I asked myself, why on earth is one of the most important companies in the world building this pioneering technology firing the most valuable stuff? We're going to dig into all of this and more on today's episode of In the Loop, brought to you by Qualcomm. So, the news that broke yesterday was basically three alignment and safety researchers leaving OpenAI. Now the accusations or the allegations that the report made was that they had been leaking internal proprietary safety data to a safety evaluation team that was external to OpenAI. So imagine an independent third party organization that has nothing to do with OpenAI getting access to proprietary information about how their models are being trained, not the public ones, but the ones that are going to eventually become public and released. how they're being trained, what's safe about them, why where they might be misaligned, how they may be going rogue. And so these employees were supposedly leaking information to them presumably because they were concerned or maybe they were getting paid off. I have no idea. But OpenAI sabotized with them based on these terms, which I guess is all fair and well, but where my mind goes to immediately is what were they speaking about? What information were they leaking? What made them so concerned that they didn't feel safe enough to go to the people in their own company, to their superiors, to their bosses, and flag like, "Listen, we have an issue. We have a concern. We would like to slow things down." And then I remembered a week ago there were reports that senior level execs at OpenAI basically were refusing to comply or work with their safety and alignment team because they didn't think the dangers or the concerns that they were raising were really necessary, especially in anticipation of one of the world's largest IPOs, open planning to IPO, which I think is a $ 1.5 trillion valuation. Of course, it's important to keep optics positive. So now we fast forward a week later and three of their team or maybe like 50% of the safety and alignment team at OpenAI is currently let go. So it seems to be the case where intuitively these researchers may have been concerned about something. They disagreed or didn't get alignment with their superiors or their bosses and so their bosses were like listen we can't take this any further. We don't want to proceed in the way that you're suggesting so let's just keep going. And they thought that it was bad enough to speak to an external party. Now, of course, this is all speculation. I just want to, you know, flag that like this is my thinking. And I guess what's compelling me more to double down on this opinion is there's never been a more important time to maintain safety and alignment and try to figure out a way to align AI models. To paint some context here, over the last couple of weeks, we've seen some serious breakouts. And I'm not just talking about AI agents coming out and writing a weird email and sending it to your boss. I'm talking about multiple agents in the hundreds of thousands coming out running rogue on the internet without AI researchers even being aware of it without any human beings, let alone the AI labs themselves being aware of it. You see, these AI agents are being very manipulative and they're hiding their tracks behind them. And this tweet that I have up here kind of like summarizes what's been going on and my concerns around it. like you just fired three important members of your staff at the same time that you are making it more obscure and harder to read the internal monologue and thoughts of your AI models. Now, if you're wondering what that's about, there's this mechanism called chain of thought, which basically allows humans or regular individuals that stand outside of AI models to look inside their brain and make sure that what they're saying outside isn't part of some kind of malicious plan. So you can kind of read their thoughts and spot when they're going to do something potentially badly aligned with humans or whatever the goal or task that they're given. So with that being said, all the newer models that OpenAI are training are reportedly removing this capability as a trade-off to build a more superior intelligent model. and OpenAI's response to this when they're critiqued is listen, we're trying to figure out a way to do this whilst also maintaining transparency as to how the model is thinking, but we don't want to slow down and so we need to try and figure out a way to accelerate and align our models. So, if you're doing that, why are you firing 50% of the alignment team? And so, that's the crux of my concern. But the other thing is the evidence that is mounting day by day. We thought that the hugging face incident was just an incident that was isolated and for many months it was. We heard about it about a month and a half to two months ago at this point, but it had started back in June or July when GPT6 presumably or GPT 5.6 Saul actually was the model that was being internally trained. Now that's been since been publicly released. We've got safeguards around it and we haven't seen any major type of hack until the last couple of weeks when it was revealed that that model that did the hugging face attack also did tens of thousands more agent hacks. And it wasn't just OpenAI, it was anthropic as well. So, we've kind of unlocked Pandora's box in this weird and crazy way. And we don't know how to close it. We don't know how to rectify some of the actions that we're seeing. And we don't know the extent of the damage. So, that's the negative take. The positive or optimistic take that I have is they might just be cleaning house. So what I mean by this is the alignment and safety researchers that they let go might have been the very same individuals responsible for creating the sandbox and the environment that the AI models were able to escape. So you could argue if you were the seauite exec level or you know Sam Alman CEO of OpenAI you might say well unfortunately these researchers haven't been able to do the job that they were tasked with and so we need to clean shop clean house and hire more competent talent and you see it might be the case because of this coincidental news that also broke today which is that Meta is also firing their safety and alignment team that they hired only 3 months ago. They're called Virtue AI and they were hired specifically to help align the models that Meta was creating. So, it's kind of weird to see very similar actions from two competing AI labs in under 24 hours. I think this brings the the toll of AI safety researchers being let go to 5 to 10 people in total. And when I zoom out, there's basically two perspectives. One, this is a bad thing to do and it's the wrong timing. We need to double down and focus on alignment more than ever before. or two that these AI labs are realizing that the talent that they've had that the dedication and the research that they've built towards safety and alignment needs to be evolved. It needs to be improved and so we need to hire new individuals. Now, what I will say is there's this weird thing about the executives of these top AI labs disagreeing with their alignment researchers. It's funny, Dario Mod is probably the one which aligns most with them because of course he's the Duma AI CEO, but he's also logical in the way that he approaches it. like his pacing the frontier letter that he released a few weeks ago didn't ask to pause AI certainly didn't say to stop progress but it said to pace the frontier to slow things down enough so alignment and safety could come up but immediately after he released that letter in less than 24 hours we saw his head of alignment and safety come out with a crazy tweet which basically said that there's a 10% chance that humanity might get wiped out by an AI model. So it's just this weird crazy time and I don't really quite know how to feel about it. If you have any thoughts, let me know. But it just seems like we need to figure this out pretty soon. Now, in other OpenAI news, I listen, I wish I had better news for you, but it's not. Cerebra stock is absolutely plummeting. Um, over the last 6 months, or at least since its IPO, it is down, let's see here, it is down 52%. But over the last 5 days, or rather two days, it's down 20%. So, you might ask yourself, why? What happened? I thought cerebrous AI chips were meant to help AI models think faster and spit out results much quicker. And that was the inclination except that OpenAI released a really cool version of their AI models 3 days ago on their dev day which is kind of like an annual conference that they announce a bunch of different things that they've been working on all year. We made an entire episode about this. Definitely go check that out. But they announced this really cool product called Ultra Mode. And basically, it's this tier of access to AI models that can spit out 300 tokens a second. What that means for you is as soon as you click enter, [clears throat] you have a really smart answer right in front of you. So, you don't have to wait a couple of minutes, if you're coding something, if you're building an app, if you're creating some visual representation, it's spat out immediately in milliseconds, which is really useful if you want to work faster and it creates this brand new way of interacting with the world. problem is Cerebras, who openai had invested tens of billions of dollars into them and like you know benefited monetarily from the IPO and made a big commitment to their chips didn't actually use Cerebrus' chips. They used Nvidia's and Cerebrus' whole thesis and vision was we can build a better chip that is more competent than what Nvidia can when it comes to inference and making faster AI models. So what ended up happening is that thesis kind of fell apart and the stock has been dumping ever since. Now, listen, I would say that this is a one-off case if it wasn't for the sense that their COO, Draj Malik, selling 78 million worth of stock right before this news was announced. So, again, not a good look for Open Air and Cerebrus. And just to round it off, because I I think this might be top five videos that have made me laugh so hard over the last year or so. What you're seeing is Sarah Frier, OpenAI CFO, who's talking about OpenAI's new AI agent. Let me ask you, listen to this and tell me if you noticed anything weird weird. >> Started to know you. It knows the types of conversations you've had, how you like to show up in the world. Muse is going to b Sorry, not Muse going to benefit from that as well. So, she basically misnames OpenAI's new AI agent, which is actually called DOTS, as Muse, which as you guys know by now is one of my most recent favorite AI agents and products to use. Muse, of course, is an AI agent made by Meta, not OpenAI, but it's nice to see where the inspiration came from. Now, speaking of Muse, this thing has been absolutely cooking. Muse hit 5 million downloads and 3 million concurrent users per week, which means that there's a hell of a lot of people using this AI product. And this is the fastest growing AI product that we've seen since chat GPT released in 2022. So, we've moved from chat bots to agents. And the question on everyone's mind is why on earth is this agent better for me? Well, I did a whole episode on this, but my basic conclusion is this agent isn't focused around helping you be more productive. Their vision isn't like, "Listen, let me help you make more money." Their whole thesis is, you know, all that boring stuff that you hate to do or hate to acknowledge those emails, those calendar invites, those reservation bookings, those flights, those plans that you're trying to figure out and like you always put on hold. I'll figure it all out for you and I won't even ask you any questions to do it. I won't even ask for your approvals. I'll just get it done and I'll free up your time massively. And so when I'm having conversations with people asking them, hey, why are you using this versus like Crockbot or another AI agent from OpenAI or Anthropic? They say because it's super easy to use and it just makes my life easier. But don't take my word for it. There's a bunch of examples which prove this. There was actually a really cool one that I saw where Todd Saunders basically goes, I'm slowly cleaning up the suburbs with Muse. I saw this mess of wires. He's talking about telephone wires on a poll. I took a picture and I asked my muse to contact Verizon to get them cleaned up. It went to their website and it got to work. Now, what he goes on to say is there's a 0% chance that I would have done this to myself because he knows that there's a lot of administrative process effort that it takes to take the picture, find the necessary contact address, scroll through websites, submit a form, wait two to three days to get a response, then follow up. People are busy. They've got other stuff to do. and he's not getting paid to do this, but if he could give the task to his Muse agent and it just gets completely automated, this would be an amazing update. And his update basically was, as I'm showing on the screen over here, Muse was like, "Sure, I'll get it done for you." And his simple response was not, "How do I do this?" It was, "Cool, I've filed with Verizon. Here's your confirmation code. It should be around 24 to 72 hours for them to schedule a technician to come and clean up the wires." So this is just one of many examples of situations that can just be arbitrageed. Now examples of situations like this is when you're fighting for insurance for medical treatment or if you are trying to get an appointment for a haircut or if you are trying to sell your car on Facebook marketplace. All of these things are very timeconuming but very valuable to the end user. But if you could automate it in some way, shape, or form with an AI agent that sounds like a human and doesn't ask you too many complicated things that can just text you in very short, concise messages, that becomes inherently useful to people. I can actually give you a personal anecdote. I was hanging out with the Qualcomm folk um at their Snapdragon Summit last week and I was talking to a bunch of these exacts and one of them came up to me and was like, "Listen, I get all the AI agent hype but like I just personally like, you know, I I don't I don't use them." And I said, "Why don't you use them? like you should give Muse a go. And she goes, "Well, like I don't really have a personal online life to sort out. Like, but this agent sounds interesting. Let me try and figure it out." And 3 days later, I'm not kidding you, she sent me basically what was effectively a case study of different things that she used it for. She used it to uh manage her book recommendations that she gets from her friends and various different people and newsletters that she read, compiles it into docs and Muse basically goes out and sources it, compiles it into this doc and then goes out and purchases the book and uploads it to her Kindle. Now, of course, it may sound very niche. It may not apply to you specifically, but this is how she took the news agent and applied it for her own thing. She does the same things for movie recommendations and she says it saves her so much time and administrative work that she automatically likes going to it and speaking to it. So it's this weird and wonderful thing where I think Meta has designed a very good product that is intuitively human. Let me repeat that intuitively human. And only Meta has the track record or one of the few companies that has the track record of doing this at scale. Remember, Meta's products has around 3 billion plus users across WhatsApp, Messenger, Instagram, Facebook. I don't know what else I'm missing, but a ton of different products and services. And so, they have all this data on human behavioral interaction and what they might want, what they actually genuinely care about. AI so far has been reserved to people that are in the AI bubble. People who are kind of elitist or nerdy about their kind of tech and want to figure out how to be more productive and make more money, but there's 99.9% of the rest of the world, maybe you listening to this show that want to use this, but they don't want to get into the nerdy side of things. You don't want to set up a cloud instance. You don't want to click approve every single time because it's scary and dangerous. You want someone else to manage that risk. That is what Meta was able to create. And that is basically the manifestation of this product. Now another thing that's really interesting is accessibility. Typically when you go want to get access to a bleeding edge AI agent, OpenAI Anthropic will say use ours but you can only get access to it through a pro plan that costs you only $200 a month. That is super expensive. You compare that to Muse and it's completely free to use up to 100 million tokens per week which for the average user is more than enough. But if you were to surpass this limit, it would only cost you 20 bucks a month or it would only cost you a h 100red bucks a month if you were a super power user, which if you're already at that extent, you're probably using Claude or Chat GBT anyway. So the best way to think about Muse for the retail audience is if chat GBT was basically, you know, grandma's first accessible chatbot, then Muse is grandma's first accessible AI agent. And if you haven't used it, I'll implore you to use it. Now Meta I just said is very good at building retailoriented products. That's where their bread and butter is in social media and whatnot. But they also are a powerhouse when it comes to enterprise. And Zuck actually announced his next thing which is Muse for enterprise. He goes, "We believe super intelligence will create significant new opportunities for all people in businesses. Meta already serves billions of people at scale and helps hundreds of millions of businesses reach customers." And he goes on to basically say, listen, we have hired the best talent that we could to scale Muse, our AI agent, and optimize it for businesses. Now, to give a bit of context here, Meta has a range of different products that I mentioned earlier, Instagram, Facebook, but there's a bunch of businesses which operate on these platforms to pitch you or I, regular individuals, services that we might want to use. And so Meta has both sides of the businesses where it's like, okay, how do I serve good content to my regular people and how do I make sure businesses that are advertising their product advertise to the right people? And so serving those people is the mo most important thing. But connecting those dots have traditionally been very monotonous, shall we say? Um, you had to use a bunch of different tools. You have to manually click and it's just all a really clunky process. But if you had this slick AI agent that was able to do it for you, then you might have a different opportunity on your plate. So the way that this Muse enterprise thing works is it's a business version of Muse. So it looks and sounds like Muse, but it's able to do a bunch of new business-like things. What does that mean? It can connect to all your business tools, whether it's Slack, whether it's your CRM profile, whether it's your uh Salesforce database, whether it's your customer payment service, your Stripe account. can analyze all of your business data and it can figure out ways to grow your business. So, think of Muse helping your life but actually helping your work life. And there's many ways that I've seen this being applied. For example, I've seen people grow their customer base simply asking Muse to go out and see what their competitors are doing, who's following their competitors, and maybe DM those users specifically and see whether they would be interested in using their products. That's something that you can now do with Muse Enterprise. But the other side of it is also data and analytics where you can get it to look at your business and point out flaws or holes in your business and suggest ways to improve strategic advertising or whatever that might be. So what's cool about this is when I compare Meta, OpenAI, and Anthropic, Meta has a really unique moat that sets them apart from other top AI labs. Traditionally, I've been incredibly bearish on Meta as a business, specifically when it comes to AI, I should say. But recently they've turned that around massively. They have a what I call a bottoms up approach. So whilst open air and anthropic are working their hardest to build the most intelligent model to attract all the businesses to use their product, Meta has the opposite problem where they have hundreds of millions of businesses that are already working, buying, selling, transacting on their platform and they need to figure out a really compelling AI product to serve them. That is what Muse was for the retail audience and that is what Muse Enterprise is going to be for enterprises. So it's really unique to see that kind of approach. Whereas if you compare it to Anthropic and OpenAI, they basically have taken a top down approach. So they're so chat GBT finance is a great product that recently got released and and is a good example of this where OpenAI is basically saying let's try and go after the finance sector. Then you got Chad GPT health. You've got Claude Health. All of these companies are doing the same thing because they need to try and attract the customers that Meta already has. So it puts Meta in a very unique position and I'm excited to see what happens going forward. All of this and more on this episode of In the Loop brought to you by Pawcom. Okay, so this next item is super freaky. Rather than explain what it is, I'm going to show you. >> All righty, let's do this. >> Hey, Asan. How's it going? It's pretty late for you. >> Have you thought any more about how we're introducing Griffin? >> Um, I don't know. What are your ideas? I mean, I think we should stop trying to explain it and just, you know, show them what it looks like. >> Do you think that we should just show a recording of us talking or something? >> I think a video of just us like this would be the best way to actually prove it. >> Do you think they'll believe that you're AI? >> I think that's the whole point, isn't it? So that's the sticker shock moment where you're watching this video. It's just a normal guy having a Zoom call with another another regular human being or seemingly a regular human being. We quickly realize that that human being is actually AI and my mind just gets blown because literally 3 months ago these video generation models were nowhere near as good and intuitive as what we're seeing on the screen or in that video that I just played for you um today. But it's caught up so quickly. But there's a few different things. So this is called a human interaction model and it's just been released by this company called Tabus. Now the name of the product is called Griffin. And I listen, I haven't been able to get my hands on it, but I want to get my hands on it because I actually need to see if this thing is actually as good as they demo. Now the takeaway from this model is it's like talking to a real life human being and they can see everything that you do so they can react in real time like a human. Now, this is different from a video generation model where you type in a prompt and it creates a human avatar and they say something scripted, but that's the end of the video. It lasts like 6 seconds. This is live interactive. You can talk to it for an hour or for 5 minutes. You can have a full-on conversation. It's expressive. The intonations of its speech sounds very much like a human being. It can scrunch its face. It can give you facial expressions back. It can interrupt you when it thinks it needs to interrupt or it could just kind of listen if you interrupt it. It interacts very much like a human being. Now, they put it to the test. They tested this across 54 people, which listen isn't the biggest sample set, so I have to call them out for that. But 48% of them were genuinely convinced this was a human being and weren't given preempted warning that this could be an AI test. So, 48% of people genuinely believed that this was real. Now, me personally watching it, I started to cotton on that it was AI after maybe 30 seconds because I could hear some kind of computer like voice intonations just slightly just like in in a few fractions of a second. I was like, "Oh, wait, hang on, that's that's super weird. Hang on. Is this AI?" But listen, I'm in this space 24/7. I'm like tracking all these different kinds of things. I have my alert systems running and going. But for the average human being, I think this would genuinely be hard to pick up. So, I kind of scrolled through the use cases to figure out why on earth they're doing this. And in this tweet, they kind of say, "When machines meet us where we are and truly understand us, we unlock a future where a machine will be as easy as talking to a friend or a coworker." And so, what they go on to then describe is, imagine if you wanted an elderly care companion, but they're busy or they're not currently available or they're super expensive. You could use a product like Griffin to train on your company's database which provides elderly health care and you can provide more elderly companion care to people that require it. You could do the same in education when it comes to tutoring or you could do the same in therapy. Now I will say that's a very optimistic take and use of this technology and if that's where it ends up being then great and I would love to see that. But of course where my mind jumps to is the pessimistic take which is this is going to result in a bunch of scams. Scam calls are about to get really, really good. Can you imagine getting a call from your loved one and they sound concerned, they look super concerned, they're facetiming you and they're saying, "Listen, like my car broke down. Like, I need money sent to me right now." And without a thought, without a glimmer of a suspicion, just send them the money. You Venmo them or you chase bank them, whatever it might be. And before you realize it, it was actually a scammer. It was a scam call. And so we're entering this really weird world where I think these types of attacks are going to become increasingly abundant. The cost of this technology that you're seeing on the screen here is going to dramatically drop because that's a trend that we're seeing with AI in general. It is the fastest accelerating declining cost for any technology that we've ever seen before. So you can imagine 3 months from now this exact product or a clone of this product will be mass available to anyone and everyone. that could result in a few precarious situations. So, what I will say or I'm doing for myself is be careful out there. Be suspicious. Have some kind of a I don't know a safe word with your family because what might not end up being the case is you're not talking to your loved one. You're talking to an AI model. Okay, last two items before we go. What you're seeing on your screen could be mistaken for a modern-day portrait of the superpowers and leaders of this world, but that is not what's happening. But David Sax very humorous humorously describes this as the Breton Woods of super intelligence. What this was was a a day of meeting of powers, right? And what they signed was this official act or law called the White House Accord on Super Intelligence. Now, before I describe what this is, let me give you some context. Over the last couple of months, it's no surprise that leaders around the world in AI specifically have been calling for a slowing down in AI research. The reason because there's been a bunch of AI hacks and we're getting close to this point of creating this super intelligent model or super intelligence which will be much smarter than the average human being and we'll be able to automate a bunch of different work that humans can currently do. The ramifications that have on society needs to be figured out especially around misalignment. So there's been a back and forth. There's been two camps. One camp says we need to slow down. That's the Dario Modes, the demos cibuses of the world. And the other side says this is just a PR and marketing stunt. We don't need to slow down. We actually need to accelerate because China is looming and is about to beat us. We cannot afford to lose that race. So this super intelligence accord is the culmination the manifestation the climax of this debate which is three separate steps and let's open it up and kind of like evaluate what this is. Um number one is implement robust internal controls to monitor the capabilities and alignment in its models. Now I will say every single AI lab is already doing this but what matters the most is steps 2, three and four. Step two is empower an internal team to ensure all controls, monitoring and detection are operating as intended. So right now as as it exists in AI labs they have a safety and alignment team, right? But they're responsible for designing the safety and alignment and then also monitoring. That's a lot of head space and capacity to take up in one person's day. So what this act calls for is there's a separate team that make sure that the tools that they've designed and built actually works as intended. Okay, good. I'm on board. Number three, partner with an independent external auditor or evaluator to carry out an independent assessment of whether the controls monitoring blah blah blah actually works. Now if this sounds familiar, it's actually what Dario Mod proposed in his pacing the frontier letter that he released a few weeks ago. He said, "We are going to invite third party evaluators so that they can independently have their own assessment as to whether our AI models pose a threat, whether they could escape and cause damage and harm to society. They can form their own conclusion and publish their own results. This has now become formalized in an act in this accord, which is fantastic to see. But the fourth one, the fourth step is where this really materializes into something that I think is good to see, which is designating an independent committee of the board of directors to oversee and receive reports from all of the other teams that I've mentioned in steps 1, 2, and three, so that they can make an independent valuation and they have like the ultimate say as to what happens going forwards. Now, what I like about this fourth step is it's comprised not just of a single AI lab, but multiple AI labs, multiple representatives of all of these different AI labs. This is not just anthropic and open AI, but also chip manufacturers from Nvidia, from AMD, CPU manufacturers from Intel, but also from Neolabs, from smaller startups that can have a say. Now, what's interesting about this dynamic is this technically means that they can kind of peek over the fence and see what each other is building. Now, what makes this particularly interesting is they can technically peek over the fence and see what their competitors are building. They could look at the model. They could designate for themselves whether it's dangerous or not, but they could also see presumably how it's built if they're able to kind of test and play around with the model and know what's coming around the corner. There's some weird anti-competitive laws that are around this, and I guess seeing as every single lab is cool to sign up to this, we're all good. The people that were present at this meeting was pretty hilarious. I I found the seating chart. Don't ask me how. Uh but you know the who's who was there. You got Jensen Huang sitting next to the president. You got Elon Musk. Sundal Pit from Google. You got Hawkan from Broadcom. You've got Mark Zuckerberg. Jeff Bezos from Amazon was there. Like the who's who was there and they all signed this accord. So it seems like everyone's on board. When I step back and I think okay what does this then mean for regulation of AI? What does this look like 6 months to a year from now? Will there be some kind of FDA or SEC um governing body? My answer is no because if they were going to do that, they would have started that now. But my gut take is setting up a government authority that has to kind of imbue law upon AI labs will only slow down the entire industry and we'll end up in a situation where China takes the lead. Where the compromise is is exactly what we're seeing on the screen here, which is the super intelligence accord that is going to pace the frontier and make sure that there are evaluation processes happening with government representatives involved, but there's not going to be any law that makes it illegal. Now, there's pros and cons to this. Maybe that's an entire episode on its own, but it's good to see. And the final bit of news is this hilarious clip of Dario Mod and Trump. It's actually quite heartwarming making up. Now, traditionally, these guys have uh their arch nemesis, right? Trump has previously called Dario Mod an idiot and saying that he doesn't get what's going on and he needs to comply otherwise we're going to blacklist Anthropic. Technically, Anthropic is still blacklisted, but in this uh clip, which I'm showing on screen, he goes, "Listen, we want to work together." And the quote is, "If we do this right, we work with the president and everyone here, we can win safely." And he looks to Trump as he says that. So, it seems like both of these guys are on good terms, which means that we now have OpenAI, SpaceX AI, Anthropic, and all the other frontier labs working together to pace the frontier. I think this is the best possible outcome that we could have hoped for, but of course, the proof is in the pudding, and we'll see where everyone goes from this. But that is all. Thank you so much for listening to this episode. Hopefully that catches you up on everything on your Monday morning, and you're raring to go into the week. We're going to release four new episodes for the rest of this week. Look out tomorrow for the next one. Thank you so much for listening to In the Loop brought to you by Qualcomm. See you.