Meta Debuts New Agent ‘Muse,’ OpenAI’s Math Drama & White House AI Whitelist Confusion

0:13Welcome everyone to the information's TI TV. My name is Akash Pasricha. It is Wednesday, September 9th. Today on the show, we are talking all about Meta's new Muse Agent. We'll get some early reactions to it. A lot of people are impressed. We'll then talk about the math breakthrough that OpenAI says one of its models has made and why that story has also been the center of some debate. We'll also discuss our latest reporting on the administration's AI regulation. There apparently is a whitelist with partners in the works. Problem is no one knows who is on it, including the partners themselves. And to close out the show, we are going to talk about Hugging Faces robotics play and what NVIDIA plans to do with that.

0:59It's going to be a great show, so let's get right on into it. Meta launched its first consumer AI agent on Tuesday called Muse. The information has been reporting on the product internally named Hatch for quite some time now. Mark Zuckerberg said the agent will, quote, understand your goals and work 24-7 to get things done for you, end quote. They're both free and paid tiers. I want to bring us someone who has been playing with it. Jaya Gupta is a partner at Foundation Capital. Jaya, welcome to the show. It's great to have you back. Thanks for having me. So I'm going to get to your reactions in a second.

1:37What is meta advertising MUSE is going to be able to do for you? Good question. I think Mark Zuckerberg and Alex Wang are coming out and saying, hey this is going to be able to do and take over your you know your tasks end to end and sort of be your personal assistant that will be able to handle admin that'll uh be able to handle like you know when you're taking your kid to soccer practice and you forget about trials like it's going to be it's going to be able to handle all the things that you know people forget to think about or it'll be it'll be handling all the things you know in life that people are charging you for money for things.

2:16Like, how do I get you money back? How do I help you shop more? And so I think it's the whole, it's the whole personal assistant mandate, essentially things that you didn't want to do in your life. Exactly. Okay. So now you've been using it. Does it live up to that hype? You know, it only came out yesterday. So, you know, we don't have a lot of data here, but what do you think? Yeah. I mean, it is, it's, it's insane. Like I would say that, I obviously had instinct first. And so, you know, that got me, you know, thinking like, oh, my gosh, this thing can it found me an apartment like it was able to negotiate my apartment with me.

2:53It was able to, you know, my building didn't have parking, we found out. So it was able to go run like an RFP or like a bake off between all the buildings and SF for like parking. And so I think you're talking about instinct right now or about Muse? That was for Muse. That was for Muse. OK, that's for Muse. Wow. That was for Muse. Yeah. And the amount of things it's been able to do has been pretty crazy. And the way you can integrate all your other apps and your emails super quickly, now you have to kind of figure out who you trust your data with. But I think that I also really like the fact how much they thought about security and privacy, all those things on the Muse side.

3:30So I'm super impressed already. So can you just maybe walk us through what is the user experience like? I mean, you download this onto your laptop, onto your phone. You have to like, do you, you have to click through and say, hey, you have access to these files or these apps and not these? Like what is the whole walkthrough? Yeah, good question. Oh my gosh, I wish I could give you a demo in real time. But basically you, from Muse, you click sign up, you get to like an app interface and you, you know, there's a sort of button where it's like you can connect things over time and you can connect like you can it's like a one click connect and you choose what to connect it to.

4:13I think for me, it was very, very fast, of course, to connect anything to like Instagram or like Facebook, anything that they already had. Super, super simple, of course. And then, you know, you can kind of connect things like your Gmail. And so then it's like sort of a chat interface like, hey uh you know it'll it'll say it's the same it says the same thing that instinct said hey give me one thing to take off your task list and i'm like hey you know i need uh help with finding an apartment um it'll go and and then it'll sort of ping you every few hours like saying hey i found this i found this i found this there's updated pricing here and so it's pretty it's pretty impressive um and then it kind of you know puts things into like little like briefs for the day like it's like when i wake up i need to know a few things it'll come and it does it does it like um how does it send these briefs to you is it in the app or is it by chat like a i message or how does it work good question so it's in the app and then also they funnel it to whatsapp so you can connect really really quickly to whatsapp makes sense meta yeah and so i get i get them both in whatsapp as plus the app because i have whatsapp on my computer and my phone and you know i want things on on both but yes you can do app or whatsapp and then so it syncs between what's you know what's happening on your phone what's happening on the computer you can sort of access it through through both exactly okay okay so so sounds pretty impressive sounds like you've had a great experience with it and i'd love to know at some point how much you managed to save on the apartment uh what that negotiation looked like but uh so where does it fall short i mean are there things that it cannot do right now?

5:52Yeah, good question. I mean, I think right now it's hard to connect it to work things. So, you know, that is something that I think Instinct, for example, has been able to get around. Meaning your work account, like what kind of work are you talking about? Like PowerPoints or? Yeah, like work emails and connecting to Outlook and connecting to like more of the, you know, foundation capital side of the house. Like, I think it's, you know, Muse has been, the guardrails are pretty tight on what you can sort of connect with on the work side. And it's also pretty tight on things that you can, you know, like say to it.

6:32So it'll catch things, you know, I think they've done a great job red teaming it because there was definitely some things that I was able to do on other apps that I can't do on this one. And when you say there are guardrails around it, does it mean that it literally will not let you connect to your enterprise Gmail account, for example, but you can connect to your personal Gmail account? Is Meta saying we do not want this to access any work systems? Is that what's going on? Yeah. I mean, look, I only tried it for a few seconds, but I was not able to connect to Outlook, and it was like, you know, we don't want you to connect to work systems.

7:09And I was like, okay, this is, you know, thank you for thinking about this in advance. But instinct, well, you can do whatever you want. You can do more of whatever you want. Yes. You can definitely do more of whatever you want. So, okay. So this is kind of an interesting and intentional choice from Meta. I mean, why do you think they're making that decision? and what's your analysis been around the safety of not just these two? Well, I mean, yeah, let's talk about the safety of these two agents specifically because, you know, we've done some reporting here at The Information about internal testing that Meta did before the release, and there were some episodes where the agent took unwanted actions, and, of course, that was in a test case.

7:58But, I mean, this is the question, right, is how safe is it? So what's your review on the safety of these two popular agents? Yeah, good question. I mean, I think I went viral for posting about my Instinct incident, which was, you know, I told it to go, hey, you know, I needed something from my UCSF account. The password and password manager was wrong. So Instinct went and literally changed my password. It hit forgot password, changed my password, and then didn't even tell me about it. And then, you know, then it tells me that, hey, actually, I changed your Stanford password. And, you know, now I'm kind of locked out of both my UCSF and Stanford accounts.

8:37But either way, like, I think you haven't you haven't fixed it yet. I haven't. But, you know, I need to call them because I, you know, the thing where it's probably a personal problem where I just can't figure out what password they use. And I should forget password. But, yes, I have not fixed that yet. um i think on the meta side what i've observed on the use side is that like that situation doesn't really happen and i know in early testing and all those things like they probably caught all these hiccups where i think instinct is catching the hiccups as it's taking off and meta probably of course bigger distribution force bigger time to think about these things they've hired privacy people i think they hired the x signal person too um so i think that they've just thought a lot of these things through.

9:19Also, Meta kind of has to. Meta not had a great reputation on that. So between the two, you do feel like the Meta product, Muse, is the safer agent to use right now? I think so. And I think that's just also because Meta's told us what their security architecture is. Each user kind of gets their own isolated secure VM. There's the Sentinel that reviews actions. They've optimized for, you know, the users like guardrails and all these things. There's an explicit like, this is not for ads. And so they've said all these things. They've posted the technical architecture diagram. I don't have that for instinct.

10:00So in some ways, like, one, like, I feel more comfortable with the fact that they kind of said, like, here's all the things that we've done. And two, like, meta. How does Muse compare to some of these other, you know, computer use agents. I know Perplexity has been making some traction there. And then, I mean, OpenAI, Anthropic, I think they have some sort of similar product. Just help me understand where Muse fits in that competitive landscape. Yeah, great question. So I think what, and I can bucket these in a few different ways, but there's GrokBot, which is Elon's product, and the cursor team's new product.

10:41And then you have like, of course, you have like OpenAI, like whether it's Codex or chat, and then you have Anthropix products. I think that what the magic behind Muse and as well as Instinct is, is that the chat is like where all your context is. Like your context is just like one long running, single threaded, you know, piece. And so everything that you do in your life kind of goes to the central interface, Whereas all the other products are done by task, like research agent or publicity agent, whatever it is. And so I think it's just two different stylistic choices, which I think will unlock consumer demand, having it in a single thread.

11:26What's your assessment here on how this plays? Do you think this is going to be a winner-takes-all type of product in the end? Do you think that there is room for six or seven leading products, and then people just pick which one they like? I mean, what do you think? Good question. I think we're still so, so, so early. Like Meta and Instinct are, you know, their U.S.-only instinct. Probably, you know, most of the VCs and founders are probably using it as their customer base right now. And so it's probably a few hundred thousand power users, is my guess, max. And so I think that we're still so, so, so early.

12:04My guess is that, like, I think a lot of people will win. Like, I think the international market is up for grabs. I think, you know, Meta will probably win that, given WhatsApp. I think that you can segment it by, look, there's global admin work, there's travel, there's bill negotiation, there's scheduling. Like, even if agents, like, take 1 % of all of those markets, like, that's a massive market. So let me ask you a more pointed question. Do you think meta is a winner in this category in the end? Because, I mean, this is the big bet that they're making. Certainly, even with the coding agent, I mean, this is not an arena they've played in before.

12:44Yeah, I think this will be meta's big win. Yeah, I think this will definitely be meta's big win. Okay. And all the CapEx will be worth it because they've got paid tiers too. So we need those to catch on. So, all right. Well, Jay, I want to thank you for coming on. It was a great review. That is Jay Gupta, a partner at Foundation Capital here on TI TV. OpenAI announced that its latest unreleased AI model solved one of math's millennium problems. This is a problem that apparently mathematicians have not been able to solve for more than 100 years. I want to bring on someone who himself has pushed these models to their furthest limits for advanced mathematics to help us understand the significance of all this and also some of the debate that is shaping up around it.

13:35Johan Lend, he is the chief product officer at Samsara. Johan, welcome to the show. It's great to have you here. Hey, thank you so much for having me. It's great to be here. So I'm going to be honest, Johan. I like math, okay? I did the times tables as a kid. I loathed the math contests that we did in middle school and high school. I hated them. I was never good at them. And so I have really glazed over every single headline that has been about this model can do this, you know, Iliad contest or whatever it is. I don't read them. I don't like them. But this one we cannot ignore. Okay. And so I want you to understand what was the math breakthrough that opening I had?

14:15Why is this so important? and then we'll go from there. Yeah, absolutely. So first of all, thanks a lot for having me. I competed myself actually in competition when I was young in the International Math Olympiad. You were probably good at them. Yeah, you were probably good at them. Yeah, exactly. So what happened is Navier-Stokes is the problem, but actually Navier-Stokes is just an equation that describes fluid dynamics. It's an established thing. It's creating great things for humanity and there's no drama about that. But there is an element of Navier-Stokes which is like, can it behave in certain way under certain conditions where you enter into singularity?

14:49Like one way of explaining this, if you've ever done like the bomb, when you jump into a pool, like you form like a bomb and then it splashes up. The fluids like push to the side and then they come together and they kind of say like it splashes up in really rapid fashion. That's like maybe while you could interpret this as a singularity. So in fluids that has a viscosity in a limited fashion, can the fluids actually reach a speed that is unlimited? Right. And so you have to prove this math problem, essentially. This is a problem that people have not been able to solve. Exactly. You need to prove either that it exists or that it doesn't exist.

15:30And what OpenAI was able of doing is that they were able of creating a scenario where this actually happens. And we also formalized it completely. And I've looked at their proof. I looked at the formulation and like, to me, it all looks accurate. And there is actually by now, there's a very small part of the math community. They're challenging it. It seems like this is accepted. Like this is going to go through. What are they challenging? What are they challenging? The most challenging piece is actually where the drama is. And this is like, I mean, if you got the popcorn out, this is where you really want to sit back.

16:04I think that's what you're here for. I mean, lay it on. So look, there's two guys. like so Tristan and Levent. And over the last year, they have been collaborating on this problem, right? And Levent is from NYU and Tristan is from Anthropic, right? And as part of their collaboration, obviously, they've used a lot of Anthropic models, but they've also used codecs from OpenAI. And then, you know, three weeks ago or so, they get a breakthrough on this project, on this problem. Like they think they probably have it. A week later, they've formalized it, but they don't really like the proof. So they like take their time on kind of seeing it through.

16:39And they probably also didn't have the full solution, by the way. But they had probably - Were they using AI to do - Absolutely. This is full on AI. Absolutely. Absolutely. But then OpenAI hears the rumor that there's a breakthrough, right? So a week ago, what OpenAI does is that they start just, it looks like in desperation, spending like$10 billion plus plus on like 10 ,000 plus agents trying to solve the six remaining Millennium problems. And they get traction on exactly Navier Stokes, right? So then they, but this gets even better, because then Levent and Tristan, they hear the rumor, right, about OpenAI side breakthrough.

17:19So then they reach out to OpenAI and here, and all this so far is like fact is well documented. And here's where it gets blurry, because it's almost like OpenAI kind of invites them to collaborate, but they kind of don't want, it seems like Levent as part of that because he's an anthropic, but they're happy to collaborate with Tristan. So Tristan rejects them. And then OpenAI just goes ahead and publishes the proof. Right. And then Tristan and Levent is like, but hang on a second. This looks similar to our proof. And we were using codecs when doing it. So, dear OpenAI, did you actually look at our work when you created the proof?

17:55And that's the drama. Wow. Wow. So let me ask you this, Johan, whose side are you on? So look, all this drama in my mind, this is just a symptom of a bigger thing that is happening here. Because what is actually happening is that the science of math is kind of falling to AI. And this kind of drama is just a symptom. Like people are trying to figure out how do we relate to this and whatnot. And the reason why math is, in my mind, the first science to fall to AI is that everything can be validated. You have a proof, you can check it with a computer. And if that checks out, it's true. It's like done.

18:35And there's nothing else like that. Like if you do it in chemistry or physics, something, you need to run an experiment in the real world. It takes time, it's uncertain and whatnot. So that's why math is the first thing. So, I mean, that's all great. And look, I, you know, we, we hear of these breakthroughs continuously happening and today it's math and, you know, yesterday, you know, may have been something like the, you know, like the MCAT or the bar or, you know, some kind of an exam that people, so the models, I mean, they keep hitting these types of breakthroughs. My question for you, Johan, is why is it that the models can do these crazy complicated tasks?

19:16And yet, I mean, you know, when you're talking with the AI in your DoorDash order, for example, I mean, it still messes that up or, you know, it hallucinates with some simple task I give it. Like, help me understand the mix mismatch here. I mean, some of those, in my opinion, is like people building bad products and not using the AI in the right way. And that will catch up. So, like, I'm like a huge believer in this. Those are like, those are growing pains that we're seeing. Like, in my mind, we have now entered into the post-AGI era. And that's why we see math being the frontier. Like, you think, when new models were released, we used to look at it like, what happened to the benchmarks?

19:58And then we looked at the percentage of this and that and whatnot. The thing is that models are kind of, they're hitting like 100 % of this. It's because the benchmarks were designed by humans to measure and how are the models doing in comparison to humans. Well, guess what? The models are now better than humans. So you need to measure something the humans can't do. And that's math. You test them on unsolved problems because humans can't do them. The models can do them by definition better than humans host AGI. So that's the kind of thing that is happening right now. And just to confirm, the model that OpenAI used to put this proof together, this was an unreleased version of Astra.

20:36So this is a model that is much more capable than what Astra is right now, right? Yeah, exactly. It is much more capable. And that's not the one. When I do my proofs, I produce this one. I've solved four unsolved problems. Not as prominent as the Millennium Prize problems, I should say. But still, problems that are like 50-year-old humans have tried to do them and they haven't been able to. And now I push them, but I'm using Astra, Fable 5.1, Gemini 3.8, Kimi K3, Gm5.3, et cetera. Like they're all of the models. And I deploy them as a swarm to really like attack this. But these are, what they are using are unreleased models that are even more capable.

21:13Do you have people that yell at you as well when you solve them with the models? So yes, I mean, look, this is a community now. No, I'm just, you know, it's a rivalry. I get it. You know, people, it's competitive. But it's like, when I think about it like this. So music, there was some time in music when people were doing it with instruments and orchestras and whatnot, and then came like electronic and digital music. And people were like, no, that's not real music. But right now, no one would question, it's like Taylor Swift or Michael Jackson, are they real musicians and artists? Everyone would be like, yes, of course.

21:49This is a transition phase where we're accepting AI as a real tool to prove mathematics. And we see the growing types of that. Johan, let me just ask you about one of the headlines. We're in this moment here where models can do all this stuff, and they're remarkable in so far as how strong they are. There's the whole safety side to this too, which is that, okay, we're in the post-AGI era. Are they good? Are they dangerously good, right? And this is what the anthropic researcher flagged this week. I mean, he resigned and, you know, he said, look, it's going to get out of control. Like, you know, I tried OpenAI.

22:28I tried Anthropic. I was not impressed with either of their approaches. What's your reaction to this? Because this headline, by the way, today, I mean, we see this happening time and time again, right? Just like the math problems, we see people who come out and are outspoken about the safety component here. So what's your reaction every time you see one of these researchers coming out and saying something like this? So look, there's a researcher at MIT and he defines like 13 potential scenarios for like the end state of this. Some of them are happy. Most of them are negative. And when I look at the happy ones, they don't seem too stable to me, right?

23:06It's like, I'm not sure that I believe greatly in an awesome outcome here, but that doesn't have to be. The decline of humanity can be a happy path too. We might live in abundance and just a slow decline when we decline as the AI takes care of us and we'll live out our lives in a happy fashion. Maybe. What do you mean by they don't, explain more of this happiness stability thing. I'm not catching it. I mean, I'm not sure that the decline of the Neanderthals was so dramatic. They must just have, maybe they lived out their lives in their caves and had their lives and they were left, but then we just out-compete in a happy fashion.

23:41It doesn't have to be brutal. And maybe the AI will provide for us. You know, they will have us live like zoo animals and give us all we need and will die out slowly. So you believe the researchers saying that it could get out of control. You believe that? I struggled finding a scenario that is logically consistent where that isn't the most plausible outcome. Yes. Now I see why you were good at the math contests, my friend. I'm sorry, I wish I could. But I think it's actually important, right? Like to the extent that I am deep in this and I believe I am from some aspects, like I think it's important to speak the truth and like speak it with clarity.

24:19And yeah, I think that the situation is alarming. So then do you think that the AI labs are doing enough safety wise? I think it's probably not their role. Like this is society. Whose role is it then? Whose role is it? The problem is that this is a prisoner's dilemma, right? Anyone can cheat. Like if you say the nations are the ones, well, any nation can cheat. Say the companies, well, any company can cheat. You take the individuals, there are so many individuals, certainly one individual will cheat. bring in open weight models in this, and then like you can sheet through them because they're like three, six months behind only.

24:53So it's like, it's hard to pinpoint whose responsibility is it? And that's why it's hard to find the logic by which you get to stable scenario where there is a outcome with the longevity for humanity. Might be a happy decline, as you say, but like longevity. Okay. All right, Johan. Well, I got a lot more questions for you, but we'll have to bring you back on to unpack it a little bit more. I want to thank you for coming on. That is Johan Lahn, Chief Product Officer at Samsara here on TITV. The White House continues to juggle its perspectives on how to regulate AI as we all are, and new exclusive reporting from our AI and politics reporter, Leo Schwartz, reveals that the administration appears to have a whitelist of firms allowed early access to frontier models, only that it is very difficult for anyone to figure out if they are on the list.

25:48I want to bring on Leo Schwartz to chat more with us about his reporting. Leo, welcome back to the show. It's great to have you here. Thanks for having me. I feel like I have to go build an underground bunker based on your last segment. I was going to say, but it's okay because we're going to have regulation. It's going to be safe. There's a list of reputable firms, I take it, or maybe there's not. I don't know. You reported the status of these discussions. What do we know so far? Well, unfortunately, the way my reporting played out, it seems like we have the opposite, which is a continued lack of clarity.

26:23We don't have a list. We have a draft. There's a notion of a list. No, I mean, I feel like I keep coming back on this program with the same message about a lack of clarity, uncertainty that the AI industry keeps harping on. But given what your previous guest said about the need to have some sort of coordinated response, ideally by the government or some central group of actors, there needs to be a way to understand, okay, we have this set of new, very advanced frontier models. Who should actually have access to them to make sure before they're live and deployed to the public that potential vulnerabilities and anything from the grid to the banking system is actually found?

27:03this was a key part of the executive order the voluntary ai framework the white house right that's what that's what we want we wanted we wanted the clarity but what's what's the current state of it so so the issue is the the framework or the eo laid out this trusted partner program this way that once a model goes through the testing program from the white house then that model can be made accessible to a number of trusted partners you early access partners whether it's banks or power utilities uh the issue is that there's a uncertainty over how that program actually works who gets access to something like mythos or astra or gbt 5.6 cyber um it seems to be the white house dictating access but without clear rules the road or criteria for how it's determined who's actually getting access to these models right Right, right.