Google Flew Me To London To Test Project Astra (Live AI Demo)

I think Google is going to blow a lot of people's minds, honestly, with what's coming. They let me walk around London and play with this Project Astra. They gave me a phone with it pre installed on it. This is one of like, I don't know, 20 phones at all right now that have it installed.

It's amazing how much they've caught up with OpenAI.

It was legitimately mind blowing. Like Linus actually tried something that they showed off and like fell back under the couch. He was standing. He like fell over. He was like such in by what he was witnessing.

The competition is really heating up. Like I think over the next year like things are going to dramatically change because like now Google is in the race.

Hey, welcome to the Next Wave podcast. I'm Matt Wolfe, I'm here with Nathan Lands and I recently had an amazing experience where I got to early test Google's newest project, Project Astra. And I got to sit down with some of the people that actually built Project Astra. So, so in today's episode, I'm going to actually demo Astra Live. I mean, Nathan and I are going to ask it questions live on this episode. We're going to show you what it's capable of and then we're going to dive in with the creators of this project and ask them all about what they built it for, where it's going, answer some of your privacy concerns. Talk about hallucinations. It's an amazing discussion. So let's just go ahead and dive right in. Google DeepMind recently flew me out to London and what they showed me at this event was their brand new Project Astra. And basically what Project Astra is, is it's their new version of like the Google Assistant. But it has vision and it was built multimodal from the ground up. Right? So it actually isn't listening to audio text, transcribing the text and then using that text, which is kind of what ChatGPT does. It actually hears the text, it actually sees the video and, and it, it has an understanding of text, video and audio. So it's like a completely different style of language model than what we're seeing from a lot of these other models that are out there. And they let me basically walk around London and play with this, with this Project Aster. They gave me a phone with it pre installed on it. This is one of like, I don't know, 20 phones like at all right now that have it installed. So I'm actually able to play around with it and use and I Also had the opportunity to sit down and talk with one of the research leads on Project Astra, as well as basically the lead project manager on Project Astra. So I sat down with two people that are running this, and we'll roll that interview in just a few minutes here. But before we did, I want to kind of talk about Project Astra and how cool it was and, you know, the experience in London and stuff like that.

Yeah, I'm jealous. I didn't get invited. I was like, maybe they heard me talking crap about Google all the time.

And then, well, you have no idea how last minute it was. I literally got my flight booked like 48 hours before I left. Like, it was like that quick. And they rolled out the red carpet too, right? Like, they flew me first class. They had like, super nice hotel. They did like, really fancy dinners in like, some of the nicest restaurants in London every night so that I could. We can all like, meet the other creators and get to know each other real well and stuff like that. And then we headed over to the DeepMind headquarters in London and basically they put us in like, this boardroom and they gave us like this. I'll have to send Darren the video footage of some of this, but they gave us this, like, presentation where there was like these big Gemini boxes in front of us. And when we opened the box, the boxes sort of opened like this way, right? When we opened the box, it, like a glowing light came out of the box and the phone popped out and it didn't make a sound, but like, you can almost hear like, like as the box opened and the whole thing lit up. But then they. They had us sit down with the DeepMind COO. I got the opportunity to talk to the CEO, Demis Hassabis, you know, and it was just. It was this really, really cool experience. But they walked us through how to use Gemini. And then they set up all these little booths around the DeepMind office. They had a booth that was like a art gallery, right? And we were walking around this booth and holding up our phone at the different paintings and asking questions about the different paintings and who painted them. And is this the original version or is this like a reproduction? And it was able to figure out all that stuff by looking at it. And then they moved us into like another booth. And in the. The next booth there was like a bar set up, right? And in the bar they had like all of these, like drink mixes. And they. They held the phone up and said, with what's here, what drinks can I make right? And it started suggesting drinks. And then the guy said, okay, I want to make this. How do I make it? And it started walking him through the recipe to make it. And he would go, okay, I did that step. What's next? Okay, I did that step and made the drink while using Gemini the whole time. And then they took us to, like, another booth where they had a bunch of books on the table. And they opened the book to, like, one of the pages on the book, held the phone over the page and said, tell me what this page is about. And, like, the phone wasn't looking at the book anymore. They just did it real quick. They went, tell me what this is about. And, like, move the phone away. And Astra started saying, oh, this is the page where this happens. And this character did this. And it, like, read the whole page in like a split second. And then they walked over to a whiteboard with a bunch of math problems on it and started going, can you solve this for me? Can you solve this for me? And it was solving the math problems. And then they pointed it at a plant, right? And they were like, what kind of plant is this? And it started telling them all about the plant that they were pointing it at. And it was just like this really, really cool experience where we got to see all of the use cases. And then they unleashed us on London to just wander around London and play with the phones.

So how was it? I mean, like, everybody's saying that the new stuff that you came out with is awesome. So, like, what's your, you know, hands on experience with it?

Hands on it is really, really good. There's. There's a few features that they haven't added to it. It's capable of doing the things, but they haven't actually rolled the features in yet. Like, it doesn't actually connect with Google Maps yet, it doesn't connect with your calendar yet, it doesn't connect with your Gmail yet, things like that. They said that it will. They have those capabilities. They just haven't rolled it into, like, the demo version yet. Right. So a lot of that kind of stuff it felt like it should do, but it didn't yet because of the sort of just beta testing of it. Like the account they set me up with on Astra is a test account. It's not using my real Gmail. So even if it was connected, it wouldn't know my calendar and my Gmail anyway because it's like a. A demo email that they set it up with for me. Right. Other than that, it was really really good. Like you can walk down the street, hold up the phone and go, where do you think I am right now? And it would look at the buildings and the landmarks around you and go, oh, based on what I can see, I think you're in London. Oh, okay. Well, based on where I am, what should I go visit? What should I look at? What's some interesting sites around where I am? And it would say, oh well, Wembley Stadium is right down the street. It's in walking distance. Why don't you go check that out? And it would start telling me things that I can go look at that were around me. I would point it at trees and go, what kind of tree was this? And it would explain the tree. And I pointed at like a car that you only see in London. Like, you know, the very, very British looking taxis that you never see in the US And I'm like, what kind of taxi is this? And they're like, oh, this is this kind of taxi. They're really only found in London. Right. And it started giving me details about the taxi. I found it to be really, really cool. One thing that we did get to demo and you know, I'm not allowed to go into too much details, but I am allowed to tell you that I demoed it was they put the same technology on glasses as well. So I did get to test the glasses a little bit as well as walking around on the phone. And this technology, once it's into glasses, that's where it's really game changing. Like I don't think people want to walk around town like with a phone all the time.

Yeah, it sounds cool, but like who's gonna do that?

Yeah, yeah, but the glasses are sort of like this always on kind of thing, just like this. And it could see what you're seeing and give you the same types of responses that this can do. It's just in that form factor now. It's still processing everything on the phone. Right. It's not like the meta Orion glasses where like you have a separate like puck in your pocket or whatever they called it. It's still using the phone to process it in the same way like a meta Ray Ban would. Once they're actually on the glasses, that's when I feel like, oh man, this is really powerful now because it's just like whatever you look at you can just ask questions about and it will just respond to you in your ear.

Oh my God. I'm sitting here imagining all the, the manga in Japan I can read.

That's another thing I Didn't mention, which I think you would really appreciate because obviously, you know, you have a language barrier with your wife.

Yeah.

Is it does like almost instant translation. So I can point my phone at like a. They actually one of the. The setups they had was the. These posters on the wall that were in different languages, like Japanese and one was in Turkish and one was in Arabic. And you could point your phone at the posters and. And say, what does this say in English? And it would translate it for you. Right. And it would just like read it out loud audibly for you. So it can do stuff like that. But one thing I thought would might be kind of cool is I might even be able to demo some of what it's capable of through this podcast.

Yeah, that'd be awesome.

Okay, let's see. So I'll hold it up to my microphone. Hopefully it'll hear.

Hello again.

What am I looking at right now?

Based on the image, you're looking at an elgato prompter displaying a video feed of someone likely engaged in streaming or online communication. The accompanying text from the source suggests this is a popular YouTube setup and was custom made.

That's correct. What can you tell me about the Next Wave podcast?

I can certainly look into the Next Wave podcast for you. Just one moment while I find some information. I've found that the Next Wave podcast focuses on AI and its impact on business. It's hosted by Matt Wolfe and Nathan Lands and aims to make AI expertise accessible to everyone.

That's pretty good.

How do I say Merry Christmas in Japanese?

The phrase you're looking for is.

My apologies. I seem to have jumped the gun there. The phrase you're looking for to say Merry Christmas in Japanese is Merry Kurisumasu. Merry Christmas.

Does that sound right?

Yeah, that sounds. That sounds right. That's why I was doing it. Yeah.

Okay, how do I say, let's see. How do I say where is the library in Spanish?

Donde estala biblioteca?

Well, I know that's right.

This is amazing. I'm sitting there thinking all the different use cases for learning.

Look, if you're curious about custom GPTs or are pro that's looking to up your game, listen up. I've actually built custom GPTs that help me do the research and planning for my YouTube videos. And building my own custom GPT has truly given me a huge advantage. Well, I want to do the same for you. HubSpot has just dropped a full guide on how you can create your own custom GPT. And they've taken the guesswork out of it. We've included templates and a step by step guide to design and implement custom models. So you can focus on the best part, actually building it. If you want it, you can get it at the link in the description below. Now back to the show.

Okay, I've got the book open to one page. Can you tell me me what the book is by looking at this page?

The book you're holding is Never Split the Difference by Chris Voss.

Is this actually available now or is it like. Or it's just for you? So like how soon?

Right now? Right now is just trusted testers. That's one of the questions I asked when I had the opportunity to sit down with Bebo and Greg from, from the Astra project. And they basically said right now we're, we've got it in the hands of testers and we're, we're basically working out bugs we don't have a specific timeline for. We're going to release it publicly, but it won't be called Project Astra when they release it publicly. It's just going to be the Google Voice Assistant.

Right?

Project Astra is just like what it's called right now, but eventually it's going to replace the assistant that's on all Android phones.

How cool.

But they didn't actually say exactly when. I guess just whenever they're done doing all their bug testing and they feel that it's good enough. I mean, in my opinion, I, I haven't run into too many issues. In fact, I was out in London with a guy at Linus Eckenstam who's, you know, on, on X. He's a big like sort of AI influencer on X or whatever you want to call it. And he was out there and he actually forgot to turn off Astra before going to bed and he said it on his nightstand in the hotel and he woke up in the next morning and just like rolled over and said, astra, are you still there? And Astra is like, yep, I'm still here. Is there anything I can help you with? And it was just on and live all night like listening for the next command, which is pretty crazy to me.

That's awesome. Yeah, I've been using the advanced Voice mode on ChatGPT more and more on the, like the desktop app for Windows. So man, if they, you know, if either one of them adds that, like whoever adds that first, I'll probably be using it. Like if they can actually see my screen and interact with things, I'm sure I'm doing on the PC while I'm doing it. It would be so cool.

Yeah. So I actually sat down with a couple of people who worked on Project Astray. I sat down with Bibo Shu, who's like the, the lead project manager. And I also sat down with Greg Wayne, who's like the lead researcher on the product. And I asked him all sorts of questions. I asked him things like, you know, what's product, what is Project Astra, where is it going? What are some of the future features? You know, should people be concerned about privacy? How are you handling privacy and what information you're giving them? You know, how do we handle things like hallucinations? When are people outside of the trusted tester program going to get access? So I asked them all of these sorts of questions. So let's go ahead and roll that interview real quick and we'll be right back.

All right, well, let's, let's just dive right in and talk about Astra. Can you, in your own words, explain to me what Astra is? What makes it different from, you know, the other AI platforms that are out there right now?

So Astra is a research prototype from Google DeepMind for exploring just sort of new capabilities of a universal assistant. So it's, it's not like a full fledged product. It's not, you know, shipping to like millions of users. Yeah, it's a typical Google product. So we're keeping the scale pretty small. Our goal is to think about what is necessary for a universal assistant, for an assistant to be really capable to understand the world, to be able to converse with you and eventually be sometime to do stuff for you and prototype this with AI and with the latest research capabilities. And we put it in the hands of users to see what they will do with it. It has a bunch of interesting capabilities like multimodal conversations, so both vision and audio. And it understands your world, it can converse with you, it's got memory, it knows how to use tools. So we have kind of like a very interesting set of things that are together. But I think Greg also has a pretty interesting way to describe the sort of division around the situated dialogue.

Yeah, I think it was like, kind of like a reduction or simplification of what it is to share experience with another and engage with them in dialogue and collaboration. That was kind of like we were trying to capture in sort of as simple a way as possible. And I was saying that there's somebody who was.

An early internal tester said it had the feeling of being with somebody and I was like, Got it? Yeah, got it.

That's what we're going for. Yeah. So is the idea to sort of turn it into like your own personal agent that can do things on your behalf and, you know, maybe do bookings for you and help manage your calendar and help manage your emails and is that sort of the longer term vision of it?

I think so. So, you know, we're, we're starting from kind of the capabilities of what like Gemini is able to do. I think first understanding you, understanding your world is one of the first pieces. But of course, if you want to call an assistant, it has to do stuff for you. So that definitely are things we're looking into.

So I have a question for you. When it comes to this is sort of off topic from Astro, but I.

Want to kind of get into the.

Weeds a little bit when it comes to something like hallucinations. Right. In some scenarios, hallucinations are sort of a feature, not a bug.

Right.

When you want AI to be creative, you want it to hallucinate, right. But in other situations where you need it to be factual, you don't really want it to hallucinate. Do you see a pathway to where we can get some of these tools to get to near zero hallucination when you don't want it to hallucinate?

So first of all, where do hallucinations come from? Part of it is the model not being powerful enough. So in certain cases, the smaller, older models had more hallucinations. They were less exact about their predictions of what language to use in various moments. Another, I think piece of the puzzle, when people talk and they generate all the data that's on the Internet, which is ultimately what grounds these models with some knowledge. They have internal states which are we're not accessing. So if I say, oh, you know, I have a stomachache or something like that, that's somebody might write that on the Internet or like and whatever. And then the models might say that. But they don't have stomachs, right? So they're replicating the observables of some human dialogue without actually having the internals. The same is very much true for states of metacognition knowledge. If I say I know or don't know something.

To replicate, that would require the model to actually have also identified that context, grounding why I said I do or do not know something. I think ultimately it's maybe it's difficult to know how to imitate that properly just from these verbables. What you really need to do is like also generate yourself and try Things and experiment and then show, realize you don't know and so forth. So in the kind of current state of the world where models are mostly, like, built without actually, like, experiencing the world, they really don't know what they know yet. And I think that once models are actually learning through their experiences, they can identify how they know things. There are these Himalayan languages wherein when you say you know something, you have to say how you know it every time. It's like a part of a marker. So you say, I saw it with my own eyes. That's a part of the morphological unit you append to the word no. Or I heard somebody else tell me that who saw it with their own eyes, or it's just rumor, or I deduced it. So the way in which we will ultimately eliminate hallucinations is models will.

Experience, generate things, be tested against their own, you know, in the world, and ground their basis for their factual assertions through how they know things. I read it. There is something they'll be able to say.

So it'll almost get to a point where if you read something and it states it as factual, but there's maybe no sort of source for it, maybe you trust that information a little bit less.

Yeah.

So, you know. Yeah.

The simplest kind of way of eliminating hallucination is just say, here's a website citation. But that's actually deeper than that. It's. It could just might not just be that's on this website. It might be you and I together, we're hanging out and we saw this together, and that's how we know that. Right. It's a whole part of learning from experience, which is actually a part of where we're getting with astro, with personalization, learning with you and so forth.

I have. I have a question about, like, the sort of training data. There's been a lot of talk lately in the media about this, like, wall. Like, we're hitting a wall. We're sort of running out of training data to make these models smarter. Do these models still get smarter? Is there, like. Is there. Do you think there's a wall? Or do you think, you know, maybe. Maybe they're getting smarter on the inference side instead of on the. The training side.

Like, how do we.

How do we continue to get smarter and smarter models if we have already scraped the entire Internet, essentially?

You know, we've trained on a lot of text. I think a lot of the code has been trained on. It is now possible also to train in different ways on similar data. So, for example, I'll give you an example code execution correctness is a possible source of improvement, you know, not just like here's predict this code. Like you know, here's a design specific, write the code and then once you write it, does it actually function in interpreter? That's a new way of getting improvement. And data back to factuality too. You're actually trying. The model's trying and learning from reinforcement, learning how to get better. So we haven't fully tapped that stuff, I think by a mile. And we use much less data for that than the Internet's worth of data. And it's much more efficient to learn from that. So that's one thing. Also there's sort of like, you know, Astras natively works with native audio. I think we barely tapped what can be done with native audio interpreting kinds of audio. So there's a sense in which maybe even if it's not sort of like more intelligent in a kind of book smart sense, it's still learning more about the kinds of things, more aspects of our reality that we haven't like worked on so much before. And that's data we already had and so forth. So we just haven't used all of it. Right?

Yeah.

And you know, there's data sources, you know, text is one kind of thing. But I think we're still only at the, you know, starting off with video effectively in time, you know, you know, Astra or Gemini models are not the same as veo, which can do real video generation. You could imagine a future in which those are in one system. What does that mean? Maybe they can make physical predictions and talk about that too. Right. So yeah, I think you'll have combinations of these things that Adamis gave an example of. Alphafold. Right. Alphafold uses special protein data banks. We didn't train Gemini models on those, but you could imagine a joint model that's trained on that stuff or on all the weather reports and so forth. And then it would be able to know about all sorts of things and still talk to you. So I think that we. There's no sense in which we've hit a data wall. We've hit a data wall in some areas, some. Some narrow niches effectively.

Right, right. And then Bibo, we were talking about this at lunch, actually is, you know, the more personalization too. Right. So as you use something like Astra, it actually starts to learn you and your own interests and your own habits. And then over time it becomes more and more tailored to just you and what you would use it for. I'm curious, what are Some of the like really interesting use cases you can see Astra being used for.

Yeah, I mean the most obvious ones that jumped to my, that just jumped out at us. I remember when we first kind of did the IO thing, everybody basically reached out us. You need to figure out a way to make this for blind and low vision users. And that's been a project actually the team has been working on and been interested in for some time. So that was kind of like the no brainer, right? So there are right now services that people call up to get some guidance about how to find a door. And you can see that, you know, this was the kind of stuff that we built for, which is this concept of someone being there with you, talking you through a thing. So I think that use case is really exciting. The other one that maybe touches a little bit on sort of memory and personalization potentially is like my favorite one has always been, you know, like, my washing machine is broken. Just like help me fix it.

Right.

And today still like there's no great solution from a user experience perspective. I have to go. It took me like five minutes to find the manual of a washing machine. I read it, I still don't know exactly which part, you know, you're, it's, it's just a, or you watch a video and you still don't exactly know what you're supposed to do. And what you want is like you want an AI to just go do that for you, go read it, go understand these things and come back and just tell me what I need to do now. I think the great thing is that like, well, if it had, if it learned to do that once, has it just like now obtained that skill?

Right.

And it now knows that that's the washing machine that you use. So next time something goes bad, it doesn't, it might not need to go through all that anymore. So I think like these are kind of like interesting things we've been thinking about. Like, okay, is there like, does Astra have like a place where it puts documents that it found on your behalf? Like how does that work and does that data, you know, that data clearly belongs to the user, but it's like kind of like Astra's low bank. But I think over time things like that will become pretty useful. You can imagine if the more we ask it to do stuff on your behalf, access information that you give it, the better it's going to be at helping you solve your particular task.

Gotcha.

Very cool when it comes to like the security and like the data privacy I know from a lot of the people that I've talked to, that's their biggest concern with AI. They're concerned that if I'm telling all of my life to something like Astra, then all of that is going to be saved on some Google server and Google's going to use that to, you know, whatever, you know, whatever Google's going to use the data for. So how, what is the approach to like the, the data security, data safety when using Astra?

I mean what we currently have is like kind aligned with a general Google product in terms of practice. I think we're a little, even more, a little more conservative. So you can choose whether you want to kind of have a certain session recorded or not. I mean it is a research prototype. So, so we encourage recordings because you can, if you have bugs, you can tell us, then we can, you know, you know, get it to fix stuff. But you can also choose not to record. And everything that is recorded is on my activities, which is kind of where all of your personal information at Google is stored. And you can go view them, you can delete them and so forth. But definitely, I know like this is generally a pretty sensitive topic Demis was talking about earlier about sort of private data secure enclave, like data that you can't see. I think this is a really interesting area. We personally haven't explored that as deeply, but our colleagues have. So there's definitely an area where you probably want some level of control in that sense in the future.

Right. And I remember Demis was actually saying that it's all going to be encrypted. And yes, it may be on their cloud, but nobody at Google can read it or access it or know what any of that information is. So my last question is, when it comes to Astra, is there any, like, what's the future of Astra? What's like the bigger picture? Where do you see it eventually going? Are there features that you'd love to have in it that maybe you haven't gotten developed into it yet? Is there anything you can share about, like where it's headed?

One thing that I might connect to, which was this core driving force for me was I was saying maybe about trying to encapsulate aspects of human communication, but we're still only beginning that sort of. So.

We talked about Astra being able to participate among groups. I think it would be nice if our technologies of communication were not only solitary technologies, one on one with the AI, but think about like a group of kids talking to Astra and it knows who's who and calls on the right one to moderate a game or something like that.

Yeah.

And I know, you know, there's probably things that you guys have in mind that you're not allowed to say yet. So this is my real last question. Any ideas of like, when Astra might be available to the general public to use?

Our current goal for Astra is really focused on trusted testers. Astra will be available to general users through Google products. So the Gemini app also aspects of Astra clearly applies to something like Google Lens. So that's our path for getting it into the hands of kind of general users. Astra itself. Our goal is basically to keep prototyping and keep trying to figure out what's going on with the latest models and how to get it to do something that's interesting and new and useful on behalf of the user, how to get it smarter in terms of intelligence, in terms of reasoning. So yeah, so we're kind of in this sort of trusted tester prototyping phase now. Prototyping I think means that we take the latest research, we create something out of it and we do want to put it in the hands of users to understand what they would do with it. So you're now a trusted tester. So we're excited to say, okay, what would you do with it? And so I think, you know, a lot of the questions around the product side, it's been really eye opening just to say, whoa, you would ask us to do that. And we've been like learning a lot from even our just small group of trusted testers so far. So yeah, so I think we're kind of still being in that phase and definitely we've been working with our close partners and other Google products to get it to more users.

Amazing. Well, thank you very much. Appreciate it.

The technology is amazing. So I mean, I think, I think it's amazing how much they've caught up with OpenAI. Like it possibly they surpassed them in some areas, you know, we don't know yet. I think we're going to know more by the time the end of the 12 days of OpenAI is done. But it's incredible how they've, they've caught up for sure.

Well, I think, I think Google is going to blow a lot of people's minds, honestly, with what's coming. You know, some of the stuff that I'm not allowed to talk about yet because it is still under NDA was legitimately mind blowing. Like Linus actually tried something that they showed off and like fell back under the couch he was standing, he, like, fell over. He was like, such in shock by what he was witnessing. And then also within the same week that we were out there, they released Genie 2, which was like their world emulation model. They released their, their weather cast model, which was all about predicting weather patterns 14 days in advance. You know, just yesterday, as of the day that we're recording this, they just released a new quantum chip that solves something that was going to take 10 billion years. And they were able to solve it within, you know, minutes. Right. So I think Google is a lot farther ahead than a lot of people give credit to. I think Google is really, actually crushing it with the. From the technology standpoint, the competition is really heating up.

Like, I think over the next year, like, things are going to dramatically change because, like, now Google is in the race, like, you know, neck and neck with OpenAI. I think Microsoft's probably going to. That probably means Microsoft's going to invest even more into OpenAI, like, once they see the stuff from Google. So that's just going to. Things are going to move even faster. And then you've got Elon Musk buying the largest purchase of Nvidia chips ever. Like, apparently Elon and Jensen are like, you know, best buds as well now. So it's going to, I think 20, 25, things are going to change a lot.

Like, all the models are, well, in Microsoft. I mean, just touching on that real quick. OpenAI and Microsoft are trying to get rid of that sort of AGI clause so that Microsoft and OpenAI can even work closer together. So, I mean, that's probably going to happen.

I mean, that makes sense. I mean, we've talked about this on the podcast several times. Like, totally. You could argue right now that they've reached AGI, like, like how much AGI. But like, you know, you. You definitely could argue they've reached AGI at some level, if not now, within a year or two, for sure.

But I don't think OpenAI can, can do anything without Microsoft servers right now. Like, so much of the investment that Microsoft made is in like, essentially server credits to be able to use their servers. Right? So it's like if Microsoft went away and said, OpenAI, we're done with you, I think OpenAI would fall apart pretty.

Quickly because that might be why stores down right now. This might be part of all this, might be part of the negotiation.

It's like, it's like, oh, you need another $100 million of servers this month. Well, look, sign this contract.

Microsoft turned the switch off on them for today and was like, get rid of the AGI clause and we'll turn the power back on.

Yeah, well, I mean, I think Microsoft will have to, because, like, like I said, I mean, like, obviously they could argue right now that they've reached AGI, and if that happens, then Microsoft has no vested interest in OpenAI. That would be horrible. So I think they'll have to.

I was actually just talking to our mutual friend Balaval earlier today, right. And Bilavel used to work at Google, right? He was. And him and I were having these conversations today about how I feel like so much of the world forgets the humanity underneath these companies, Right? They just look at these companies as like, these. These corporations that are just trying to maximize dollars for shareholders. And, you know, in some respects, the companies are designed to maximize profits for shareholders, but the people in the companies are. Are not that. Right. Like, I. When I was out at DeepMind, I met 40 different people that worked at Google, right? I got to meet the person who put out that Genie 2 project, which was that sort of world simulation model. And he was super excited that. That he was. He was putting that out while we were there. And he couldn't wait to see, like, the world's response to, like, this new research that he's, like, poured his blood, sweat and tears into.

Man, I was surprised more people didn't talk about that. And I think that maybe even inspired a dream I had, like, literally a day or two after that, I had a dream which is. They probably hate this because it was. I wasn't using a Google product, but literally in my. Literally in my dream, I was using Chat GPT. And I know just because of, like, the color scheme of it, like in a dream, it looked like chatgpt. And as I was typing stuff, everything I was typing was like, being visualized in front of me. So as I typed in, everything was. Was there, like almost like a world in front of me. And as I typed, would change instantly. As I was typing, I was like, God, that's like.

So now.

Now that you're saying it, I'm remembering the GENIE thing. Like. Yeah, it's kind of like that, right? Where, like, you're gonna be like, generate worlds.

Yeah. And then imagine that in AR or VR, right now you're generating worlds that you're actually sort of immersed in. Like, what you envisioned will probably be a reality in some sort of VR form at some point.

And that's gonna create, like, entirely new experiences as a consumer, but also as a Creator. Like, imagine being a creator and creating something now and you're able to, like, see all this stuff as you're creating it. Like how, how fun that's going to be to create things like that. Like, imagine creating video games or movies or there'll be new forms of entertainment, you know, that people will create probably more interactive entertainment. It's gonna be so much fun.

Yeah, Yeah, I, I'm, I'm excited about where all this is going. I actually had a conversation with Linus when we were out in London that right now all of this technology is like, so amazing and so exciting and so fun to play with. The big problem these companies have to figure out is how do we make money off of it? Right. Because they're pumping tons and tons and tons of money into it and they don't want to be that company. That's, like, cool. Now we're, you know, we've got a chatbot in your ears all day long and, you know, every 15 minutes we interrupt you to tell you you need to go buy some chicken nuggets at McDonald's. Right. Like, they don't want to be that company. They want to figure out other ways to profit off of this stuff.

Yeah, hopefully they do. You know, like, the Google guys didn't want to be that originally either, with their ads and Facebook.

Facebook.

Mark Zuckerberg did not want to do ads like all those founders. My understanding is originally they were against the idea of ads. Originally.

Yeah. Sam Altman flat out said he was against ads as well.

Yeah.

Now rumor has it they're exploring the possibility of ads. So.

And it's, it's because, like, when you have a service that's expensive to run and you want it to go as large as possible to reach as many people as possible, you have to have a free tier. And if it costs a lot of money to run, that typically means, you know, unless you're, unless you're making so much money somewhere else that covers those free people you typically, you know, you cover with ads.

Yeah.

I honestly don't know how it's all going to play out, but, you know, the, in the state that we're in right now, the, the researchers that are building it, they're just excited about building something cool and game changing. Right. They're not in that wavelength of like the monetization side. But anyway, we should probably wrap this one up. This. That was my experience at Google. You know, hopefully you learned something from the interview about what Google's doing, what their motives are, where this project astra is going. Thank you everybody who tunes into this podcast. We really appreciate it. We've got some cool interviews coming up in the near future, some cool discussions that we've got planned. So if you love conversations like this, make sure you subscribe. Wherever you listen to podcasts, I highly recommend checking out on YouTube. We're very visual with this podcast when we can be, but we are also available wherever you listen to podcasts as well. So make sure you're subscribed and hopefully we'll see you in the next one.