Limitless: An AI Podcast - This Week in AI: Tesla Ending Model S, Gemini in Google Chrome, Grok Imagine

Episode Date: January 30, 2026

Tesla's shift from the Model S and Model X to focus on the Optimus Gen 3 humanoid robot is definitely surprising... if you haven't been paying attention.In other news, Figure’s revolutionary robot can autonomously unload a dishwasher, and Google’s integrated Gemini into Chrome for enhanced productivity. We also touch on Tesla's capital expenditures and OpenAI’s $100 billion funding round.------🌌 LIMITLESS HQ ⬇️NEWSLETTER:    https://limitlessft.substack.com/FOLLOW ON X:   https://x.com/LimitlessFTSPOTIFY:             https://open.spotify.com/show/5oV29YUL8AzzwXkxEXlRMQAPPLE:                 https://podcasts.apple.com/us/podcast/limitless-podcast/id1813210890RSS FEED:           https://limitlessft.substack.com/------TIMESTAMPS0:00 Intro0:46 Tesla Earnings Report Highlights8:20 Robotaxi Updates9:09 SpaceX IPO Announcement13:31 Figure's Humanoid Robot Breakthrough16:37 Google's AI Browser Innovations21:47 Alpha Genome27:30 Claude's Interactive Tools32:39 AI Operating Systems41:38 OpenAI's Big Raise45:27 Closing------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

Welcome back to this week in AI. It's been a pretty big week for the robots taking over. Tesla are retiring two of their most iconic cars, the Model S and the Model X, to make room to scale out their new Optimus Gen 3 robot, which is targeting 1 million units this year, very ambitious, let's see if they make it. Figure, another robotic startup,

also launched a new humanoid robot, which did something that no other robot has done before. Clean the dishes, which sounds like a really simple task, but actually requires 61 very precise local. locomotive operations, but the best part is that it did it completely autonomously. And in final news, the Chinese and Google both released two AI models. One can convert video recordings into production-ready apps, and the other might just make

the first major scientific discovery since we discovered DNA. Google also introduced their leading Gemini model to 3.8 billion people in the form of a personal AI agent in their browser. An exciting week. Let's get into it. Let's start with the Tesla earnings report because that was the highlight of the day for me, at least. I've been following these reports forever. In fact, it's funny, six years ago, I was watching a video that popped up six years ago this day talking about Tesla earnings.

They were a blowout quarter in terms of forward-looking guidance, a little bit less so on the actual numbers. The stock's down for today, but that's not the story. The story is the forward-looking future of what this company looks like, which is very much centered around autonomy and robots and manufacturing. In fact, there's a world in which five years from now you can't even really buy a Tesla. And that's signified slightly and early from the Model S and X depreciation. But I think you'll continue to see more of that as we go because Tesla is fully leaning into autonomy. So much so, in fact, that they're doubling their capital expenditure this year. I believe last year it was $8 billion.

This year they're going up to $20 billion in cap expense just to scale these production lines for things like the CyberCav and for things like the Optimus Humanoid Robot. which is really exciting because I think this year, paired with the Helix News, we're seeing a lot of humanoid robots, a lot of automation actually coming to the physical worlds, like entering meat space, we're going to be able to walk down a street and not only have a self-driving car roll right by you, but also see a humanoid robot maybe walking down the street or doing something actually useful this year. And Tesla has a plan to scale these at a rate that no one else really is, starting with a million units this year, and then going to 10 million, 100 million, and then they project up to a billion of these.

The idea is that these will be able to offload enough of the human capital workforce. It'll unlock significant amounts of GDP for the actual country. So this is a gigantic project that's getting underway. And what we're seeing now is the early seeds being planted through these huge capital expenditures happening across the globe at all of their gigafactories. This feels like a very prescient moment because Elon's clearly saying that Tesla in five years or 10 years is not going to be known to be a car company. It's going to be an autonomous robot company. And we've kind of seen the signs of that over the last year where Elon's kind of just spent all his time on full self-driving

and improving that neural network and focusing on his kind of cyber trucks, autonomously working in SpaceX and then this new Optimus Gen 3 robot. That to me is like a kind of like crazy switch, I guess, to still kind of like materialize in my head. And I don't think it's coincidental because this year was also, or rather, 2025, was the first year that Tesla's annual revenue went down. They only put out 1.64 million cars. I think that's down like around 9% from the previous year. And the fact that he's shutting down two of his most early models, I think it was like back in 2012 and 2015 that he released both the S and the X. He's now going all in on humanoid robots.

I feel like, Josh, that this is the same moment where Elon was like, no, FSD is going to be the biggest thing for Tesla cars. he's doing the same thing for robots now. So people might kind of take it as this is kind of crazy right now. This humanoid robot isn't really useful for me or can't automate a bunch of manual labor. But I think Elon sees something that we don't. Do you agree with this? Do you feel kind of like solemn that we're going from cars to robots now?

Has this always made sense to you? Yeah. Well, this is always made sense because they've telegraphed this years in advance. Tesla famously has released these master plans. Or Elon has released the master plan. Once every three to five years for the last 15, years. And it just very clearly states exactly their plan and what they're going to do. And going back and looking at the old content from me, the old content from the master plans, it's very clear that

this has always been the plan. They're just accelerating the advent of sustainable transport. And now they're kind of moving to a sustainable abundance through the optimist humanoid robot. But the idea is always just to move the world towards an autonomous future. And cars were the gateway drug to doing that. It started with the roadster. The roadster was the very high-priced model that wealthy people were able to buy to fund the production and development of the Model S and the X, which funded the production of the scale vehicles, which is the three in the Y, which is funding the production of full self-driving, as well as serving as the actual network in which they could distribute the software on.

So it makes sense to depreciate the S&X. I'd like to take a moment of silence for these unbelievable cars that changed the world. I mean, I would argue the Model S is one of the most important cars that was ever built and ever shipped because it came out in 2012. just five years after the iPhone. And it proved that electric cars were not only possible, but they can be great. And they can be better than gas vehicles across the board. And not only that, but they could be capable of autonomous transport in a world where all of those things were complete contradictions and believed to be impossible. And the Model S did that. And it went an amazing

15-year, 13-year run where they sold almost a million of these vehicles. It proved that it was possible. It funded the development of where we are today. And they're just unbelievable vehicles. A lot of people don't understand that the Model S and the X actually only account for about 3% of Tesla's vehicle sales now. It's a very niche audience. They're fairly expensive. They're over-engineered to hell. They have the craziest features, particularly the Model X. So it is very sad to see them go, but it is very much expected and very much exciting in the sense that that production line starts turning out humanoid robots this year. Well, it's also important to remember that robots, as it was, isn't just going to materialize as humanoid robots. It's also going to look like the

cars themselves, right? You mentioned like they only account for 3%. The rest of the vehicles that he's going to be producing are mainly going to be targeted for just completely autonomous vehicles. They won't even have a steering wheel in there. They won't have a steering wheel. It's a good point, right? Which is these vehicles are going to become thought of as less of a car where you sit in the front and steer something to kind of like a third space, to kind of like a living room where you can be productive or enjoy and consume entertainment when you get from as you go from point A to point B. And we're seeing this with his plans to roll out the Robotaxie Service. So, you know, at the end of last year, he was kind of like, there were numerous sightings of him trining Robotaxie. He's going to be

rolling out or planning to roll out the Robotaxi service to nine more cities this year, which is super exciting to see of nine cities in total. And I know that there are several reports at this point where it is good enough to drive much better with lower incidental damage than humans. I don't know if you saw this, Josh, but Lemonade Insurance, which is a very popular insurance startup here in the US, are offering drivers 50% off their driver's insurance if they enable FSD on their Tesla cars. And I kind of thought of this fun idea where it would be funny if Elon just kind of like packages his new FSD subscription that we spoke about on a previous episode.

last week with this kind of like insurance discount to kind of like increase FSD subscribers and stuff like that. Those are ways that you can mix and match this. I'm excited to see more robots roll out this year. Yeah, for the people who are listening, the new cities that you will be able to get a robotaxie are going to be Austin, Dallas, Houston, Phoenix, Miami, Orlando, Tampa, Las Vegas. If you live in any of those places by the end of this quarter, you will be able to try this for yourself, which I think is so cool. As it rolls out, I would encourage everyone to try. Like I mentioned last week I was in L.A. and using Waymos. And it's unbelievable if you've never experienced it before to see and feel what it's like to have a computer driving you around. So

far they have 600,000 autonomous miles driven on the Robotaxi Network. That number is going to increase. It is safer than a human driver as reflected with the insurance. Tesla actually also offers their own insurance, which incentivizes people to drive safely. So really cool updates from Tesla across the board. Very excited about that. And this isn't the only news from the Elon atmosphere, I guess, this week. there's also more SpaceX news, particularly as it relates to the IPO, which seems as if they're going to be raising even more money than previously thought. $50 billion, which is going to be, I think, the largest amount of capital ever raised for an IPO and the largest market cap of a company

to ever go public all in one. This is going to be huge. They're targeting June as the announcement and release date for this IPO to go public for people to actually invest in SpaceX. And SpaceX, to me, I mean, this is a huge opportunity. This is a huge company. I'm so excited for this to get into the open market and see how the public reacts. It's also one of the few IPOs that are raising at such a large valuation

that I think justifies the amount that they're raising. We're going to talk about Open AI raising $100 billion later on this episode. But like, I would say 50 billion is probably too low in an IP. And I know he's raising other rounds, but like he probably needs more money to achieve his mission of getting to Mars, the moon, and setting up, you know, these interplanetary energy factories and stuff like that that he's spoken about on X. So this is really cool and exciting. I think 2026 is going to be the year of big IPOs, Josh. We were talking about this

before we started recording. We've got what, the Anthropic IPO. You've got Open Air raising an $850 billion round now, and they're going to be presumably IPOing more than that. Like $1.4 is what $1.4 trillion is what some of the rumors are saying. Yeah, this is just a just, you're going to be, you just crazy to see. Yeah, we might get Stripe this year too, a lot of IPOs. And part of the value in which enabled Starlink and SpaceX to go public is the Starlink network, is the internet network. And as they improve this network through Starship launches, which allows them to put their new version three satellites into space, and for reference, those version three satellite launches are 20 to 40 times more effective than the current versions that they're pushing out right now with the

version two. So it's a huge increase in bandwidth. And that bandwidth is going to make its way into cell phones in our pockets because the direct-to-cell network is growing very quickly. If you're a user of T-Mobile, you can try this now. But Apple is now reportedly planning to integrate Starlink directly into the iPhone 18 Pro, which is coming later this year in September. So this is really exciting news coming from Apple. Again, outsourcing the software stack to someone else who is more capable and competent in doing this, in this case it's Starlink. And it solves one of the huge pain points for the perfect mobile device where my hypothetical

perfect mobile device has unlimited battery, unlimited connectivity, no matter where you are. And now this solves the connectivity problem where there's nowhere you can go where Starlink will not be. And that's a really exciting partnership

that I'm looking forward to. In other robotics news, I saw one of the coolest, but maybe unsexy to some, demo of a humanoid robot. What you're watching on the screen right now, and if you're listening to this, I'll describe it to you,

is a humanoid robot. which looks very much like a human, I have to say, he's just missing some clothes, unloading a dishwasher, walking across a full-sized kitchen, placing all these dishes and crockery into the cabinet, very delicate, precise movements I needed for this, by the way, walking back and refilling the dishwasher with all the dirty dishes.

These include wine glasses, by the way, okay? So for all of you haters out there that are saying they're using plastic stuff, no. He's using delicate glasses as well. Now, what you're looking at is the new robot from figure, a very hot robotic startup here in the US called Helix O2. And this humanoid robot is different from the other humanoid robots in a few different ways. Number one, it is the first robot to perform a four-minute fully autonomous task of this level.

Just get that into your head. That may not sound very impressive right now, but I remember a time where AI model LLMs could only work autonomously for 10 minutes at a time. And now they're doing weeks at a time. This is the same kind of moment that we see for, robotics, and I think that Helix is kind of like nailing that. To emphasize, it is completely autonomous. There are no humans. There's no teleoperation. You know who dares leverage teleoperation? One KX. Also, Tesla Optimus, dare I say, although the new generation should be fully autonomous

in some way, shape, or form. This is a really impressive feat, because the technicalities and the mechanisms that you need to achieve this is nothing short of just insane. So 61 locomotive manipulations are required for this robot to do this, but that's not even the most impressive part. Typically, these robots are pre-programmed to execute on a set number of different moves. So those 61 moves I just mentioned, you would need to hard code it.

And when it faces a scenario, it'll then activate that movement. This is completely different. So what Felix did was they launched something called System Zero, I believe, which is their new neural net, which is only 10 million parameters, but it basically ingests

what it sees around it through its palm senses, through its eyes, and reacts in real time. This is different from how the model previously used to work, which is take in some kind of input, process it through its hard code, and then output in action. And hey, Presto, the kind of like trophy win here is that it's fully autonomous and it's the first humanoid to be able to do this. Now, to remove all that code and transform it into neural net is cool enough. But what I also like about it is that they have this new form of tactile sensors. And I think we spoke about something about this on our previous run, or maybe it was even last weeks.

But because of its palm sensors and its new tactile sensors, it's able to know when to apply pressure and when to be delicate. For example, lifting a wine glass and placing it into a cabinet. So I personally took this kind of boring task and thought, huh, this is actually something that I would actually have in my home because I can trust it. Also, look at the speed that it's moving. It's definitely quicker than my sister that I've seen her

loading the dishwasher at home, at least. Do you think this is ready to be in your home? Is this something you'd be interested in buying? Well, I would need it to do something more than just dishwasher stuff. But presumably if it can do this, it can also fold my laundry, Josh, and load the laundry basket as well. So depending on what they're charging, if it's a subscription or if I have to buy it outright, would influence my district. If I could have a live-in housekeeper, sounds pretty good. I'm a pretty messy guy.

Yeah, when evaluating robots, I think the things that are most important to look at are first the hands, the dexterity, the sensors within the hands, how capable the hands are, because that's by far the most complicated thing. The second thing is the intelligence. How general purpose is it? We've seen this example over and over again of loading a dishwasher or doing simple tasks like sorting clothes. How complex are those tasks able to get and how quickly are they able to roll that out is the second thing? And then the third is how quickly and how effectively are they able to actually manufacture this at scale? Are they going to be able to create enough of these at a price point that makes enough sense with a feature set that makes enough sense to encourage people like me and you to want to buy

one. It seems like there is a very long and challenging path for figure to get there. I'm really looking forward to Tesla's Optimus 3, which is coming this quarter and will be rolling out at scale this year. And I am hopeful that figure can keep up because competition is good. And I hope they don't turn into like the ravine of electric cars. I want them to stay on the forefront. And I hope Brett and the whole figure team continue to keep their foot on the gas. Yeah, I mean, I think this demo is cool, right? But it's pitched towards... It's cute.

It's cute. But I think the real application, the reality is these robots aren't going to be in your homes first. They're going to be in factories. In fact, their Helix O1 robot has been working at BMW controlling their operating system, or one of their manufacturing plants for, what is it now? Four months now, 24-7? There's like a live stream of this.

Hopefully, maybe we can link to this and people can go watch it. but Brett keeps talking about it every single week. So I don't know, I'm bullish on it. I think Optimus is also going to be in factories first before people's homes. But we shall see. We have some big news from Google. Josh, you were showing me some pretty cool demos. What's going on here?

Yeah, big news. If you remember from a little while ago, we had the CEO of Perplexity on the show, and he was talking about the new pivot to the AI browser. Since then, OpenAI has released their own AI browser. Claude has released a plugin for Chrome, but now Chrome, where everyone actually uses their browsing, has rolled out their own Gemini-agentic controller inside of the browser. And it's pretty incredible. I mean, like you mentioned at the intro, 3.8 billion people on Earth use the Chrome browser. And now all of those people have just been given a magical update overnight that

allows this agent in and allows it to actually control and do things on behalf of the user within this browser. So we're watching on screen an introduction demo video that shows some examples, and I thought they were really amazing. The first one that they did was with Nanobanana Pro, which is the image generating model, and they were just looking at a home and they wanted to add furniture to it, like looking at an empty apartment. And you could just ask Nanobanana to stage the apartment. Yeah, here's the demo. Where it's empty, you could ask it to stage the apartment and it shows you what it looks like with furniture on the fly as you're browsing. The follow-up feature to this is they were looking for an apartment that accepted dogs that was in ex-proximity

to this place that had a whole list of parameters that were pretty complicated that would have taken the average person a long time to do. And they typed in that prompt and it actually clicked through the browser and navigated through this for them, putting in all the filters and figuring out which results were best for them. And I think this is such a great use case because it's just so prevalent in what you do every day as a user of Chrome and as someone who uses a Chrome browser. And I'm excited to see it integrated where people are, where you don't actually need to download a separate piece of software, move everything over. It's just meeting people where they are and building on top of this existing knowledge base that Google has about you through your Gmail account.

What I like about this is they're not overpromising or they didn't over promise Gemini 3 being pretty much wherever you are on the internet. Like I've used Chrome browser for years now. I've used Gmail for longer and the fact that, you know, these demos show kind of places that I visit every day. Because like, let's face it, like most of my online life or rather most of my computer life doesn't live on my desktop, right, which is my, which is separately my point around Claudebot. I'm like, it's not useful for me because I'm always on the internet. But Google is bringing Gemini and AI, cutting edge AI to me on the internet, on the browser where I am every day. Now, I might not be someone that kind of edits images all the time and says like, hey, like, show

me in this outfit, but I am someone that writes a lot of emails. I am someone that spends a lot of time in documents. And in some of these demos, Josh, it doesn't just like get me to prompt it. It prompts me and says, hey, I see that you have like updated some of these goals for the next month. Do you want me to email so and so on your behalf? I've generated you a pre-script if you want me to do that for you. All I need to do is look over to it on my task panel on the right and click send. And that's just like such a beautiful experience. It sounds again, unsexy, but I think this is going to make people way, way more productive and most importantly, loyal to Google. We mentioned their major release of personal intelligence, which is basically a

personal AI assistant that is prevalent across every single Google product that you use, not just on your browser, but on Google Maps, Android, wherever you are. This is it coming into reality just for Chrome. It's frigging awesome. Yeah, really nice feature. It's just, it's, again, it's another muscle that people are going to need to train as they get used to these new tools. in reaching for the agent in your browser versus trying to do it all yourself, where I would encourage anyone, as you're browsing and you're doing something, you're trying to achieve a goal, try to click that Gemini button and see if it can achieve that goal for you. And the process of doing that over and over kind of trains your brain to rely more on this. And it just leads to a far better quality of life

because it's better at doing a lot of these tasks than we are. The other thing that I thought was interesting to note is that they updated their Gemini AI Pro plans, the Pro Plan and the Old plan, which are either $20 a month or $250 a month. If you have those now, the big thing on this long list of things that they include is Google Cloud Credits. So now you can actually use, you could build your own environments for these AIs. You could start querying more of their API to get more, I guess more just requests on more intelligence. They're including a lot more tools in this package for people who are interested in just toying around with these products. And for me, this is really exciting because I'm not running large scale products. I don't need a

server that costs more than $10 a month. And essentially for one membership of $20, I get all of these really cool perks. You get to access the entire Google AI ecosystem. And I just, I liked this pivot that they're moving towards kind of giving people the tools to tinker at a very low price point at $20 a month. So in other Google news, they release what is probably the most impressive scientific AI breakthrough ever. And no one kind of noticed it because it got published in nature, which is a scientific journal, but it was 100% open source and super cool. So this is, the best way to think about this is this is kind of like a DNA sequencer AI. Now, the reason why this is cool is 95% of our DNA is actually not that important. It doesn't actually lead to protein production. The reason why

protein production is important is because it influences pretty much how your entire body, mind and world works, how your perception is, right? The issue is, we don't understand the 98% of that dark matter because it controls some types of genetic expressions. So for decades, humans have basically just kind of like mapped out the genome themselves and then like just tried a bunch of different tests to figure out what proteins or what genes are expressed in a, B, or C type of condition. Now there's an AI model that can not just sequence a couple hundred thousand of these base pairs for your DNA, but up to one million base pairs.

Now, there are four main base pairs in your DNA strand. This is, we're taking them back to like, you know, fifth grade biology here. And if you connect them together in different types of combinations, you can result in different types of proteins, which kind of make up your genetic makeup. Now you have an AI model that can just sequence this entirely and predict what the expression is going to be. Now, the reason why this is super cool is you can apply this to so many different things. You could apply it to preventative medicine.

You could apply it to personalized medicine to create personalized pharmaceuticals for yourself or you could recognize genetic disorders before it manifests, before it becomes an issue and kind of snip that baby out. So it is such a cool thing to see. It kind of harkens back to my biology degree where I'm just like, oh my God, I can't believe that they've produced a tool that can like, that could have saved me like months of work when I was doing that degree. But I just think it's super awesome. Josh, are you impressed by something like this? Yeah, this to me is the coolest news of the week, actually, which is funny enough. The biology stuff is so unbelievably impressive, and I think it's easy to overlook it because it's kind of complicated and difficult to understand

for those, I mean, like myself, who just haven't spent much time in the world of biology. But in the research and learning about this segment, I discover that the DNA sequence, it's three billion letters long. It's this ginormous sequence of letters, but it essentially acts as an instruction manual for the human body, for the makeup. It tells you everything that right with you, but also everything that's wrong with you. And previous AI models before this, they were like, it was like reading a manual with either a magnifying glass where you can see tiny details, but only a few segments at a time versus binoculars where you could see like a lot, but it's kind of blurry. This is able to do both simultaneously. Like you mentioned, it can read

one million letters at once with perfect clarity. So while it's not the three billion, it is one million, which is a significant improvement because when doctors find a weird genetic variant, and they ask, is this variant causing a problem? Alpha genome has beat the the previous models on 25 out of 26 of these tests, which means there's a lot of fewer, we don't know the answer, answers for patients with rare diseases, and it's able to kind of diagnose things that previously were not possible. And as we see progress through this, there's going to be a lot of unbelievable examples. It's released to like cancer and splicing these DNA sequences and improving things. It's just, it's a huge quality of life improvement

for the people who understand it and for the people who stand to benefit from this, which is, I mean, in a way, everyone. We mentioned, on the previous episode, you can upload your DNA sequence to something like Claude and you could ask your questions. This is that on steroids. This is going down to the letter and diagnosing what's wrong and what's right based on a single variance in the three billion letter data set. It's really amazing. For those of you who haven't spent time in many of the scientific journal world, there are tens of thousands to hundreds of thousands of scientific papers published every single year.

But the common pattern across all of these is that some of these conclusions are simply inconclusive. They don't have enough data or they just don't know. They haven't got enough tools or technology to figure it out. Now, a tool like this,

to your point, Josh, I don't think it's going to suddenly have a crazy discovery this year, but you're going to see the discovery in some of these papers that come out from the researchers that are using. In fact, we're probably going to see

Alpha Genome be quoted as a main author produced in 2026. There are already 3,000 researchers using this. I think one million API calls every single day. So it's really been kind of like crushing a bunch of research across 160 countries. So I am curious to see what this gets involved in and what discoveries it makes. To your point, maybe it's cancer or maybe it's something smaller. But one thing is for sure, there's going to be something major this year. Would be my prediction. Yeah. The velocity in progress of this is just an, it's unbelievable. It's really exciting to see. I'm looking forward to all the

continue breakthroughs that scientists are now able to use because this is just open weights, open source freely available for anyone who's involved in research to actually go and use for themselves. There's another interesting breakthrough, or at least release this week, that's slightly less eccentric. This one's more practical and something you can use today, which is Claude releasing interactive tools within Cloud itself. So if you are a user of Slack or Figma or Asana or any of these programs, you actually never need to go to those websites ever again. they will be integrated into your cloud interface on your desktop for you to actually engage with through the chat interface to do whatever you want them to do. So the example we're seeing on

screen now is Figma. If you're a designer, this is how you design things. This is how you create websites, create renderings of things. And built into Cloud Now is tooling control that allows you to type in a prompt or even manually adjust things yourselves from within the interface in a way that's smart. It understands the context and it writes back to the main server. So if you, like for For example, we use Asana for production workflows for scheduling episodes. I don't need to go to asana.com anymore. I just ask Claude, when do we need to publish this by? Who is responsible for publishing this?

And it has all the answers built in, and it can write the changes to that. And this is on the back of the integrations that they've recently added, things like iMessage, which I think is really cool. Previously, or early in this week, we released an episode on Claudebot, which is this whole desktop controller that allows you access to your whole desktop, but it is a bit technical. It does cost a bit of money. this is so easy. If you are a Mac user and you have Clod downloaded on your laptop,

you can connect Claude to iMessage and then have it send messages on your behalf or read messages on your behalf or integrate it into any of these other services. And what they're doing is they're just creating, I think what Open AI was trying to do, which is this full stack place in which you can go for any of your needs. It's just slowly embedding itself into more and more of your day-to-day workflow. To me, what this announcement says is it's confirm that we are heading towards an AI operating system that is ephemeral. Now, we've mentioned this many times on the show before, but I believe in the future you're not going to navigate to an app or a website or scroll.

It's just going to appear ephemerally in front of you, also generated by tokens generated by an LLM or any kind of AI model that is multimodal, right? And the reason that description that you just gave just now, actually, Josh, is, you know, it all kind of converges into one synonymous experience where you could be on different devices or you could be on your browser or you could be on a chatbot and somehow the same LLM or AI model follows you around and is there to be helpful and expressive in many different ways. This is Anthropic confirming it. OpenAIA actually released this feature themselves in chat GPT last year.

Actually, I think it was mid last year. So they were super early with the exact same partners. that uptake didn't really take off. I'm curious whether like six months later we now see a similar pattern with Anthropic or whether people enjoy it and find it more useful here. Another kind of subtle point that this is using underneath is something called MCP Model Context Protocol, which is something that Anthropic, I believe, founded.

It's an open protocol that allows you to connect to a bunch of different apps. Now, the change here is that it's allowing the generative experiences to happen within the chatbot, which is a UX decision, which I think is really cool. Now, I think this is a very strategic move for Anthropic because they created the protocol themselves. So Google famously produced Android and open source the entire thing or rather was like, hey, it's just for everyone. But they ended up becoming the dominant company that control that.

And now they manage like millions and millions and millions of devices. I think this is Anthropics attempt to do the same thing. Now, they're not planning to release a new device, but it is yet another step for them to own the entire operating. stock for whatever AI becomes. Yeah, you said this good word. I actually looked up because I don't know what it means. Ephemoral.

You said ephemeral operating system, which is something that lasts for a very short time is fleeting or transient. Often describes things that are beautiful but temporary, like a moment of feeling. So I guess the idea is that this is, it's dynamically generated, right? Like it's kind of built on the fly. It doesn't actually exist as a permanent fixture in our workspace,

which is, yeah, super interesting. This is cool. There is one use case that I particularly loved, which was the IMessage Rapped. I think one of my favorite parts of the year is Spotify Rapped, when we can see all of the most played music. Everyone talks about how their listening age. It's very fun. Because it has access to these more intimate personal data sets, it can actually use that data to generate these fun things like IMessage Rapped, which is a cloud code project that this woman did, which I thought was really fun. It showed how many messages were sent throughout the year, who the top.

messages were with what it showed heat graphs and charts of when you are most relationship over time yeah you could see relationships over time when a relationship got hot you were texting a lot when it faded away it's this really fun exciting experiment it shows grammar who gets your best English most formal versus most casual your writing style over time and it can really break it down because it has this this deep understanding and intelligence but also because it has this connection to your personal intimate data set. And I thought this was so much fun

because it shows you even the heat chart of when you were most likely to be sending messages. Oh, I like this. The dream dinner party is good. Very cool. Who's sitting at your ideas table? Who your rider dies? Who's the industry dinner?

It's a really fun example of a use case that I hope we see a lot more of. And again, the hardest problem with these is just figuring out how to extract value from it. And whenever we get a cool example like this, it's awesome. It's like, okay, here it's open source of my GitHub. You can go and copy it for yourself

and go have fun. And this is a really awesome use case that I see for Cloud in particular. As we're filming this episode, Josh, and we've given a bunch of tool demos and examples right now, I'm realizing that we have made so much progress over the last year when it was incredibly manual. And now it's much, much, much less manual. But in my opinion, it's still manual enough. Do you know what I mean? And I think there's going to be a dissection of audiences where you have the more kind of, I guess you can call them semi-technical folk that will

spin up a Claudebot or try out a new AI native thing. And to your earlier point, like, you should be proactively trying to use these tools so you can kind of evolve your brain and thinking into using these tools that you're not caught off guard. But I think eventually, majority of people are just going to use some kind of package system. Like Google is like the perfect placement for this, right? Where it's like, hey, don't worry, Gemini's just going to come to wherever you are. And we've got you. We're just going to give you the features instead of you trying to figure out the features. This is an example of someone figuring out the feature. And I'm actually more of an advocate for people to go out and do that. But yeah, it just occurred to me that I think a lot of people are just going to kind of take the lazy option and just wait for that prepackaged thing, which isn't a bad thing because there's less security risk. It's probably going to be cheaper at that point. It's probably going to be a better user experience. Well, it's also like what is the actual delta between the time that early adopters use it versus the time that the general public uses it. And it's becoming increasingly small. It used to be months to years. Now it's a matter of days. I mean, Claudebot was exciting because it had access to your eye message and you could talk with it. And,

now we have a connection and a really cool use case that's built right into your cloud desktop app without any of the technical know-how or installation. And that was a matter of days. The compression is happening very quickly. Yeah. I agree with you on the software front. I maybe don't agree if I look at it from an investment angle. Because to your earlier point, you knew Tesla was going to be a robot company, a humanoid specific company six years ago with that video that you recorded. Right. So it's crazy to see how a lot of these timelines are collapsing for some things, but in other things they're not, or maybe it's just lagging.

I don't know. I think people mostly understand the trajectory. The question is whether they can deliver on that or not. And in the case of Tesla, they have actually delivered on it. In the case of these other companies, they don't have set timelines, they don't have set delivery dates. It's kind of a fuzzy thing,

but everyone is generally working towards the same goals. So what we're seeing is just incremental progress, some faster than others. And XAI right here is another example of incremental progress that perhaps leaves a little, bit to be desired here. The new Grogh Imagine model, EJez, let's talk about this for a second. Okay, I'm going to be honest because I'm a massive Elon Bull. That is no secret on this show. But the video and image models from Grock, Rock Imagine specifically have kind of been subpar.

Like when I look at like what Mid Journey is putting out, what I look at like what V-O-3 is putting out, when I look at like what runway is putting out, there's just a higher quality and definition that GROC just isn't giving me. And that wouldn't even matter because they are, again, like only a two and a half year old startup, but it's the fact that Elon keeps on shilling it like it's the next best thing.

And I'm like, bro, it's not, right? Anyway, the news here is we have a new GROC Imagine API, which is basically their latest image or video model. And as you can see, it is actually quite impressive. And I might be being a bit of a brat here because this is next level compared to like what we had last year. And the cool part is it's cheaper, it's quicker to generate, and it's just simply a better model that is available everywhere.

Whether this inspires me to use it isn't really the point. I'm sure there are people that are involved in video production that will find this way more useful. Well, actually, let me ask you, Josh, as someone that plays around with a lot of this stuff, is this a tool that you would use? Are you impressed by it? No, they didn't need to release this.

This, like, didn't need to happen. They could have just shelved this and kept going with GROC 4.2 and then working on GROC 5. It's, had this been released a year ago, it would have been amazing. It's a great image video generation model. Given today's standards, it's not. And while it's an improvement for the XAI team, it is not an actual improvement pushing the frontier of video or photos forward.

So while I look forward to the next iteration of this, as they work towards their game development engine and building these graphics in real time, this is kind of an incremental progress update that really didn't need to come out. I don't think many people are going to be super excited about this. Well, what if I framed it in a different way, Josh? What if I told you that this was actually never meant for video production, but rather game development. So one of our predictions at the end of last year was that XAI is going to release one of the best gaming AI models, which will allow for real-time game immersion and gameplay, which is kind of like some of the demos that we're seeing on this video right here.

And I'm wondering whether that might be their angle. Maybe they're just not going for video production at all. Yeah, and I suspect that's probably why Elon is abnormally excited about this because he understands the vision for this. and he understands where on the trajectory this incremental data point falls, whereas we're seeing it at face value and being like, eh, this isn't that impressive. We've seen this before.

But gaming is certainly where they want to go. Gaming and animation and graphics and building real-time three-dimensional spaces, which are going to compete with Google's genie, the world building model. So I'm looking forward to that war as it comes along because Logan Kilpatrick from Google is actually teasing Jeannie 3 this week,

which hopefully we'll be getting very, very soon and we'll have a real comparison of a true frontier world building model very soon. In other news, we put out an episode earlier this week, I actually think it's our last episode, on this new Chinese model, Kimi K2.5. Now, we're not going to get into it on this episode, but the TLDR is it is a very impressive model.

It's 100% free, open source. You can download it, amend the model weights, and run it locally, but it might cost you a few very expensive bits of hardware to be able to do that, but you can access it for free on their website. And the coolest part about it is you can kind of like convert video recordings into live production ready apps. But there was an update that we wanted to make based off of this.

Yeah. So we actually said that it would be immensely difficult to run this locally on a machine because it would have required a tremendous amount of compute. And the reality is that someone actually got this running with far less than we thought. You don't need a couple of Gb200s or H-100s or any cutting-edge Nvidia technology. this person actually did it with two Macs on his desktop. So he bought two M3 Ultra Mac studios and strung them together, which costs about $20,000 USD,

and he was able to generate 24 tokens per second using these two computers and the new KimiK2.5 Pro model. So I think that is a testament to how accessible a model this good is and essentially making it free. So now, it's funny, tying it back to the Tesla Autopilot example where you can buy a license currently for $8,000, and that gives you lifetime free miles, or you can do a subscription.

For $20,000 now, you can basically get a lifetime membership to pseudo-AGI, to one of the highest-level frontier AI models, and you could run that for free through infinity at the cost of these two localized models. And I think that's a really impressive and fun breakthrough, where, I mean, assumedly, the server costs to run this would be significantly lower,

and you could just run these open-source models now that are very much at the frontier for a very reasonable cost relative to what some people are paying to use like an API like Opus 4.5, which is charging you $25 per million tokens. Yeah, I mean, the funniest part about this model release has got nothing to do with the model, but the fact that you can never rest easy because your competition is always cooking up something better.

And literally within hours after recording that episode of Kimi K2, Google pretty much announced an identical copy or functionality where they can turn, they call it agentic vision, but they can turn any screen recording into a production ready application. Now, they don't state that specifically on here, but that's technically what it should be able to do. And it just is a testament to how quickly and almost synchronously these AI labs are working together. Now, in my opinion, it's more impressive for the Kimi K2.5 team, the Moonshot Labs team, because they have less infrastructure, Although someone made the point in our episode in the comments, Josh,

I don't know if you saw this, that they have much, much cheaper energy. So I don't know how that translates necessarily. Maybe they just run these servers for way, way longer, and it just cost them less, so it kind of like matches out. But it's interesting that the open source teams and the centralized teams are kind of like moving at parity right now. It's cool that the open source teams have even caught up. But yeah, Google's come up with the same thing if you want to try it and use a Western AI

Lab product. It's amazing how quick they were to respond. And also that comment is very correct. Chinese electricity costs significantly less than the U.S. electricity does because they have so much more of it. So the cost per token, the cost to train these models is significantly low because of that cost per kilowatt as it relates to just generating all of this energy. Yeah, yeah. But they had really nice reply from the commenter, but also Google in a matter of seemingly days. It feels like they had this ready to go.

And they were like, oh, all right, well, I guess we should probably just release this. now because Kimmy did something similar. And now here we are. Agenic Vision for all. Well, if Moonshot AI, Kimmy K2.5, is the model example of spending your resources widely, open AI probably needs to be on the opposite end of that spectrum with the breaking news that they are raising a hundred billion dollar round, which is just Gargantuan, to value them at around $730 billion. I've seen $8.50 as well. well. I can only imagine that they are raising this amount of money to blow it all on training various different models, not just one model, or to kind of build out a new social network

that we saw on the timeline. Apparently they're building that and their hardware device. And what else am I missing, Josh? We've got the SORA TikTok app competitor. It just seems like at this point, and I know this might be a negative take, they've been raising so much money and endlessly blowing it on myriad different things. So they're spreading themselves very thin. This doesn't signal confidence to me, actually. It actually signals that they are kind of being reckless at this point. Yeah, it's probably helpful to look at past fundraising history to get a better perspective of how much money this actually is in a way that feels almost existential. There's so many vested interests in Open AI succeeding. If they don't, what does that look like for the rest of

the industry? That's an important question to just kind of sit on for a little bit. But if you go back to the funding rounds, even April 2023, they were raising still in the millions. They raised $300 million at a $27 billion valuation. Then in October 24, they raised 20 times that, $6.6 billion at $157 billion evaluation. And now they're raising eight times that at $40 billion at a $300 billion post money valuation. Now, this is interesting because EJAS earlier in the episode, we mentioned SpaceX was looking to raise $50 billion at about a $1.5 trillion. valuation. So they're raising 20% more at 500% the market cap. And we're starting to see, like, Open Eye is giving up a lot of equity here. They're really giving a lot of vested and controlled

interest to a lot of companies that may not want that liability. And I wonder if this is a strategy in where he's trying to, Sam and the Open AI team are just trying to get as many vested interest as possible in their success in a way that they become too big to fail, where so many people have so much money tied up in this thing that they own a lot of, that they are forced to give whatever it takes to make this work because the downside effect impacts the entire industry in such a large way. And I can't help but read these headlines and not think that, I mean, this is just outrageous. They have a tremendous amount of debt. They haven't made a single profit. They're trying all these things like ads. They're trying to take revenue cuts. They're trying

social media things, seeing what sticks, and it's giving desperate vibes. Yeah, I mean, I have a simple take, which is they are on track to make and also burn $20 billion of KAPX this year. I think they need someone to foot the bill to also spin up a bunch of the compute data centers, $1.4 trillion worth over the next three years. And I think that panicking that they didn't spend enough time and resources on building out a better coding model to compete with Anthropic, which is kind of taking over all of enterprise and eating out their market share right now. And needlessly spending it on other stuff like social media apps and stuff like that. So my take is, listen, I hope they raise the round. Good for them. But I don't

trust Sam and the team to spend it wisely for now. And that might be super bearish take. I'm willing to be proven wrong. But that's just where my head's out. Well, I feel like we have to root for them because we have no choice. If open AI goes down, so does everything with it. So we're rooting for you Sam and the Open AI team, I hope everything works. I hope this money goes to good use. And I think that probably wraps up our roundup for this week. Of all the top news in AI, as you enter the weekend, you can now feel properly satiated, that you are fully up to date and aware of all the hottest news that you needed to this week. We released a few pretty great episodes. They're doing really well on Kimmy K2 and Cloudbot. So if you haven't had a chance to go listen to those,

please go and check them out. The ask for you, perhaps, this week, is to try out the Gemini Chrome extension situation and use it to subscribe to our podcast. That's like a fun little demo. You can say, hey, subscribe to my favorite new AI in Frontier Technology podcast. And it will actually go and click the button for you. And then if you go on Claude and you connect your eye message, you could ask it, send the limitless podcast to 10 of my best friends and figure out which are my 10 best friends who would be most interested and share the link with them. And I think these are really great use cases that you can try in order to demo this new tech. A number of other Google accounts and do the exact same thing, right, Josh? Let's just create a,

entire platform. Exactly. But I mean, this is a bit of a longer episode. So if you made it this long, thank you so much. These are the real ones that are still here at the end. Thank you for listening. We will be back again next week with plenty new episodes, lots of new news to cover, and a lot of exciting topics that we will keep on keeping you up to date with. So thank you for watching and we'll see you guys next week. See you guys.