Limitless: An AI Podcast

Josh:
The CEO of Google just got on stage yesterday and told everyone that there's no new Frontier model.

Josh:
In fact, he said they've only made it 90% of the way there. So on the surface,

Josh:
this seems fairly disappointing given the fact that it's been many months since

Josh:
our last model release and Gemini was doing pretty great.

Josh:
So while he didn't claim they had the lead, they did proceed to release new

Josh:
models in nearly every single product we use every day and in some pretty creative ways.

Josh:
I mean, three of the highlights are a brand new model that's pretty prevalent

Josh:
throughout all the software we use, an OpenClaw competitor that is built right

Josh:
into Google, and this new model named Omni.

Josh:
And that's where we're going to start today, because Omni is nothing like we've

Josh:
ever seen in the world of AI before.

Josh:
Omni is a pretty impressive model, and I think probably the highlight of this

Josh:
event. So, Ijaz, what's going on with Omni?

Ejaaz:
So they had this really cool moment during IO where they just started showing

Ejaaz:
this announcement video.

Ejaaz:
And people were kind of looking at this and were like, oh, some of this looks

Ejaaz:
AI generated, but some of this actually looks quite realistic.

Ejaaz:
And the reason was they released this brand new model called Gemini Omni.

Ejaaz:
Now, it's a very unique model. It's

Ejaaz:
not something that we have seen before because it's a hybrid of an LLM,

Ejaaz:
and a world model. And world models, if you've heard that term before,

Ejaaz:
is really unique because it can produce video that not only looks realistic,

Ejaaz:
but is physically accurate.

Ejaaz:
So something that LLMs struggle with, it's like that analogy we've used before,

Ejaaz:
Josh, where you have a student in a library that's read all of the books and

Ejaaz:
understands the world, but actually has never stepped outside of the library

Ejaaz:
and experienced the world for himself.

Ejaaz:
World models are basically an upgrade from LLMs where it can teach the LLM to

Ejaaz:
understand how the world works, the consequence of an action.

Ejaaz:
So say if you like hit something or if you jump how gravity works and all those kinds of things.

Ejaaz:
So the unique thing about this model is it can take any input medium.

Ejaaz:
So we're talking about words, we're talking about images, we're talking about

Ejaaz:
video itself and it spits out only one output, which is video.

Ejaaz:
Now the videos that you're seeing on screen here is physically accurate.

Ejaaz:
Now, obviously, it's generated by AI, but the vortex that you're seeing on your

Ejaaz:
screen now, the way that the person draws a circle,

Ejaaz:
the way that the paper kind of crinkles is basically all meant to be physically accurate.

Ejaaz:
And the reason why this is such a powerful model is, well, there's a few.

Ejaaz:
Number one, when you're learning about something, the number one medium that

Ejaaz:
people or humans ingest information is through video. Words is actually kind

Ejaaz:
of slow. It takes a while to kind of like go from left to right.

Ejaaz:
People don't like read all the time in today's world, but they do watch videos.

Ejaaz:
There's several different sensory instruments that come in. There's sound, there's like visuals.

Ejaaz:
You kind of like ingest data way quicker with video, and that's the bet that

Ejaaz:
Google was making with this.

Ejaaz:
And the second thing I'm going to say is, this is pretty world-changing for

Ejaaz:
a lot of different applications outside of just LLM.

Ejaaz:
So, for example, Google has been a massive proponent of robotics.

Ejaaz:
The issue with robotics is, these robots don't understand the real world.

Ejaaz:
And so far, all the robot models have been fed with words about how the real world works.

Ejaaz:
Well, what if you could place that robot brain into a simulation that is physically

Ejaaz:
accurate, like a world model, And you can start teaching people or robots specifically

Ejaaz:
how to interact and they can train and learn a lot faster.

Ejaaz:
So I just think there's many different applications for what this world model is.

Ejaaz:
And it's the first mass scale deployed one. So I'm excited to get my hands on this personally.

Josh:
Yeah, and it's available today live for basically anyone who uses Gemini.

Josh:
It's available in the Gemini app, and it's pretty cool. One of the things that

Josh:
I've enjoyed personally is that you can create an avatar of yourself.

Josh:
So you can actually build a digital clone of yourself and inject it into these.

Josh:
You could think of this model sort of like Nano Banana, but for video,

Josh:
where for the first time ever, you can really create fully custom videos that

Josh:
either take base reality and build something on top of it, or just create a

Josh:
total AI demo in general.

Josh:
So we have this pretty cool demo from the CEO of DeepMind, Demis Hissabas,

Josh:
which show you kind of how this works, where it takes his actual video that

Josh:
was recorded on his camera roll and then turns it into five different scenes

Josh:
that are all completely unique, seeming somewhat realistic.

Josh:
And I think this combines the fidelity of something like Nano Banana with the

Josh:
complexity and physics understanding of VO3. And that's where it really shines.

Josh:
What's also really fun is you have the ability to refine these things too.

Josh:
So you can actually tweak the model, tweak the outputs.

Josh:
And the result of this is something fairly comprehensive, particularly if you're

Josh:
trying to learn things. I find that education is probably one of the most interesting

Josh:
demos and use cases of this.

Josh:
And we have one, which is prompted from a single question of how does photosynthesis work?

Josh:
And now as a student of anything, as someone who's trying to learn,

Josh:
you're given this fairly comprehensive answer that starts with an image.

Josh:
The image is generated by Nano Banana, and then it gives you the chemical recipe.

Josh:
And then under this how it unfolds section, we actually get a fully generated

Josh:
video from the new Omni model that shows you a like actual visual representation

Josh:
full of graphs and charts and seemingly realistic

Josh:
version and an easy way to learn how this works.

Josh:
So I think this is probably the one of the most powerful use cases if you're

Josh:
a teacher, if you're a student, if you're someone who's just curious about any

Josh:
topic and your preferred way of learning is visually, this is a really compelling

Josh:
product for you. Now this also works...

Josh:
With real videos that you could just upload to Gemini. We did this this morning.

Josh:
I had this video on my camera roll that was just filmed outside of a window.

Josh:
And the prompt was just, hey, make some meteors fall down from the sky and put

Josh:
the Loctis monster in the East River.

Josh:
And the reality was this output, which looks pretty good.

Josh:
I mean, if you're looking at this, you'll notice their sound.

Josh:
You can hear the ambient sound of the city.

Josh:
There's a large splash when the dinosaur or the loctus monster hits the water.

Josh:
It looks pretty good, but also this is very clearly AI generated.

Josh:
It feels like CGI from perhaps the early 2000s. So not quite what we would expect

Josh:
in 2026, but very cool nonetheless that it can actually take one of your videos

Josh:
and generate something pretty powerful on top of it.

Ejaaz:
So if you listen to this, you're probably wondering, like, how can I use this?

Ejaaz:
Why is this applicable to me?

Ejaaz:
Well, if you're just a regular person that kind of wants to play around with

Ejaaz:
video, So this is plugged or integrated right into YouTube Shorts.

Ejaaz:
About two months ago, Google made a very important decision,

Ejaaz:
which is integrating AI into their major video platform, YouTube.

Ejaaz:
A lot of people had mixed feelings about this. Artists in general in the creative

Ejaaz:
world tend to hate on a lot of different AI tools.

Ejaaz:
Google takes a different approach where they're saying the kind of like proactive

Ejaaz:
efforts that a lot of artists are making.

Ejaaz:
If they can produce more content that is applicable to various different niches

Ejaaz:
of audiences, then you can kind of like create a sustainable platform.

Ejaaz:
They see the number one distribution for this being YouTube Shorts.

Ejaaz:
So this is plugged into YouTube Shorts right now.

Ejaaz:
You can go out there, create an eight second video and show us what you've got

Ejaaz:
if you actually listen to this and you have an account, like we would love to see it.

Ejaaz:
I will say that these things do look quite janky right now, as you mentioned,

Ejaaz:
Josh, but this is also V1.

Ejaaz:
And it's important to stress that world models up until now weren't really applicable

Ejaaz:
or distributed to a mass audience.

Ejaaz:
If you remember, Google released, I think it was Genie 3, which is like their

Ejaaz:
really powerful world model.

Ejaaz:
That is what powers this model that you're seeing on your screen here today.

Ejaaz:
But it's not a full implementation. Why?

Ejaaz:
Because it's actually quite expensive to serve.

Ejaaz:
It requires a different type of infrastructure to train.

Ejaaz:
And so once that kind of scales up with Google, they'll be the first distributors of this.

Ejaaz:
And I'm excited to see more of this because Demis at Davos earlier this year

Ejaaz:
basically said that he thinks world models is the future evolution of LLMs. He thinks that words are.

Ejaaz:
Ultimately constraining for AI models. And in the future, if you have a model

Ejaaz:
that can speak, but also understand how the world works, that will be the all-consuming model.

Ejaaz:
Now, of course, that's in the future.

Ejaaz:
For now, we have to focus on the major frontrunners of AI, which is Anthropic

Ejaaz:
and OpenAI, which have the world's leading LLM.

Ejaaz:
So the question becomes, did Google release an amazing LLM.

Ejaaz:
Although they did release an LLM, it's called Gemini 3.5 Flash.

Ejaaz:
And to quote Logan over here, it's their most powerful model to date.

Ejaaz:
Now, it pushes the frontier in many different ways. We've got a chart here which

Ejaaz:
kind of like compares Gemini 3.5 Flash to the older Gemini Flash models,

Ejaaz:
as well as Opus 4.7 from Anthropic and GPT 5.5 from OpenAI.

Ejaaz:
And you'll notice that Across coding in general, it doesn't trump or beat GPT

Ejaaz:
5.5 or Opus 4.7, which I guess some might say is a little disappointing, but also unsurprising.

Ejaaz:
Google has fallen behind in coding AI for a while. So it's not as good,

Ejaaz:
but it's kind of near the marker of them.

Ejaaz:
But if you look just below here, this is where this bottle actually excels.

Ejaaz:
It's known as agentic tooling or agentic AI in general.

Ejaaz:
What Google has been able to achieve with 3.5 Flash is the ability to spin up

Ejaaz:
multiple sub-agents that work in parallel to get your prompt resolved.

Ejaaz:
So if you think that like having one singular brain to do all the work for you

Ejaaz:
really, really well, which is what Claude Opus 4.7 does, which is what GPT 5.5

Ejaaz:
does, Google took a different approach and said, let me have multiple brains,

Ejaaz:
maybe even hundreds of brains working on the same problem and collectively,

Ejaaz:
they can produce a better answer.

Ejaaz:
And that's what they've achieved with the Gentic tooling here where you've got

Ejaaz:
the tool-a-thon and the MCPP Atlas benchmarks, which absolutely crushed GPT 5.5 and called Opus 4.7.

Ejaaz:
So you're probably wondering, okay, well, what tasks can I use this for?

Ejaaz:
It's typically long horizon tasks, like long tasks that you can kind of shut

Ejaaz:
your laptop down and kind of let it work away for like 10 to 20 hours this is the model for you.

Josh:
So you'll note that they said that it is their best model

Josh:
ever not the best model ever and that

Josh:
it is fast but it is not the fastest and that

Josh:
it is affordable but it is not the most affordable and when

Josh:
you're kind of observing the parameters on the Pareto Frontier the things

Josh:
that we actually measure to evaluate a model it's not

Josh:
really the best at anything and in fact

Josh:
we have this kind of comparison that I saw with Cerebris which

Josh:
I really enjoyed and Cerebris posts this kind of trolling them which

Josh:
is Cerebris versus Gemini flash and it very

Josh:
clearly shows where cerebris sits versus flash

Josh:
you could see 3.5 flashes out pretty fast in terms

Josh:
of the tokens per second output but when you're trying to

Josh:
optimize for one of these things along the frontier clearly cerebris has the

Josh:
speed thing completed so it's not the fastest by a large margin it's not the

Josh:
smartest by a pretty large margin it sits kind of in the middle and i think

Josh:
this is the most interesting part of this story today is that a lot of these

Josh:
models are not the best but they are unique in their own way one of the cool demos that

Josh:
I liked that they had mentioned during the presentation is that they actually

Josh:
took this new Gemini 3.5 flash model to build an operating system from scratch

Josh:
over the course of a series of hours.

Josh:
I mean, that's pretty incredible. It says it took 12 hours, 93 parallel agents,

Josh:
15,000 plus model requests, and then $2.6 billion in tokens process.

Josh:
So you are essentially able to build an entirely new operating system from scratch

Josh:
for less than $1,000 in API credits. That is pretty phenomenal.

Ejaaz:
That's actually super hard to do. Do you remember we, I think, who did it?

Ejaaz:
I think it might've been a GPT 5.4. It was, it was GPT 5.4.

Ejaaz:
And Sam posted this video of a team building an OS from scratch.

Ejaaz:
And that took them, I believe, 36 hours, like over a day.

Ejaaz:
Like in that time, so I think that was like two and a half months ago.

Ejaaz:
You now have Gemini Flash 3.5, which isn't as good a model as GPT 5.5.

Ejaaz:
Do this in a fraction of the time by creating this novel approach of spinning

Ejaaz:
up almost 100 agents to do that.

Ejaaz:
Josh, I had a comment on the speed thing here because you're right.

Ejaaz:
Cerebrus that runs GPT 5.5 versus Google is not just a question of speed.

Ejaaz:
It's a question of infrastructure.

Ejaaz:
Now, I have a demo here which shows actually how quick Gemini Flash can be.

Ejaaz:
And to be honest, I just want to start this from the start. Look at this.

Ejaaz:
That's crazy my brain can't ingest that

Ejaaz:
much information in real time and i think although

Ejaaz:
it's not as fast as some of these other new chip architectures it's still good

Ejaaz:
enough for the majority of humans for sure right the other thing i'll say is

Ejaaz:
this is built off of google's latest gpu infrastructure known as the tpu v8

Ejaaz:
specifically the 8i which stands for eight inference it's inferencing chip,

Ejaaz:
which basically is specialized around

Ejaaz:
giving you quicker responses to your prompts. Now, that being said...

Ejaaz:
This is incredibly expensive. Demis Hassabis dropped some stats here where he

Ejaaz:
kind of shows off the speed.

Ejaaz:
He says it's 4x faster than any of the other Frontier models.

Ejaaz:
And actually, if you use our IDE environment, which is basically Google's equivalent

Ejaaz:
of cursor, it performs 12x faster.

Ejaaz:
What he doesn't mention is it's four times more expensive than all the other Frontier models.

Ejaaz:
So there's a trade-off here, right? I think you mentioned it before we started recording, Josh.

Ejaaz:
There's a trade-off. If speed is something that you necessarily need,

Ejaaz:
which you could very well want in financial services where you're like trading

Ejaaz:
algos for whatever, speed becomes a necessity.

Ejaaz:
You're willing to pay that premium because you end up making more money on the top end.

Ejaaz:
That's completely fine. But for the majority of users where speed doesn't necessarily

Ejaaz:
matter that much, but thought and intelligence matters more,

Ejaaz:
you want to make sure that that cost per watt per intelligence or whatever that unit might be,

Ejaaz:
is kind of like kept at a certain bracket. Like I watched an episode where they

Ejaaz:
interviewed the CFO of Anthropic and he mentioned that these enterprises are

Ejaaz:
spending like 100%, sorry,

Ejaaz:
500% more than the initial budget that they had at the start of the year by the end of the year.

Ejaaz:
So the point is like people are spending a lot of money on this.

Ejaaz:
And at some point we need to kind of like reduce budgets to an extent.

Ejaaz:
So yeah, it's quite expensive.

Josh:
All right, who's an OpenClaw fan who's listening? Or who has given up on OpenClaw?

Josh:
Because if you are a current OpenClaw user, you're probably going to hate this.

Josh:
But if you've never used it before, this is a pretty compelling product.

Josh:
And it's called Gemini Spark.

Josh:
This is the third major announcement that they made.

Josh:
And Gemini Spark is fascinating because it is that 24-7 personal agent.

Josh:
They're claiming that it helps you navigate your digital life,

Josh:
taking actions on your behalf and under your direction.

Josh:
Essentially, this is Google's answer to OpenClaw. This is their version of the 24-7 agent.

Josh:
And the reason why I said you may hate this is because you are beholden to Gemini 3.5.

Josh:
Now, one of the best things about OpenClaw or if you use Hermes is it's model

Josh:
agnostic. you could choose models that you want to use, and it will go and tackle specific tasks.

Josh:
In this case, on Gemini Spark, you're stuck with 3.5.

Josh:
So is this good? Is this bad? We'll see, but I find it really interesting because

Josh:
this just runs directly in your browser.

Josh:
So you don't actually need to have a laptop open. You don't need to run this on a remote server.

Josh:
All you need to do is log into your Google account, set up a Gemini Spark instance,

Josh:
and then it will just run permanently in the cloud.

Josh:
A lot of times you'll see there's memes going around to people who are walking

Josh:
around with their laptops cracked open because they're running these agents

Josh:
that are running long-term tasks.

Josh:
This is basically your own little private server to go off and build anything that you want.

Josh:
And it's pretty interesting. I think for people who are less technical or for

Josh:
people who just don't love using OpenClaw, this is a pretty easy way of accessing

Josh:
a 24-7 server to just run commands, to run prompts over very long durations of time.

Ejaaz:
I like this and I hate this. The reason why I like this is I use a lot of Google

Ejaaz:
products to do work, Google Docs, Sheets,

Ejaaz:
Gmail. So the fact that I can have an OpenClaw that is seamlessly integrated

Ejaaz:
into all of those things gives me hope.

Ejaaz:
The other side of this is I just don't think it's going to be as good a quality

Ejaaz:
product as OpenClaw or Clawed Cowork.

Ejaaz:
In fact, there was news that broke, what was it, like a couple of weeks ago,

Ejaaz:
that revealed that majority of Google employees were using Clawed to do their work, right?

Ejaaz:
And then Google kind of like put a ban on this. And the reason why they were

Ejaaz:
doing this was presumably to learn how to build a better agentic or coding model

Ejaaz:
that can do exactly what we're seeing on the screen here.

Ejaaz:
So they're late. They're lagging the front runners. I'm glad they put something

Ejaaz:
out there. Google Desk has the distribution.

Ejaaz:
They do have the data mode. They do have all the users.

Ejaaz:
For now. And so I think they have a shot at creating something good here.

Ejaaz:
But I get meta vibes from this, Josh, like the same reason why like meta might

Ejaaz:
have like a good model with Muse Spark.

Ejaaz:
I'm not using it because it's not on the platforms that I actually want to use the thing.

Josh:
Yeah, and I think that's one of the trends of all the announcers that we're talking about today.

Josh:
They are all kind of existing in seemingly different places.

Josh:
It's not clear and obvious which features I'll want to use. We're

Josh:
about to rattle off maybe five or six more things that they've announced

Josh:
and i think one of the themes as i i mean this is

Josh:
a common complaint that we've had with google or at least me personally is

Josh:
i'm not really sure the best way to use all of these

Josh:
tools like with anthropic i go to the claw desktop app

Josh:
and it has everything in one roof and i know exactly how it all works with open

Josh:
ai the same thing is true and i even have a companion mobile app with google

Josh:
there's a lot of this kind of ambient ai that exists at the seemingly random

Josh:
series of touch points like it exists in my gmail but i kind of hate the ai

Josh:
written emails that it makes.

Josh:
It exists in my Google Drive, but I don't really need it there.

Josh:
And then you have the Gemini app, which is getting better, but it still doesn't

Josh:
include a lot of the functionality that we're mentioning today.

Josh:
I mean, the next thing we could talk about here is their search,

Josh:
because for the first time in over 25 years, they're really changing how search is going to work.

Josh:
And again, it's just kind of like this ambient AI that you don't really seek out.

Josh:
I'm not really sure where to find all that a little confusing.

Josh:
But in terms of the AI search, it's very different.

Josh:
Now Google search has gemini 3.5 flash

Josh:
baked into it by default so when

Josh:
you search for something it will be routed through gemini 3.5

Josh:
instead of the traditional google search engine and this is a pretty profound shift

Josh:
in just the way the internet works the way people get information on the internet

Josh:
the first person i think about is the sites that are kind of optimizing for

Josh:
seo on google now it seems like the whole meta has shifted to seo for ai models

Josh:
because now when you search something the reality is is that you're going to

Josh:
be routed through an ai model and not the

Josh:
indexing algorithm, at least for now.

Ejaaz:
So one thing a lot of critics on Wall Street said about Google three months

Ejaaz:
ago was that they're worried about AI eating their dominant market share in search.

Ejaaz:
And then they released their Q1 earnings about a week ago, and it revealed that

Ejaaz:
not only did it not eat their search market share, it increased it.

Ejaaz:
They ended up earning more money. And I think that's because,

Ejaaz:
Google Search has been integrating AI for a while now. This isn't the first

Ejaaz:
time that I type something in Google and it gives me an AI-generated prompt.

Ejaaz:
I love that we have the latest model here and that they're making it more native into the homepage.

Ejaaz:
I think this is amazing. So this is my second favorite release from Google, from I.O.

Ejaaz:
Hundreds of millions of people, probably even billions of people,

Ejaaz:
use Google every single day. They ask a bunch of different random questions.

Ejaaz:
I would actually argue that it has more reach than your ChatGPT home interface

Ejaaz:
or Claude home interface.

Ejaaz:
And if Google is able to successfully pivot this into something that serves

Ejaaz:
up their latest LLM prompts and somehow get people to use or pay for it,

Ejaaz:
then this could end up being a home run.

Ejaaz:
Of course, as you mentioned earlier, like Google is like free at this point,

Ejaaz:
right? The whole point is like distribution is the mode.

Ejaaz:
And so I think they're going to end up making most of their money from this

Ejaaz:
particular feature through ads.

Ejaaz:
Now, it's important to say that no AI company right now has figured out the

Ejaaz:
SEO strategy. In fact, the leader of this is Meta.

Ejaaz:
Who has been embedding a bunch of AI agents and models into their ad services

Ejaaz:
on the backend to help advertisers reach more of a wider audience.

Ejaaz:
If anyone knows the data of their users, the best is definitely Meta.

Ejaaz:
They've been training on this for a while.

Ejaaz:
Their new model, MuseSpark, does exactly this. I have a feeling Google's going

Ejaaz:
to go down the same route, but they're going to be a little more careful about

Ejaaz:
this because they are the dominant market platform for advertising.

Ejaaz:
So I'm excited to see where this goes from here.

Ejaaz:
And then there was this other post, which you were commenting on,

Ejaaz:
Josh, which is basically like, you get like such good value for money with a google ai pro bundle.

Josh:
This is my new favorite subscription for twenty dollars it comes

Josh:
with youtube premium light which doesn't include music videos but basically

Josh:
an ad-free experience on youtube in addition to five terabytes of storage you

Josh:
get access to all the new gemini pro models it has image video generation access

Josh:
anti-gravity notebook lm pro which is really good for deep research and long

Josh:
context stuff It's a really compelling bundle at $20.

Josh:
And I think this is one of the most understated

Josh:
things that was released yesterday,

Josh:
because this is what I would probably recommend most people use.

Josh:
It gives you access to the entire suite. And even if you just watch YouTube

Josh:
videos, I mean, you get access to the ad-free version of that.

Josh:
And that alone makes it worth it with the five terabytes of storage.

Josh:
So very compelling value coming out of them.

Josh:
And speaking of YouTube, they also did the same search thing that they did with

Josh:
Google on YouTube as well, where now there's AI search baked right into your search results.

Josh:
It turns out that Gemini has been trained on lots of youtube

Josh:
data and actually has context on the types of

Josh:
videos that you want to watch so

Josh:
if you ask a question about a specific thing that you

Josh:
want troubleshot it will not only find the right videos for

Josh:
you but it will find the correct moments in the videos and

Josh:
serve those up to you and i find that we've started to experience this in google

Josh:
search recently where like i can type something it'll kind of time stamp me

Josh:
into a video now it's fully baked into the youtube experience and i think that's

Josh:
going to be a really compelling product as well so google in terms of upgrading

Josh:
their core products that everyone uses on a day-to-day basis looking really strong.

Ejaaz:
For this final section, Josh, I want us to zoom out because the question that

Ejaaz:
I have on my mind is Google has a decision to make on which strategy they want to pursue for AI.

Ejaaz:
Right now, OpenAI has made it very clear they want to catch up with Anthropic

Ejaaz:
on coding. And in fact, you could argue that they have done that.

Ejaaz:
Anthropic focused on coding for the longest time. And the reason why both of

Ejaaz:
those companies are doing that is they believe if they nail or coding AI,

Ejaaz:
then they can build everything else.

Ejaaz:
Everything else becomes downstream of coding AI, right? There's an argument there.

Ejaaz:
Now, Google, with this latest launch or series of launches.

Ejaaz:
Hasn't really made that explicit statement as well. In fact,

Ejaaz:
I think they're making a few other statements.

Ejaaz:
Number one, they're betting that video or Omni models are going to be the future.

Ejaaz:
That's why they spent a lot of compute and time invested in their new Gemini

Ejaaz:
Omni intelligence model.

Ejaaz:
Number two, I think that Google is realizing that the agent harness or the harness

Ejaaz:
that goes around And an AI model is equally as important as the fundamental

Ejaaz:
foundation model itself. Now, I have proof of this.

Ejaaz:
There was this tweet that I saw yesterday, which on Cursor, you can measure

Ejaaz:
the intelligence of different models across different tasks.

Ejaaz:
And it showed that the recent Gemini 3.5 flash was worse than Cursor's own foundational

Ejaaz:
model, the old version of it, Composer 2.0.

Ejaaz:
And yesterday, Cursor released 2.5, which is on par, you can see on the top

Ejaaz:
here, with the other models, the Frontier models.

Ejaaz:
So the question I have is, like, Gemini kind of missed here.

Ejaaz:
Like, they should be focusing on the agentic side. As I mentioned earlier,

Ejaaz:
3.5 Flash is great at agentic tooling, but what they've missed here is the harness.

Ejaaz:
The harness is basically the wrapper, the muscle memory that goes around your

Ejaaz:
model that orchestrates the model in a really ideal way.

Ejaaz:
What OpenAI, what Anthropica realized is that that is super important.

Ejaaz:
And Elon Musk has figured that out, and that's why SpaceX is looking to acquire

Ejaaz:
or moving to acquire Cursor for $60 billion.

Ejaaz:
I think Google missed here, and what I come away feeling about the series of I.O.

Ejaaz:
Launches in general is, it's good, but it's not great enough,

Ejaaz:
and if you're Google, which was.

Ejaaz:
Two weeks ago, the most valuable company in the entire world,

Ejaaz:
you need to be making bigger bets. And it's a shame that they haven't made it.

Ejaaz:
That's that's currently how I feel.

Josh:
Yeah, when you compare the prices to I mean, it's four times the cost of composer.

Josh:
So it's like on a price basis is losing on a score basis is losing.

Josh:
I will say this is 3.5 flesh, not 3.5 pro, which we can expect to be released next month.

Josh:
So that might change things and shake everything up in a pretty big way.

Josh:
It's just not currently available. So we can't judge it we can't be

Josh:
critical of it there is one final announcement i just want to

Josh:
squeeze in because as a hardware guy i get excited about anytime

Josh:
someone releases hardware google has some new hardware offerings it

Josh:
comes in the form of glasses they are able to see the world and interpret it

Josh:
some of the examples that were used that i found pretty cool was um she could

Josh:
take a selfie the person who was giving a demo of the crowd and then prompt

Josh:
the glasses to not only capture the image but create a variation of it using Nano Banana Pro.

Josh:
It gives directions. It is agentic in the sense that it can work with your phone,

Josh:
order you coffee on the way to work.

Josh:
It's pretty cool, pretty interesting. I'm very excited to see more about glasses.

Josh:
They actually announced that they're planning to add a HUD, a heads-up display,

Josh:
basically a visual display into the glasses as soon as next year, which is exciting.

Josh:
So Google seems to be stepping pretty seriously into the hardware market.

Josh:
I think the current versions are kind of a joke, similar to all the others,

Josh:
but the ones with the heads-up display are going to be very exciting next year

Josh:
very much could be the year of ai glasses but with that

Josh:
I think we got it covered. Is that IO covered? Is that all of it?

Ejaaz:
I think that's just about everything. There's a bunch of other stuff,

Ejaaz:
but it's kind of tertiary.

Ejaaz:
There's some YouTube upgrades and stuff like that. But I think we've covered

Ejaaz:
all the hot topics, Josh.

Ejaaz:
I think I come away from this episode thinking Google is not out of the race,

Ejaaz:
but I do think that they need to double down on coding specifically.

Ejaaz:
And the crazy part is like they have the moat. They're currently in the number one position.

Ejaaz:
The only other company that could compete with them is Apple.

Ejaaz:
And Apple has been doing nothing.

Ejaaz:
They're getting a CEO swap. And then I think later this year with John Turner's

Ejaaz:
coming in and then I think we're going to see like them step up.

Ejaaz:
But the truth is, it's them versus these two startups, which.

Ejaaz:
Are aggressively increasing ARR by the month.

Ejaaz:
I don't know if you saw this, Josh, but someone projected out Anthropik's rate

Ejaaz:
of revenue increase, which means that by 2028, they would have surpassed Google.

Ejaaz:
So Anthropik's projection for this year was 10 billion ARR.

Ejaaz:
They are about to hit 45 billion ARR this month.

Ejaaz:
And if that projection or if that rate of increase keeps happening and they

Ejaaz:
start consuming more enterprise customers, etc, etc,

Ejaaz:
you can feasibly say that they will become as large as Google,

Ejaaz:
which then starts to justify all these crazy, you know, $1.1 trillion valuations

Ejaaz:
or the secondary round that they're raising right now for $1 trillion.

Ejaaz:
It kind of makes sense. So I think Google still has a shot.

Ejaaz:
This wasn't the killing blow, but maybe next time, maybe when Flash,

Ejaaz:
not Flash Pro, but 3.5 Pro comes out.

Josh:
Yeah, well, if I could give any feedback on the presentation,

Josh:
it would be to focus. It feels like Google doesn't really have any focus there.

Josh:
It felt like I was watching 15 separate presentations instead of one coherent one.

Josh:
And therefore, as someone who is an AI fan who literally follows this for a

Josh:
full-time job, I was unable to keep up and orient myself on where all these

Josh:
products fit into the ecosystem.

Josh:
So if I could leave one criticism for Google, it's just to focus.

Josh:
When I think about Anthropics offerings, I have a very clear picture in my head

Josh:
of what they're good at and how I want to use it.

Josh:
Same is true with OpenAI. With Google, that's not the case and i

Josh:
think if they can focus if they can compress all these

Josh:
releases down into a few key things that would meaningfully shift

Josh:
the way that people think about gemini and engage with the products and i hope

Josh:
that for their next releases their next presentations will get a little bit

Josh:
more of that focus but that is the episode for today that is google io you are

Josh:
now fully up to date and informed on all the things gemini all things google

Josh:
any final parting thoughts you just before we head out of here

Ejaaz:
Please spin out a really good version of Google Glass. That's one.

Josh:
Oh, man, I love that.

Ejaaz:
I really want a good hardware device. I'm like begging for it.

Ejaaz:
Josh and I saw another leak last week. We didn't mention this on last week's

Ejaaz:
episode, but it was those weird OpenAI over earbuds with the weird disk thing.

Ejaaz:
And I was like, I need that device. I just need something.

Ejaaz:
I realized I need hardware in an ever-increasing digital world.

Ejaaz:
So Google Glass, I almost think it's good if they would just spin it out and

Ejaaz:
make it a cool consumer company. But yeah.

Josh:
Well, you know, no one wants hardware more than I do. I'm dying for a good pair

Josh:
of glasses, just anything physical that I could touch.

Josh:
That's not my iPhone. That's not made by Apple.

Josh:
But alas, no one has completed that yet. So we will keep our eyes peeled.

Josh:
We will continue to watch. If you enjoyed this video, please don't forget to

Josh:
share it with a friend who might also have enjoyed it.

Ejaaz:
Don't forget to subscribe to the channel. Loads of new subscribers.

Josh:
Josh. Leave a comment. Yeah, we had our biggest episode ever.

Josh:
Drop on Monday, the Leopold Ashenbrenner portfolio, the new 13f filing so if

Josh:
you haven't seen that i would highly advise we had a lot of new people joining

Josh:
so for the new people welcome we publish new episodes four times a week about

Josh:
20 minutes sometimes we go a little bit longer when we get excited but there

Josh:
will be many more to come this week so thank you guys so much for watching this one and we'll see

Ejaaz:
You next time see you guys.