Today on the AI Daily Brief, Gemini3 has officially arrived.
The AI Daily Brief is a daily podcast and video about the most important news and discussions
in AI.
Alright friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, Blitzy, Rovo, Assembly, and Robots and Pencils.
To get an ad for your version of the show, go to patreon.com/aidailybrief, or you can
subscribe on Apple Podcasts to learn about sponsoring the show, and pretty much anything
else about the show, including jobs, speaking opportunities, etc., visit aidailybrief.ai.
As I mentioned yesterday, we are in the last week of accepting submissions, so if you want
the full readout and report, go to roisurvey.ai and contribute your use cases now.
Related to that, between this new benchmarking study and some other exciting things coming
up, I have a few exciting projects for researchers/analysts.
If you have immediate availability and a research or data analysis background, shoot us a note
at
[email protected], put research in the subject line, share your background, and I'll
tell you a little bit more about what we're thinking.
With that, though, let's talk Gemini3.
Welcome back to the AI Daily Brief.
Happy Gemini3 day to those who celebrate.
It has been a long time coming.
The final anticipated model drop of the year barring some big surprise from open AI has
finally arrived as Google has announced Gemini3.
Now the focus has been on Gemini3 for at this point months, really ever since GPT-5 launched.
In fact, the launch of GPT-5, which initially had a bit of a bobble around first impressions,
ratcheted up the pressure in some ways on Gemini3 for the entire industry.
Over the last few weeks, though, the rumour mill has been swirling and hype has been building
incredibly for this.
An example from just a few days ago, Marmaduke091 on Twitter says, "Okay, I'm seeing it now
with full confidence.
Gemini3 will make every other LLM irrelevant and Nanobanana2 will make every other image
model irrelevant.
It is not close in Google 1.
This is just my honest feelings after seeing things.
Nobody is ready."
There were effectively infinite posts of that type, including people who said that they had
access and were seeing things.
So much so, in fact, that in the last couple of days, you've actually started to see a
bit of a shift where some folks have nudged in the other direction.
Peter Wilderford shared the meter chart for the length of autonomy that a model can perform
at, saying Gemini2.5 released in June was a great model, but it was a bit disappointing
on the meter's benchmark below trend at 39 minutes.
Where will Gemini3 be?
If on trend it should be around 3 hours.
My forecast is Gemini3 remains a bit below trend at 2.7 hours.
Even ThwaV DD went much farther.
On Monday they wrote, "Almost every day I hear and see more things that lower my expectations
for Gemini3.
That is once again the case today, and honestly I'm not that excited for this launch anymore.
The model regresses with every new checkpoint.
DeepMind should be managing expectations.
Instead we get more hype posting with little substance or transparency.
Hopefully it will be a lot better in the real world, but my expectations there are similarly
low."
Now to be clear, this was the exception going into this release, not the rule.
Most people were incredibly hyped, and if you needed any more confirmation of that,
last night the normally highly reserved Google DeepMind CEO Demis Asabis tweeted it's nearly
three here, my favorite part of the night shift, locked in, "Demis does not hype post."
One little coder writes, "You guys even made the calmer Demis hype up."
But let's talk for a second about the stakes of this launch, for Anthropic and OpenAI.
Google has been surging throughout the year, picking up market share, picking up mind share,
and picking up brand relative especially to ChatGPT.
For the first time earlier this year, the Gemini app actually jumped up over ChatGPT
for a little while in the Apple App Store charts.
There is a lurking narrative surrounding OpenAI and Anthropic, that the 800 pound gorilla
that is Google will eventually just have too many resources and has a certain ultimate
inevitability.
And so the question coming into this for those companies would be, would Gemini 3 blow them
out of the water?
Would there still be areas where their models, especially for Anthropic their coding models
were on par or still better than Gemini 3?
And in general what did this say for the state of the consumer AI app race?
This launch also had some implications for Nvidia.
While it wouldn't be as direct as the comparison between Gemini and ChatGPT for example, Google
is also building out the full stack around AI including their own chips called tensor processing
units or TPUs.
If Google trained this model primarily with their TPUs and it was more performant than
other models that were trained on Nvidia chips, maybe that would have implications for how
the market viewed Nvidia.
Speaking of the market, AI bubble talk continues to surge and continues to be the dominant
theme in markets right now.
Google CEO Sundar Pichai actually contributed to this saying in a recent interview with
the BBC, there is some irrationality in the current AI boom.
The growth of AI investment has been an extraordinary moment.
When asked what happens if the AI bubble pops he said, I think no company is going to be
immune including us.
We can look back at the internet, there was clearly a lot of excess investment but none
of us would question whether the internet was profound, I expect AI to be the same.
So this is a theme we've heard from lots of leaders before, that yes there might be a
rational exuberance when it comes to valuations and there may even be a correction but ultimately
the technology underneath will be every bit as significant as people say it is.
Now that plus broader macro factors including Federal Reserve expectations contributed on
the day of the Gemini launch to a big tumble in stock prices led by the AI tech sector.
And so the reason that Gemini 3 matters from that macro market perspective is that one
of the things that would confirm for some investors their concerns about an AI bubble
would be if it appeared that we had reached some meaningful plateau.
In fact, the conversation around GPT-5 being underwhelming when it came to expectations
helped ratchet up this latest generation of AI valuation and AI bubble fears.
So you've got implications for their AI model competitors, implications for their infrastructure
competitors, implications for the market as a whole.
For Google itself, honestly, I kind of think the stakes were the lowest.
Google is basically locked in an incredibly good year on their AI products.
The number of users has gone up, the amount of tokens their processing has gone up and
unless Gemini 3 was an absolute total flop, it's hard to see that trend changing.
Now of course, that doesn't mean that they couldn't use this moment to really reinforce
a new leapfrog position.
But when it comes to downside risk, I was actually least concerned about Google heading
into this announcement.
Now pretty much everyone knew it was coming today.
And indeed, at 8am Pacific Time, 11am Eastern Time, Sundar Pichai tweeted introducing Gemini
3.
It's the most powerful model in the world for multimodal understanding and our most
powerful, agentic and vibe-coding model yet.
Gemini 3 can bring any idea to life, quickly grasping context and intent, so you can get
what you need with less prompting.
The companion blog post was called a New Era of Intelligence with Gemini 3.
In the note, we got a few statistics.
The Gemini app is now up to 650 million users per month, that's about 50 million more than
last we heard, and Sundar also said that 13 million developers have now built with their
models.
Now one important thing, it's clear that Google has learned from their own launches
in the past that the market does not like when you make announcements and say "coming
soon."
Sundar writes, "Starting today we're shipping Gemini at the scale of Google.
That includes Gemini 3 and AI mode in search with more complex reasoning and new dynamic
experiences.
This is the first time we are shipping Gemini in search on day one.
Gemini 3 is also coming today in the Gemini app to developers in AI Studio and Vertex
AI and in our new agentic development platform Google Antigravity, which I will get into
in a little bit."
In Demis' section of the announcement post, he calls this another big step on the path
to AGI and gets into the monster set of benchmarks that we'll get into in just a moment.
Now interestingly, the blog post is really focused on what you can do with this model
organized around learning anything, building anything, and planning anything.
In his announcement tweet, Demis said that beyond the benchmarks it's been by far his
favorite model to use for its style and depth.
In an example, he writes, "I've been doing a bunch of late night vibe coding with Gemini
3 in Google AI Studio and it's so much fun.
I recreated a testbed of my game theme park that I programmed in the 90s in a matter of
hours, down to letting players adjust the amount of salt on the chips."
In his announcement post, Google VP Josh Woodward discusses six additional features that come
with the launch.
The first is generative interfaces, which is a new type of experimental interface that's
generated on the fly and adapts to the user's needs based on the prompt.
Gemini Agent, which he writes is an experimental tool that orchestrates and completes complex
multi-step tasks, a new look for the Gemini app, better shopping results, support for
23 more languages, and a new promotion where U.S. college students can get a free year
of AI Pro.
This episode is brought to you by Blitzi, the enterprise autonomous software development
platform with infinite code context.
Blitzi uses thousands of specialized AI agents that think for hours to understand enterprise
scale code bases with millions of lines of code.
As engineering leaders, start every development sprint with the Blitzi platform, bringing
in their development requirements.
The Blitzi platform provides a plan, then generates and pre-compiles code for each task.
Blitzi delivers 80% plus of the development work autonomously, while providing a guide
for the final 20% of human development work required to complete the sprint.
Public companies are achieving a 5x engineering velocity increase when incorporating Blitzi
as their pre-IDE development tool, pairing it with their coding pilot of choice to bring
an AI native SDLC into their org.
Visit blitzi.com and press Get a demo to learn how Blitzi transforms your SDLC from AI-assisted
to AI-native.
Meet Robo, your AI-powered teammate.
Robo unleashes the potential of your team with AI-powered search, chat, and agents, or build
your own agent with studio.
Robo is powered by your organization's knowledge, and lives on Atlassian's trusted and secure
platform so it's always working in the context of your work.
Connect Robo to your favorite SaaS app so no knowledge gets left behind.
Robo runs on the teamwork graph, Atlassian's intelligence layer that unifies data across
all of your apps and delivers personalized AI insights from day one.
Robo is already built into Jira, Confluence, and Jira service management standard, premium
and enterprise subscriptions.
Know the feeling when AI turns from tool to teammate?
If you Robo, you know.
Discover Robo, your new AI teammate powered by Atlassian.
It started at rov as in victory, oh.com.
If you're building anything with voice AI, you need to know about assembly AI.
They've built the best speech-to-text and speech understanding models in the industry,
the quiet infrastructure behind products like granola, dovetail, ashby, and cluly.
Now, as I've said before, voice is one of the most important modalities of AI.
It's the most natural human interface, and I think it's a key part of where the next
wave of innovation is going to happen.
Assembly AI's models lead the field in accuracy and quality so you can actually trust the
data your product is built on.
And their speech understanding models help you go beyond transcription, uncovering insights,
identifying speakers, and surfacing key moments automatically.
It's developer first, no contracts, pay only for what you use, and scales effortlessly.
Go to assemblyai.com/brief, grab $50 in free credits, and start building your voice AI product
today.
Today's episode is brought to you by Robots and Pencils.
When competitive advantage lasts mere moments, speed to value wins the AI race.
While big consultancies bury progress under layers of process, Robots and Pencils builds
impacted AI speed.
They partner with clients to enhance human potential through AI, modernizing apps, strengthening
data pipelines, and accelerating cloud transformation.
With AWS-certified teams across U.S., Canada, Europe, and Latin America, clients get local
expertise and global scale.
And with a laser focus on real outcomes, their solutions help organizers work smarter and
serve customers better.
They're your nimble, high-service alternative to big integrators.
Turn your AI vision into value fast.
Stay ahead with a partner built for progress.
Partner with Robots and Pencils at robotsandpencils.com/aidailybrief.
Now one thing that was interesting to some folks is that Google launched this huge model
with just blog posts and videos.
No live stream, no presentation, but as it turns out, there was plenty for people to
talk about even without that.
If you are a regular listener, you'll know that I'm very skeptical of the ultimate value
in benchmarks when it comes to a new model.
I think most of our benchmarks are getting fairly saturated, and they don't necessarily
tell you how a model is going to interact with your particular use cases.
In other words, while they can be useful guides, they're no substitute for just getting in
there and trying the thing out.
However, to the extent that benchmarks can make a splash, boy do Gemini3's benchmarks
make a splash.
Vracerex sums up the attitude of basically everyone that I've seen online, writing,
"The Gemini3 Pro benchmark results are genuinely unreal.
Google didn't just catch up today.
They walked into the arena and rewrote the difficulty settings."
To take a few examples, on the academic reasoning-focused humanities last exam, GPT-51 was at 26.5%.
Gemini3 Pro is at 37.5%.
On MMMLU, one of the more saturated benchmarks for multilingual Q&A, Gemini3 Pro nudges out
over GPT-51 91.8% to 91%.
On GPQA Diamond scientific knowledge, GPT-51 is at 88.1% versus 91.9% for Gemini3 Pro.
And basically this story repeats itself kind of across the board.
The only benchmarks where it either shared the title or was behind were on the AIME
2025 with code execution, where it tied with Sonnet 4.5 and 100%, and SWE Bench verified,
where it was just ever so slightly behind both Sonnet 4.5 and GPT-51 at 76.2% compared
to 5.1 76.3% and Sonnet 4.5 77.2%.
Now before you think that means that it's worse at coding, on another coding benchmark,
terminal bench 2.0, which is a gentic terminal coding, Gemini3 Pro set a new standard at
54.2% compared to Sonnet 42.8% and 5.1's 47.6%.
Now beyond the biggies, some people noticed a few where the jump really stood out.
Matt Schumer writes, "One of Gemini3's biggest leaps is its towering score on ScreenSpot Pro."
This by the way is a measure of a model's ability to understand what's going on on a
screen.
The previous state-of-the-art was Sonnet 4.5 at 36.2%, Gemini3 Pro got 72.7%.
Schumer says, "It just massively accelerated my timeline to full computer using agents."
Another one that stood out to many was the performance on ARC-AGI 2.
ARC-AGI is of course explicitly designed to be difficult for computers, and while GPT-51
mustered a 17.6%, Gemini3 Pro got a 31.1%.
ARC Prize founder Francois Chalet called it "impressive progress."
On the VPCT Spatial Reasoning Test, Gemini3 absolutely smashed the test at 91% compared
to GPT-5's high 66%.
And when it comes to user preference overall, LM Arena writes, "Breaking!
Google DeepMind's Gemini3 Pro is now number one across all major arena leaderboards.
Number one in text vision and web dev.
Number one in coding, math, creative writing, long queries, and nearly all occupational
leaderboards.
Massive gains over Gemini2.5."
And indeed, Gemini3 also released a DeepThink mode that had even higher results.
For example, the DeepThink mode drove the score on ARC-AGI up to 45.1%.
Again, Matt Schumer writes, "The last time we saw a capability jump of this magnitude
was the release of GPT-4 in March 2023.
We're entering a new era."
In addition to LM Arena, we also got the independent review from Artificial Analysis.
They put it quite clearly, Gemini3 Pro is the new leader in AI.
In their aggregate score, Gemini3 Pro shows up three points ahead of GPT-5.1.
Going back to that concern from the market bubble people, Simon Smith writes, "So I
guess we haven't hit a wall."
Now, at the time of recording, this model has been out to the general public for less
than an hour.
In fact, outside of Google AI Studio, I'm not seeing it yet in any of my literally four
different Gemini accounts.
So aside from early preview tests, people haven't really had a chance to get their
hands on it.
There are folks who have had a few reps though and here's some of the things that they've
shared.
Dan Shipper and the Every Team were testing it earlier this morning and said that so far,
on coding, it's extremely fast and the quality seems very high.
In long context understanding, it has a, quote, "massive context window" and it can
use it.
Dan writes, "I found it was able to find, synthesize and use pieces of information in
a long book draft that other models couldn't."
Now interestingly, Dan says that when it comes to writing, in our early test, it's not as
good of a writer or editor as Sonnet or Haiku.
Dan says it appears worse at telling whether writing is interesting.
Far Al started to play around with it on AI Studio and said it's so freaking fast and
is actually pretty good.
Shumer again writes, "Gemini III's biggest differentiator, in my opinion, is that it's
far easier to get it out of the standard slop style for writing, coding, etc.
Just ask for what you want and it'll make it happen."
Referencing the tendency of AI coding models to design purple shaded interfaces, no more
purple gradients.
He also gave the example of it redesigning his personal website.
Matt actually wrote a blog post about his last three days using the model in preview.
The TLDR, he writes, "Gemini III is a fundamental improvement on daily use, not just on benchmarks.
It feels more consistent and less spiky than previous models."
He says it's fast, intelligence per second is off the charts, often outperforming GPT-5
Pro without the weight.
Front end capabilities, which we were just discussing, he says are excellent, with it
nailing design details, micro interactions and responsiveness on the first try.
This might not be great for the people who love 4.0, but for the work user, Matt writes,
"It respects your time and doesn't waste tokens on flowery preambles."
Lastly, he says, "Creative writing is finally good.
It doesn't sound like AI slop anymore.
The voice is coherent and the pacing is natural."
Now, of course, that sounds a little bit different than what we heard from Dan, and also different
than what Murtacan Coylan found.
He wrote, "I'm doing my writing with constraints tests and Gemini III Pro is not living up to
the hype at least for this task."
He basically talks about a writing test that he's created and goes in depth about how
he thinks it fails, coming to the conclusion so early to say, but this model lacks the
taste restraint and structural intelligence necessary for challenging creative work.
My big takeaway is that writing is one area where we're going to have to test it a little
bit more.
Now, a lot of the folks who are sharing their first impressions are focused on improvements
in its coding ability and improvements in its design sensibility.
The PHO Charano writes, "I asked Gemini III Pro to create a 3D Lego editor.
In one shot it nailed the UI, complex spatial logic and all the functionality.
Reiterating something that we heard from others, PHO concludes, we're entering a new era."
He also said, "It's also amazing at games.
It recreated the old iOS game called 'Ridiculous Fishing' from just a text prompt including
sound effects and music."
He continues, "It also did something I've seen other LLMs struggle to do before.
It built a fully functional Game Boy emulator, and yes, it even drew the Game Boy as an SVG."
Flavio Adamo also tested it building games, showing off not only the ability to build
a game engine, but also its understanding of spatial physics, with a game where you rotate
a ball running through a tunnel to try to avoid oncoming objects, with Flavio saying,
"Built this fun game in literally five minutes and it's way better at coding than I expected."
Now on the coding topic, one of the things that came alongside Gemini III was a new Google
native IDE called Anti-Gravity.
In their announcement post, Google wrote, "As Model Intelligence accelerates with Gemini
III, we have the opportunity to reimagine the entire developer experience.
Today we're releasing Google Anti-Gravity, our new agentic development platform that
enables developers to operate at a higher task-oriented level.
Using Gemini III's advanced reasoning, tool use, and agentic coding capabilities, Google
Anti-Gravity transforms AI assistance from a tool in a developer's toolkit into an active
partner."
While the core of Google Anti-Gravity is a familiar AI IDE experience, its agents have
been elevated to a dedicated surface and given direct access to the editor, terminal, and
browser.
Now, agents can autonomously plan and execute complex end-to-end software tasks simultaneously
on your behalf while validating their own code.
Finally, they point out that in addition to taking advantage of Gemini III Pro, Anti-Gravity
also has access to their latest computer use model for browser control, as well as nano-banana
for image generation.
Now, this might have been the part of the announcement that a lot of at least the developers
and builders were most excited for.
In his post-Google AI studio lead Logan Kilpatrick wrote, "Anti-Gravity is a faster way to develop.
You act as the architect collaborating with intelligent agents that operate autonomously
across the editor, terminal, and browser.
These agents plan and execute complex software tasks, communicating their work with the user
via detailed artifacts.
This elevates all aspects of development from building features, UI iteration, and fixing
bugs to researching and generating reports."
So is this the end for cursor?
Hyperbole aside, people do seem excited about this new launch.
The AI for success account wrote, "I was one of the early testers and this thing is
crazy good.
Steren says an IDE isn't actually the right framework.
He writes, 'I've been using Google Anti-Gravity for a few days, and to me it's not an IDE,
it's a coding agent UI powered by Gemini III Pro.'
Notably, he writes, 'The agent can control a Chrome browser, enabling it to validate
open and use the app that it builds.'
A surprising example of anti-gravity accessing a browser, Steren writes, 'I asked to generate
an SVG.
It did of course then ask to transform to PNG.
Anti-gravity looked for image magic, couldn't find it, then looked for FFMPEG, couldn't
find it.
It rendered the SVG in Chrome and saved the pixels.'
Max Weinbach writes, "The anti-gravity IDE from Google is my favorite one now, been
using it for a few days and it's been outperforming cursor and Windsor for me, even most CLIs."
Richard Serrata writes, "I've been using Google Anti-Gravity for a few weeks and it's
wild.
It's an early preview so you'll find quirks, but you'll be blown away by the agent stuff."
So where is this all net out?
Well, firstly, I certainly think that the AI bubble narrative did not get worse today.
Maybe we will see over the next week people quibble that there aren't some monster, crazy,
highly noticeable advances, but this appears to be at first glance a pretty significant
jump in capabilities.
When it comes to the competitive dynamics, we'll have to see.
Gemini 3 looks like a great model, but I've been absolutely loving 5.1 as well and it's
going to take a lot to displace it for me.
Ultimately, what we have here is what appears to be by all accounts a great new model from
Google, one that is shifting behaviors already for some, and even new tools like anti-gravity
around it that could shift how people interact with AI.
Next up for me, of course, is getting in there and actually trying it.
That's basically what I plan to spend the whole rest of the day doing, and I will be
back with what I find in the next couple of days.
For now, as I said at the beginning, happy Gemini 3 day to those who celebrate, and until
next time, peace!
PodcastTranscript AI