All right, Google just made a bunch of Gemini announcements. Gemini 2.0 is here; we have Project Astra, we have agents that can browse the web, we have deep research, and so much more, plus a bunch of other news stories that happened this week. Let's get into it. All right now,

right, so first, the biggest news story of the week: Gemini 2.0 is here, and Google is making a big push to let it be known they are in the AI race very much so. Let's go over some of the announcements because there were a lot. The first thing

is Gemini 2.0, their newest Frontier Model, and they have a Flash version, which is incredibly fast, very cheap, but still performs really well. Gemini 2.0 Flash outperforms Gemini 1.5 Pro. Yes, you heard that right: the 2.0 Flash model,

the small model, the fast model, outperforms the previous Pro version of 1.5, and it's twice as fast. It supports multimodal inputs like images, video, and audio. 2.0 Flash also supports multimodal output like natively generated images mixed

with text and steerable text-to-speech, multilingual audio. It can natively call tools like Google Search. This is a critical differentiator because not many people have that native search feature; they're using either the Brave browser or some other third-party search versus having the index themselves, and this is

really where Google can shine. And it has code execution, which, of course, ChatGPT just announced, and other tool calling. So how does it perform? Well, here are all the benchmarks on the left side. Here's Gemini 1.5 Flash and Gemini 1.5 Pro, both

of which pretty much across the board. Gemini 2.0 flash, experimental, beat. And if you've been keeping an eye on the top leaderboards, you know that the Gemini models have been in the top three for weeks, if not months now. Now this is the

one that's really impressive: nearly 93% on natural-to-code. That is, being able to generate code from natural language. And they have live code bench right here at 35%. Now, one thing that model providers sometimes do that I don't really like is just compare their models

versus other versions of their models. Now, I want to see it compared to Llama, ChatGPT, Claude, and so on. I don't want to just see one family of models compared to itself. So, just a little piece of feedback to Google, if you

happen to be watching this. Gemini 2.0 flash is available right now, so try it out. And not only is it available now to developers, but it's also available in their Gemini app, which is their AI assistant. And I have an iPhone, so I'm

not really using it, but I bet it's better than Apple intelligence. But that's not all. Gemini 2.0 flash is especially good at native user interface interaction; basically, it can browse the web on your behalf. And that is a big theme of all the announcements; they are

very much seeing the future as agentive, which, of course, I know I sound like a broken record, but bullish on agents. So all of the features announced with Gemini 2.0 help it to be able to navigate complex UI environments; so, multimodal reasoning, long context

Understanding complex instruction following, planning, compositional function calling, native tool use, and improved latency—they all work together for a new class of agentic experience. Next is Project Astra, and we saw a preview of this a long time ago, but now it seems like it's

finally here. So, if you're not familiar with Project Astra, it is a way for AI to basically use your camera, look around at the world, and be able to tell you things about the world. So you can ask it questions, and it can see, and it

seems really cool. And it's much more than just a vision-capable model; it can remember things and it can reason about your environment based on that memory. There are a lot of potential use cases that can be unlocked; it's essentially like having a second brain. Fellow AI

YouTuber Matt Wolfe was actually invited to London by Google to try it out, and he just posted a video, so definitely check that out. So you can kind of think of it like advanced voice mode, but with vision, memory, better reasoning, and basically all the things. So it seems

really impressive. Now, the only problem—I think it's only available on Android. So, some of the improvements: better dialogue, new tool use, better memory, improved latency. So these are all very welcome features of an AI assistant that's really supposed to help you. Thanks to Emergence AI for partnering with me on

this video. If you've watched this channel at all, you know I'm bullish on agents, and especially agents that can accomplish real-world tasks for you. That's why I'm excited to tell you about Emergence AI. Emergence AI just launched their enterprise-grade multi-agent

orchestrator, and they showed off the first demo of a real-world use case where these agents can actually browse the web on your behalf. This is enhanced web automation. This means multiple agents can go out and dynamically interact with different elements on the web under this intelligent

orchestration to bring human-like interaction and navigation, but at machine-level scale. And what's really cool about this is these agents can actually perform complex and sophisticated web interactions that previously required a human. They can navigate dynamically loading menus, fill out forms, adjust settings, process

embedded files, and extract relevant data from PDFs and HTML. Emergence AI's orchestrator offers a combination of design-time flexibility and runtime determinism, which basically means that these agents can heal themselves. So if it makes a mistake along the way, it can figure it out and then succeed on the

next try. And Emergence AI has put a huge emphasis on privacy and security. They offer a fully hosted solution where you can access it via API, or you can host it on your own virtual private cloud. So if you're an enterprise business and you're looking to automate a lot of your

processes, Emergence AI is a great solution. Integrate their agent API and seamlessly orchestrate multiple agents to accomplish tasks for you and your business. This includes interacting with both modern and legacy enterprise applications. So Emergence AI is just starting to invite developers to try out their

platform. Definitely check them out; tell them I sent you, of course! Go to their website, emergence.ai, or simply email them: contact@emergence.ai. I'll drop all of the links in the description below. So thank you again to Emergence AI for partnering with me on this video. Now,

Back to the video. All right, so I already talked about Gemini's ability to navigate complex user interfaces, and that's what Project Mariner is. It is Google's foray into giving agents the ability to actually control your browser, and it comes in the form of a Google

Chrome extension. So it is a research prototype; it is experimental; it is not publicly available, but hopefully it is soon to beta testers because I definitely want to test it out. And right here it says Project Mariner achieved a state-of-the-art result of 83.5% working

as a single-agent setup. So basically, you gave it a task to accomplish on the web, and it was able to do so. This is the future of agents; I could not be more sure. Now we already know Operator from OpenAI should be coming out in the new

year, maybe during the 12 Days of OpenAI, we'll see. We saw Runner, which is a similar project to navigate the web, and we've seen Claude compute use. So really, giving agents a way to control your computer, control your browser, is definitely coming soon. And then next,

they also released Jewels, which is interesting because Devon was just released publicly this week as well. I'll get to that in a moment, but Jewels is agents for developers, and it integrates directly into your GitHub workflow, which is nice, and you can just tag it and let

it accomplish things for you. It can tackle an issue, develop a plan, and execute it all under a developer's direction and supervision. So very similar to GitHub workflows, very similar to Devon, very similar to Pygmalion and CodeBuff, and a lot of other projects that are just making it so much easier

for developers to accomplish so much more, and then, in a surprise announcement, they're also allowing agents to watch you play video games and actually give you tips, which seems kind of wild to me. I did not expect this. So, here it says Google DeepMind has a long history of

using games to help AI models become better at following rules, planning, and logic. Just last week, for example, we introduced Gemini 2, our AI model that can create an endless variety of playable 3D worlds, all from a single image. Now, they've built agents using Gemini 2 that

can help you navigate through a virtual world, and it can reason about the game based solely on the action on the screen. They're collaborating with one of my favorite game developers, Supercell. I love Brawl Stars and I love Clash Royale, and so these agents can basically watch

you play and give you suggestions. So they are really thinking about putting Gemini 2.0 in everything, and I've said this a few times, but not only that, they're putting it in embodied robots in the real world. So we'll see. I mean, it

seems Gemini 2.0 is very impressive. So that's the summary of all of Gemini's updates. Let me know what you think in the comments below. I would say I'm very impressed, but I want to know from you. Let me know. All right, since we're

talking about Google, let's talk a little bit of drama now. It seems, according to the information, Google has asked the FTC to basically end Microsoft's deal with OpenAI, and specifically their cloud deal. Now, the whole deal between Microsoft and OpenAI is weird to say

The least, and all of it really hinges on what OpenAI's definition of AGI is, because as soon as they hit AGI, the entire deal changes pretty much overnight. Google recently asked the US government to break up Microsoft's exclusive agreement to host OpenAI

technology on its cloud servers. Exclusive deals aren't necessarily illegal, but companies that buy OpenAI technology through Microsoft may face new costs if they don't already use Microsoft servers to run part of their business. Microsoft gets a 20% cut of OpenAI's revenue and, as you probably

know, they also own 49% of the overall company. And you've probably already heard OpenAI transitioning to being a for-profit, so there's just so much going on with the corporate strategy of OpenAI and Microsoft, and we're going to see how all of this plays out. Now, this is

all at a time where the US government is looking very closely at Google and potentially, and it seems likely, going to break them up and maybe they're going to spin off the Chrome browser. I don't actually think we have the final details yet, but it's interesting how all of

these massive tech companies are all aggressively playing legal battles. Next, moving on from Google, we're now talking about General Motors. Now, when I was in San Francisco for about a decade, I remember this little company that started called Cruise, and they made autonomous vehicles, and they were a

really popular company within that insular environment of Silicon Valley, and then they were acquired by General Motors. Now General Motors is shutting it down, which is kind of crazy because they were bought for billions of dollars. GM plans to realign its autonomous driving strategy and prioritize development of

Advanced driver-assistance systems are on a path toward fully autonomous personal vehicles. So, rather than having an autonomous fleet of taxis, for example, they're betting more on personal autonomy—basically, your own car driving itself. This is both surprising and not so surprising. Whenever a non-tech company

acquires a hardcore tech company, I haven't really seen it work out all that well many times. Now, if you can think of a good example, let me know in the comments. Now, it's definitely disappointing to see. I'm excited for Tesla's robotaxi. I know Google has Waymo,

there's this other company, Zoox, which looks really cool. So there's definitely a lot of competition out there still, but as always, more competition is better. All right, so I talked about it earlier; let's talk about Devon. Devon is now publicly available, but there's a catch: it's

expensive—$500 a month gets you Devon. Now, if you put it in context of what it is actually delivering in terms of value and time savings, maybe it's not expensive. If you just think about it as a junior developer, it's actually quite

cheap at $500 per month, $6,000 per year. Devon seemingly is very impressive, but transparently, I have not tested it myself yet. I do plan on doing it; I just haven't done it yet. And as part of their announcement, they actually linked to a

full Devon solution for a GitHub issue, and the GitHub issue happens to be for mCP (model context protocol), which is from Anthropic. So it's kind of cool to see this. So this person on December 6 says, "Hi, please figure out a good solution for"

This issue linked the issue. Use the local copy of the inspector repo; don't make a PR yet. Devin starts doing it, grabs everything, makes a plan, and executes on the plan, pushes the code, and creates a PR. All of this is done autonomously; very impressive, so very

cool interface, very easy to use seemingly. But I always look at it and think, what is their differentiator? If these models are publicly available for any developer to build on top of, what are they really doing differently? Do they have a special version of the model?

I don't think so. They definitely have a really cool interface; they probably have a lot of agentic scaffolding on top of the model, but I don't know how far that gets you. We'll see. Either way, seems very cool. Congrats to Dev on the public

launch. All right, next, it seems Gro has a new image model; it's called Aurora, but then they removed the Aurora title. But either way, it's out. It can create people, famous people, and it looks really accurate, really high quality. It is an autoregressive mixture

model, a mixture of experts network trained to predict the next token from interleaved text and image data; that is Aurora. So let's take a look at some examples. So here's the locked SR-71 Blackbird in an abstract style. Here's the Optimus robot wearing an Xmas

costume. Here is a Grock logo, some flowers, which looks really good; an origami Cybertruck. Here's Jackie Chan in Donald Trump's hairstyle. So it obviously has no problem reproducing celebrities and other well-known people. So I've seen Bill Gates, I've seen Donald

Trump, I've seen Kamala Harris, I've seen everything, and it doesn't refuse ever. And I think Grok is kind of known as being the most permissive in terms of what it will do for you. Google was really on a roll this week; they also

announced Willow, which is a completely game-changing new quantum chip. Quantum is not something I'm very familiar with, but I'm starting to learn more, and I definitely want to talk about it on this channel. If you want to hear about it, let me know. So, quantum chips

are able to compute certain types of use cases infinitely faster than traditional CPUs or GPUs can. And I'm not going to get into the details of how they do that, but one of the biggest problems of why quantum computing hasn't gone mainstream is because of the error rate, and

previously, as the number of qubits—essentially, the bits of the quantum computer—increased, the errors increased. And so, at a certain point, it was just unusable. But now Google released Willow, which actually has the opposite effect: as you increase the qubits, the error

rate decreases exponentially. This is a fundamental unlock for quantum computers and has a lot of potential consequences for the entire world, whether we're talking about research or weather prediction or encryption. All of your passwords might be at risk; cryptography in general—crypto, Bitcoin, Ethereum—these

things might be able to be just, quote-unquote, 'solved' by quantum computers. Now, here's the good thing: it is extremely expensive, both resource-wise and expertise-wise—basically everything—to create a quantum computer. There's only a handful, or less, companies in the entire world

That can actually pull this off, and only one so far has figured out how to get rid of those errors really well. So we're not anywhere close to some hacker in a basement being able to build a quantum computer and break everybody's passwords, so don't worry about that, but...

a really cool update nonetheless. Listen to this stat: Willow performed a standard benchmark computation in under 5 minutes that would take one of today's fastest supercomputers 10 septillion years—that's 10 to the 25th, a number that vastly exceeds the age of the universe. So it's

basically impossible to compute on traditional computing, and it only took five minutes. And if you have any articles or books you suggest that I read to learn more about quantum computing, let me know down below. All right, next, it seems like December is AI

month, because there have just been so many updates and releases. Of course, Zuck has something to say; they have released Llama 3 37B, and he also talks about the future of AI. Let's watch. All right, hey guys, I've got our last big AI update of the

year. Our goal at Meta is to build general intelligence and open-source it, that way everyone benefits. Meta AI now has nearly 600 million monthly actives and, as promised, is on track to be the most used AI assistant in the world by the end of the year. If you haven't

gotten a chance to try voice mode yet, try it; it's awesome. Now, I started 2024 talking about how we had a bunch of H100s, that this was going to be a big year for open-source AI, and it has been. Llama has quickly become the most

adopted model with more than 650 million downloads, and it is starting to become an industry standard. We have one more release for the year: Llama 3.3 is a new 70 billion parameter text model that performs about as well as our 405 billion parameter model, uh, but now it is easier and more efficient to run. So that is the last Llama 3 release; the next stop is Llama 4.

And you know, if we started the year talking about the size of our clusters for training Llama, I want to start to wrap the year by doing the same. Uh, we just announced that we were building a 2+ GW data center in Louisiana that we're going to use to train future versions of Llama. Uh, we are in this for the long

term, and we are committed to building the most advanced AI in the world. All right, thank you for a great year so far. Looking forward to sharing more soon. All right, so let's look at Llama 3 37B, because there's a lot here. It is better

performing than 405B, but easier and more cost-efficient to run. So we can see the benchmarks right here. Here's Llama 3 17B and Llama 3 37B, and it looks like a modest improvement from 3 17B, but the best part: open source, so

you can download it, play around with it, and use it, and it has a very permissive license. All right. And then next, this happened a little over a week and a half ago, so it's kind of old news, but I just want to mention it quickly:

Tech investor and operator in Silicon Valley has been chosen by President Trump to be the AI and crypto czar for the United States. So, kind of interesting. I'm a big fan of the All-In podcast; I watch it pretty religiously. And now seeing David Sacks being the head of

such an important part of the government as it relates to AI and tech and cryptocurrency, I have mixed feelings about it, but we'll see. I definitely like that he chose somebody that is very technical, very forward-thinking. Sacks was also PayPal's COO, kind of the founding

era of PayPal. He built Yammer and sold it for billions. He founded Craft Ventures, which has made a number of investments in every area of tech, but mostly known for SaaS. And so we'll see how it plays out. And so that's it for

today. A lot of news! I know, a lot of cool stuff. Of course, we still have multiple days left in the 12 Days of OpenAI, so I'll be here for it. I'll be reporting to you. If you enjoyed this video, please consider giving a like and subscribe, and

AI News: Gemini 2.0, Devin, Quantum Computing, Llama 3.3, and more!

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

Gemini 2.0 launch and Google's emphasis on its role in the AI race.

Gemini 2.0 Flash surpasses Gemini 1.5 Pro in performance and speed.

Multimodal input/output capabilities (images, video, audio, text-to-speech).

Native Google Search integration and code execution.

Benchmark results showcasing Gemini 2.0 Flash's superior performance.

Availability to developers and in the Gemini AI assistant app.

Project Astra's world-understanding capabilities using camera input.

Potential use cases and comparison to an advanced voice assistant with vision.

Improvements: better dialogue, tool use, memory, and latency.

Android exclusivity mentioned as a limitation.

Emergence AI's enterprise-grade multi-agent orchestrator.

Enhanced web automation with multiple agents interacting dynamically.

Ability to perform complex web interactions previously requiring humans.

Self-healing agents and emphasis on privacy and security.

Hosted solution via API or on private cloud for enterprise use.

Project Mariner and other Agentic AI projects

Project Mariner: Google Chrome extension for browser control by agents.

State-of-the-art result of 83.5% in single-agent web task completion.

Discussion of similar projects from OpenAI and others.

Jewels: agents for developers integrating with GitHub workflow.

Agents providing tips while watching video game play.

Google's FTC request regarding Microsoft and OpenAI

Google's request to the FTC to end Microsoft's deal with OpenAI.

Focus on OpenAI's AGI definition and implications for the deal.

Microsoft's 20% revenue cut and 49% ownership of OpenAI.

OpenAI's transition to a for-profit company.

US government's scrutiny of Google and potential breakup.

General Motors' decision to shut down Cruise, its autonomous vehicle subsidiary.

Shift in strategy towards advanced driver-assistance systems.

Focus on personal autonomy rather than autonomous taxi fleets.

Discussion of other companies in the autonomous vehicle space.

Devon's public availability at a cost of $500 per month.

Discussion of Devon's value proposition and cost-effectiveness.

Example of Devon solving a GitHub issue for Anthropic's mCP.

Analysis of Devon's differentiators and competitive advantages.

Grok's Aurora image model and other AI advancements

Grok's new image model (Aurora) and its capabilities.

Examples of images generated by the model, including celebrities.

Grok's reputation for permissiveness in generating content.

Google's announcement of Willow, a new quantum chip.

Willow's potential impact on various fields and cryptography.

Discussion of the high cost and expertise required for quantum computing.

Willow's benchmark performance compared to traditional supercomputers.

Meta's release of Llama 3 37B and its open-source nature.

Meta's goal of building and open-sourcing general intelligence.

Llama's adoption as an industry standard with over 650 million downloads.

Llama 3 37B's performance comparable to a much larger model but more efficient.

Meta's commitment to building advanced AI and its data center expansion.

David Sacks' appointment as AI and crypto czar for the United States.

Discussion of David Sacks' background and experience.