Google Cloud Next - Gemini 2.5 Pro EVERYWHERE

Summary

Highlights

• The Rubik's Cube demo by Gemini 2.5 Pro shows it's not just fun and games; it's deep reasoning and coding prowess. 🧩
• Ironwood TPU sets new benchmarks in AI chip performance, promising unprecedented speed and efficiency. 🚀
• Agent interoperability is set to redefine cross-platform communication with the introduction of a new development kit. 🌐
• Google announced new models for text-to-image (Imagine 3), voice generation (Chirp 3), and music creation (LIA). These innovations enhance creative output across media. 📸
• Improvements in energy efficiency underscore a sustainable approach to powering cutting-edge AI models. ♻️
• V2's video generation abilities show a leap from static images to dynamic, 3D video imagery, expanding creative potential. 🎥

Key Takeaways

• Google's Gemini 2.5 Pro AI model is making waves with its expert coding and reasoning abilities, demonstrated through a complex Rubik's Cube simulation. 🎲
• A new, seventh-generation TPU, dubbed Ironwood, offers 3600 times the performance of its predecessors, making it a powerhouse for AI applications. 💪
• Google is pushing forward with multi-agent systems, using an open-source framework to ensure seamless agent interoperability across platforms. 🤝
• The event introduced Imagine 3, Chirp 3, and LIA for state-of-the-art text-to-image, voice generation, and music creation tasks, positioning Google at the forefront of AI-generated media. 🎨
• AI energy efficiency is a major focus, with improvements ensuring that powerful models do not come with untenable energy costs. ⚡

Overview

Chapters

• 00:00 - 00:30: Introduction and Rubik's Cube Demonstration The chapter introduces a Rubik's Cube simulation coded by developer Matt Berman, highlighting its complexity as a reasoning challenge. It features adjustable dimensions, scrambling options, and keyboard controls, all powered by Gemini 2.5 Pro. The innovation represents a major advancement in creating interactive code. The significance of this leap is underscored by a mention from the Google CEO at the Google Cloud Next event.
• 00:30 - 03:00: AI Announcements at Google Cloud Next The chapter discusses the recent AI announcements made at the Google Cloud Next event. Highlights include new advancements in AI technology such as agent interoperability, new text-to-video and text-to-image models, and new speech models. They also partnered with Box for agent-to-agent interoperability. The narrator humorously states that while they are on vacation, AI doesn't take breaks, implying the constant evolution and discussion around AI.
• 03:00 - 05:00: Gemini 2.5 Pro Capabilities The chapter introduces the concept of a new tensor processing unit (TPU) named Ironwood, which has been specifically designed to enhance AI infrastructure. It is announced as the seventh generation TPU from Google, promising to offer a significant boost in performance, achieving 3600 times better performance compared to the first publicly available TPU. Ironwood is set to be the most powerful chip released by the company and is expected to launch later this year.
• 05:00 - 08:00: Agent Creation and Interoperability This chapter explores the advancements in AI model infrastructure, emphasizing the development of high-performance chips. These chips have shown significant improvements in performance, measured in flops, and are 29 times more energy-efficient compared to previous generations. The focus on power efficiency is highlighted as being equally crucial to performance in the evolution of AI technology.
• 08:00 - 09:00: Demonstration of Agent-to-Agent Communication The chapter discusses a significant constraint on AI advancement in the United States, focusing on the energy limitations that impact the development and operation of next-generation AI applications. Efficiency improvements in both hardware and software are emphasized as pivotal solutions. Furthermore, the chapter shifts to highlight the capabilities of the Gemini 2.5 Pro model, renowned for its exceptional coding and reasoning abilities. The chapter concludes by mentioning a specialized test, seemingly named Ella Marina, that the model was subjected to, although further details on this test are not provided in the transcript.
• 09:00 - 16:00: Media Models: Text to Image, Voice, and Video The chapter emphasizes the development and accomplishments of AI models in text-to-image, voice, and video technologies. A significant highlight is the introduction of Gemini 2.5 by Eric Hartford, touted as the most intelligent AI model developed. The model is recognized for its advanced reasoning abilities, allowing it to process and think through its responses before interacting. It has gained recognition as the best model in the world according to the chatbot Arena Leader.
• 16:00 - 16:30: Conclusion and Reflections on Google’s Advancements Google has achieved state-of-the-art results across various benchmarks of advanced reasoning, including achieving the highest score on an industry benchmark known as humanity's last exam, which tests the frontier of human knowledge and reasoning.

Google Cloud Next - Gemini 2.5 Pro EVERYWHERE Transcription

• 00:00 - 00:30 take a look at this Rubik's cube coded by developer Matt Berman you might think of it as a toy but it's actually a really complex reasoning challenge adjustable dimensions scrambling the squares keyboard controls and Gemini 2.5 Pro can simulate it all it's a significant leap and shows the ability to produce robust interactive code all right the Google CEO just talked about our Rubik's Cube simulation done with Gemini 2.5 Pro at the Google Cloud Next
• 00:30 - 01:00 keynote that's pretty awesome and that event just finished it was all about artificial intelligence they made some incredible announcements including some new agent stuff new text to video text to image some speech models and even agentto agent interoperability in which they partnered with our partner for this video Box so we're going to go over all of it right now and if you're wondering no that is not an AI background i'm on vacation but AI doesn't take vacation so I guess neither do I all right the first
• 01:00 - 01:30 thing they're going to talk about is a new tensor processing unit this is a chip specifically designed to run their AI infrastructure and so let's take a look today I'm proud to announce our seventh generation TPU Ironwood is coming later this year compared to our first publicly available TPU Ironwood achieves 3600 times better performance an incredible increase it's the most powerful chip we have ever
• 01:30 - 02:00 built and will enable the next frontier of AI models all right that is an insanely fast chip obviously it's just relative to the previous generations of chips but as you can see on the Yaxis this is performance in terms of flops so a massive massive improvement there the same period we've also become 29x more energy efficient and am mean will share more later today and power efficiency is just as important as performance because
• 02:00 - 02:30 one of the limiting factors especially in the United States for AI is energy we simply don't have enough energy to power the next generation of AI applications so the more efficient we can get both with hardware and software the better it's going to be all right next he's just going to show off some of the qualifications of Gemini 2.5 Pro which as you know is an absolutely incredible coding and reasoning model and he talks about things like Ella Marina and as I mentioned at the beginning of this video a test that we put it through which
• 02:30 - 03:00 really lends an air of legitimacy to our little community as Eric Hartford said "A couple weeks ago we released a new model Gemini 2.5 a thinking model that can reason through its thoughts before responding it's our most intelligent AI model ever and it's the best model in the world according to the chatbot Arena Leader
• 03:00 - 03:30 it's state-of-the-art across a range of benchmarks requiring advanced reasoning that included the highest score ever on humanity's last exam one of the hardest industry benchmarks that's designed to capture the human frontier of knowledge and reasoning there's a lot of impressive words but let me show you what it can do take a look at this Rubik's cube coded by developer Matt Berman you might think of it as a toy but it's actually a really complex reasoning challenge adjustable
• 03:30 - 04:00 dimensions scrambling the squares keyboard controls and Gemini 2.5 Pro can simulate it all it's a significant leap and shows the ability to produce robust interactive code now the thing that he didn't mention which I'm really surprised is Gemini 2.5 Pro did this with one try there was zero iteration it was zero shot i had no examples i simply prompted it and it gave me that i loaded it up in my code editor and there it was
• 04:00 - 04:30 and yeah he kind of skipped right over that which in my opinion is the most impressive part of that demo all right and next they're going to announce a faster version of it Gemini 2.5 Flash and by the way if we're at 2.5 imagine what three is going to look like but anyways let's look at 2.5 Flash now Gemini 2.5 flash our low latency and most costefficient model with thinking builtin with 2.5 flash you can control how much the model reasons and balance
• 04:30 - 05:00 performance with your budget 2.5 flash is coming soon in AI Studio Vert.Ex AI and in the Gemini app we'll be sharing more details on the model and its performance soon i'm pretty excited by it and can't wait for you to see it for yourselves all right next we're going to talk about the thing I'm personally most excited about you know I'm bullish on agents they now have a new agent creation platform as well as agentto agent interoperability that means in the
• 05:00 - 05:30 future you will have your agent and your agent will be able to talk to other agents from other platforms other software and they will easily be able to communicate and work with each other this is really the underlying architecture needed to have this agentic future that we all know is coming we're announcing today a new agent development kit it is a new open-source framework all right and that is the key word that I love to see opensource it is an open-source framework now he does talk
• 05:30 - 06:00 about using Gemini models but if it is open source technically you should be able to use any model you want all right let's keep watching that simplifies the process of building sophisticated multi-agent systems now you can build sophisticated Gemini powered agents help them use tools do complex multi-step tasks including reasoning or thinking you can also discover other agents learn
• 06:00 - 06:30 their skills and enable agents to work together while maintaining precise control agent development kit supports the model context protocol that's huge so model context protocol is everywhere the Google CEO just about a week ago asked should we support it and obviously I think he already knew he was going to be doing it and everybody said yes so now Google Microsoft OpenAI Anthropic pretty much everybody is supporting MCP and I
• 06:30 - 07:00 love it standards are good for us which provides a unified way for AI models to access and interact with various data sources and tools rather than requiring custom integrations for each and every one all right next we're going to learn about the agentto aagent protocol which sounds incredible we're also introducing a new agentto aagent protocol that allows agents to communicate with each other regardless of the underlying model
• 07:00 - 07:30 and framework they were developed with this protocol is supported by many leading partners who share a vision to allow agents to work across the multi- aent ecosystem and with agents built on other agent frameworks including Langraph and Crew AI so happy they mentioned Langraph and especially Crew AI you know I'm a huge Crew AI fan and I'm really glad that these products are
• 07:30 - 08:00 all working together really well because having agents talk to each other that aren't built on the same system will be incredibly important and one of their launch partners is Box who partnered with us on this video all right so let me show you a demo of Google Agent Space which they also just announced this is kind of the UI of this agentto agent interoperability platform and this one is showing off Box and so check out how cool this is so let me show you so he's going to type in can you help me create a claim report and cost summary with my content in Box and my pricing database
• 08:00 - 08:30 in Google Cloud two different platforms all in one place now so you can see one of the data sources over here on the right is Box the other one is Big Query and with agent space agents from these two different platforms are going to be able to talk to each other and solve this problem together okay so we see it queried Box it queried Big Query it's putting them together can you please provide the claim ID all right so then the claim ID is provided continues to think and again it's using
• 08:30 - 09:00 tools that kind of tap into both of these platforms there's all the relevant documents some from Box some from Google and now it's putting it all together box AI agent has generated the report look at that we have photos and then it generates an incident report so it's finished you can send it to Box right there you can look at obviously the chain of thought that just happened so this is super cool i'm really excited i cannot wait to test this out myself and I also recommend checking out Box AI
• 09:00 - 09:30 because Box enables you to use AI to extract useful information from all of the documents that you're already storing on Box they are compatible with leading models including Gemini 2.5 Pro and have a very easy to use API that you can build on top of right now and they handle the full rag pipeline for you it is dead simple they're trusted by 115,000 enterprise organizations with enterprisegrade security compliance and governance so check out Box i'll drop a
• 09:30 - 10:00 link down in the description below okay next they're going to talk about Imagine 3 which is their newest text to image model that really has incredible quality and then they also introduced Chirp 3 which is their voice generation model you just need 10 seconds of example audio and then you can generate it obviously this is a competitor to 11 Labs they also talk about LIIA which is text to music so really Google is going all in on all different types of media let's take a look over the last year we've made huge improvements to Imagine
• 10:00 - 10:30 3 our highest quality textto image model which generates images with better detail richer lighting and fewer distracting artifacts than previous models imagine delivers accurate prompt adherence bringing your creative vision to life with incredible precision we also introduced Chirp 3 to help you create custom voices with just 10
• 10:30 - 11:00 seconds of input and to weave AI powered narration into your existing recordings today we're making LIA available on Google Cloud to transform text prompts into 30-second music clips and with the first hyperscaler to offer this capability let's hear a clip from LIA but in my opinion V2 is the most
• 11:00 - 11:30 impressive you give it an image and it will generate a video from that image but not only that you can give it direction you can say pan across the screen or zoom in and it looks incredible and from that single image you're getting a 3D video look at this v2 is our industry-leading video generation model it generates many minutes of 4K video watermarked with synth ID to ensure they can be
• 11:30 - 12:00 identified as AI generators it gives creators unprecedented creative control with new editing tools including camera presets to direct shot composition and camera angles without complex prompting first and last shot control to define the beginning and the end of a video sequence with VO seamlessly bridging the gap and dynamic inpainting and outpainting for video editing and
• 12:00 - 12:30 scaling with Gemini Imagine Chirp LIA and VO google is the only company that offers generative media models across all modality and all of them are available to you today on Vert.Ex AI all right now check out this live demo that they're going to do using V2 it is so cool we're going to generate video but here's the new hotness check it out camera presets
• 12:30 - 13:00 built right into VO panning left panning right timelapse tracking shots and even drone shots so let's go ahead and submit a drone shot of the city skyline there we go we'll go and submit this now normally this would take a few seconds i ran this earlier today so it's cached so it's going to be a little quicker than normal all right let's look at video number one absolutely spectacular
• 13:00 - 13:30 we have the ability to see the fountains the Eiffel Tower now let's go ahead and take a look at video number two a different angle that VO creates for us again stunning imagery you can see the clouds in the background and look at the cars driving up and down Las Vegas Boulevard absolutely incredible now one video is not going to do it for the concert promo
• 13:30 - 14:00 we want to do so I want to show you some of the other videos that I created i have one here of the stage being set up all through the power of VO i even have one of the audience actually clapping for what they're about to see this will be a good reminder for all of you now something very interesting happened it turns out that VO can do something that my 12-year-old can do and that is be an expert in
• 14:00 - 14:30 photobombing it turns out that this great video we just saw has a crew member and we love our crew members however in this case I'd like to feature the guitar because the guitar is the most important part of the band so let's go ahead and use VO's new inpainting capability and I'm sorry sir i apologize i know you're very good at your job but I am going to have to remove you from this image we will send flowers to you and
• 14:30 - 15:00 your family though sir let's use the new inpainting capability wait a couple of seconds and let's see what we see now if this does what I think it does it should preserve every single aspect of what we saw before just without our stage hand look at that so a lot of major announcements google is absolutely on fire since
• 15:00 - 15:30 launching Gemini 2.5 Pro i think they saw that and they were like "Oh my god I think we jumped in the lead." And now they are just firing on all cylinders if you would have asked me even 6 months ago if that were the case I probably would have said no but here we are and Google now has the best model on the planet so that's it for today if you enjoyed this video please consider giving a like and subscribe