# Google I/O 2025: artificial intelligence at the heart of innovation and the future

Author: Fernando Nieto Lobato
Original publication: 2025-05-28
Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-87-el-retorno-de-google
English URL: https://elcontemplador.github.io/estrategia-english/essays/087/
Status: Published translation

This is a translation of the original Spanish essay published on 28 May 2025. Its claims, examples and forecasts retain that historical context.

English publication: 2026-09-29

*Historical context: this article was published on 28 May 2025. Product availability, comparative rankings and forecasts reflect that moment.*

[Google I/O 2025](https://io.google/2025/), held on 20 May, marked a turning point in Google’s trajectory, placing AI not merely among its key technologies but at the centre of its strategy, as the driving force behind all its future innovations, with the aim of making AI very easy for everyone to use. It was also the first moment in many years when Google appeared to be leading AI worldwide, with products outperforming its competitors in almost every field. Google, although it may have spent years “asleep”, is the great pioneer of generative AI, [with its foundational 2017 paper on transformers](https://arxiv.org/pdf/1706.03762).

[![Google I/O presentation slide claiming first place across all LMArena categories; its complete rankings are transcribed below.](https://elcontemplador.github.io/estrategia-english/assets/images/88b17f88-ce66-42a5-8a35-1e2e19ded835_1600x855.png)](https://substackcdn.com/image/fetch/$s_!iFLQ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F88b17f88-ce66-42a5-8a35-1e2e19ded835_1600x855.png)

*Text of the slide: “#1 across all LMArena categories”. The following are the ranks displayed at the time of the presentation, not current rankings.*

| Model | Overall | Overall with style control | Hard prompts | Hard prompts with style control | Coding | Maths | Creative writing | Instruction following | Longer query | Multi-turn |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| gemini-2.5-pro-preview-05-06 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
| o3-2025-04-16 | 2 | 1 | 2 | 1 | 1 | 1 | 5 | 2 | 5 | 4 |
| chatgpt-4o-latest-20250326 | 2 | 3 | 3 | 3 | 2 | 5 | 2 | 3 | 2 | 1 |
| grok-3-preview-02-24 | 2 | 5 | 2 | 4 | 2 | 4 | 2 | 3 | 2 | 3 |
| gpt-4.5-preview-2025-02-27 | 4 | 3 | 2 | 3 | 3 | 2 | 2 | 2 | 2 | 2 |
| gemini-2.5-flash-preview-04-17 | 4 | 5 | 2 | 3 | 2 | 2 | 2 | 2 | 2 | 4 |
| deepseek-v3-0324 | 7 | 5 | 7 | 4 | 3 | 5 | 5 | 7 | 6 | 4 |
| gpt-4.1-2025-04-14 | 7 | 5 | 5 | 3 | 7 | 8 | 5 | 6 | 2 | 5 |
| hunyuan-turbos-20240416 | 7 | 13 | 5 | 6 | 7 | 10 | 4 | 7 | 5 | 4 |
| deepseek-r1 | 8 | 8 | 8 | 4 | 8 | 3 | 6 | 7 | 8 | 6 |
| gemini-2.0-flash-001 | 9 | 16 | 8 | 20 | 8 | 10 | 6 | 11 | 9 | 10 |
| o4-mini-2025-04-16 | 9 | 5 | 7 | 3 | 6 | 1 | 14 | 11 | 12 | 8 |
| o1-2024-12-17 | 10 | 8 | 7 | 5 | 8 | 5 | 7 | 7 | 7 | 11 |

## Gemini, the brain of Google’s new era

Gemini, Google’s foundational AI model, took centre stage with the official presentation of **Gemini 2.5 Pro**, a version that had been available for several weeks. [Our students were able to use it extensively in the experiment creating a political party and election campaign that we described last week](https://open.substack.com/pub/estrategiabyaleph/p/estrategia-86-un-ano-despues-creamos?r=2tyybn&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false). It offers improved reasoning and multimodal understanding. Features such as *Deep Think* mode for complex problems, the nimble Flash version optimised for fast tasks, and Gemini Live’s conversational interactivity seek to establish Gemini in our daily lives, integrating it into everyday tools such as Gmail, search and Chrome. This is a strategic move to make AI not only powerful but also ubiquitous and useful.

## New diffusion models for text and video: Gemini Diffusion and Veo 3

In content generation, Google presented two advances that struck us as especially significant: Gemini Diffusion for text, an incredible model for the technical progress it promises for the future; and Veo 3 for video, a genuine revolution that is already here.

- **Gemini Diffusion:** this new text diffusion model is extremely fast. In demonstrations, it generated text five times faster than Google’s fastest model to date, while maintaining strong performance on coding tasks. Unlike traditional language models, which generate tokens sequentially — and this is the particularly significant point — Gemini Diffusion refines noise over multiple steps, using a method inspired by image generation. It can thus begin with an approximation and improve it iteratively, making it especially useful for tasks that benefit from this approach.
- **Veo 3:** this is the latest generation of Google’s video model, capable of creating highly realistic videos with visual effects and, among its most notable new features, native audio generation. Veo 3 can therefore generate videos containing dialogue, ambient sounds and synchronised music directly, without additional editing steps. This puts Veo 3 a step ahead of competitors such as Sora or Kling. Although initial access to Veo 3 will be limited and require an AI Ultra subscription, **its potential** for film and video creation **is enormous**.

In the recommendation of the week section, we have collected some of the best creations users are already generating with Veo 3. They are extraordinary: do not miss them. This particular development has also driven an extraordinary increase in traffic to the application:

[![Similarweb chart of daily visits to Deepmind.Google from January to May 2025, showing a sharp rise near the end.](https://elcontemplador.github.io/estrategia-english/assets/images/bf75e6b3-df14-4f97-a626-b7aa008c90ff_700x574.png)](https://substackcdn.com/image/fetch/$s_!dK1N!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbf75e6b3-df14-4f97-a626-b7aa008c90ff_700x574.png)

*Chart text: “Daily visits to Deepmind.Google”; series: “Website Visits”; scope: “Worldwide | All Traffic”; source: Similarweb. The vertical axis runs from 0 to 900K in steps of 100K. Dates labelled on the horizontal axis run from 3 January to 19 May 2025, with the series continuing slightly beyond the final labelled date. Visits fluctuate mostly around 100,000, with some peaks below 200,000, before rising sharply to approximately 850,000 and then falling to approximately 780,000 at the end. These are visual estimates: the individual points have no printed values.*

*Editorial note: the recommendation section mentioned above is a separate part of the original Spanish newsletter and is outside this main-article translation.*

To complement Veo 3 and Imagen 4, Google’s improved image-generation model, Google introduced Flow. Flow is an AI filmmaking tool that allows users to design complete narratives, scene by scene, combining Veo, Imagen and Gemini to maintain visual and narrative coherence. It also includes tools to control the camera and transitions, helping give clips a more cinematic look.

## Redefining search and interaction

The search experience, Google’s flagship product and principal source of revenue, is also being reimagined as AI advances as a new way to find information. A new “AI Mode” will enable more complex conversational searches, while “AI Overviews” will provide AI-generated summaries at the top of results.

## Other notable announcements

Alongside these developments, Google I/O 2025 brought a wave of AI-powered innovations:

- **Project Astra:** an ambitious initiative seeking to create a universal AI assistant that can understand the context of the world around us, help with complex tasks, find information and even act on the user’s behalf.
- **Imagen 4:** the new version of Google’s image-generation model, with improvements in photorealism, cleaner detail and a wider variety of artistic styles, plus the promise of better spelling and typography in generated images. In our initial tests, however, it is not yet at the level of OpenAI’s model.
- **Gemma 3n:** an open, fast, efficient multimodal model, **designed to run without a network connection on devices such as phones**, laptops and tablets, supporting audio, text, images and video. I have already been able to test it on my own phone [with Google AI Edge Gallery](https://github.com/google-ai-edge/gallery). For us, this was another of the event’s major advances. It makes clear that, trailing the most powerful giant AI models by only a few months — or around a year in the case of phones — we will be able to have artificial intelligences at almost the same level running locally on our own devices, with the enormous advantages for privacy and control that this brings.
- **Google Beam, the evolution of Project Starline:** a new AI-based video communications platform using a state-of-the-art video model to transform 2D video streams into realistic 3D experiences in real time, aiming to create a more authentic sense of presence in video calls.
- **AI advances for developers:** given this newsletter’s nature and focus, we have chosen not to cover the enormous advances in programming, tools and technical protocols here. But it is worth highlighting the major competitive advantage Google gains from having its own “computers” to train models and serve them to users: its seventh-generation TPU, Ironwood, enables faster, more efficient models at competitive prices.

Ultimately, [Google I/O 2025](https://notebooklm.google.com/notebook/953b658a-579b-4b3c-b280-43b3781babf3) made clear that the company is fully committed to an AI-powered future in which artificial intelligence is not merely a tool but an essential “partner” in almost every aspect of our lives. The advances presented — especially Veo 3, already transforming the audiovisual industry completely, and Gemini’s evolution towards models leading competitors in benchmarks — will continue to redefine how we create, interact with information and communicate, marking a significant milestone in AI’s continuing, dizzying progress.

## Connecting the global vision with local impact: Google Cloud Summit Madrid 2025, which we will cover in the next issue

[![Google Cloud Summit Madrid event graphic, dated 22 May 2025.](https://elcontemplador.github.io/estrategia-english/assets/images/3dd8d35e-30e0-4f46-9691-f1ceb00b3178_1300x1200.png)](https://substackcdn.com/image/fetch/$s_!G7PD!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dd8d35e-30e0-4f46-9691-f1ceb00b3178_1300x1200.png)

*English text of the event graphic: “Google Cloud Summit”, “22 May 2025”, “Madrid”.*

The transformative vision and wave of AI innovations presented at Google I/O 2025 were not confined to global announcements from Mountain View. These advances have a direct echo and practical application in local settings, exemplified by **[Google Cloud Summit Madrid 2025](https://cloudonair.withgoogle.com/events/google-cloud-summit-madrid-2025)**, [held in Madrid on Thursday 22 May](https://cloudonair.withgoogle.com/events/google-cloud-summit-madrid-2025). To offer a first-hand perspective and analysis of what happened there, our next issue will feature a valuable contribution from [Gabriela Ortega](https://www.linkedin.com/in/gabrielaortegaj/), Director of Strategy at ALEPH Educational Institution, who attended the event in Madrid. Next week, she will tell us about her experience of getting to know Google’s approach to artificial intelligence first-hand.

[Fernando Nieto Lobato](https://www.linkedin.com/in/fernandonietolobato/)

*Director of Digital Innovation, ALEPH Educational Institution*