5/24/2025 AI Express | What's New in AI: An Overview of Veo 3 Precision Control and Multi-Model Updates

Cripes!!! The Veo 3 can be precisely controlled by the second. ！！！！ What more bike!!! Epoch-making friends!!!

Recently, the ZH3O team released a test video of the Veo 3, demonstrating its ability to accurately execute commands over different time periods. This technological breakthrough not only enhances the flexibility of content creation, but also provides creators with unprecedented control precision.

Here are the test results:

1. 0-3s: Fighting (✅)
2. 4s: freeze frame (✅)
3. 5-6s: Back in Time (✅)
4. 7s: turn + foresight (❌)
5. 8s: rain + go + text (✅)

This feature of Veo 3 enables creators to control every movement and scene in a much finer detail, greatly enhancing the presentation of visual effects. The realization of this second-by-second control signals a more efficient and high-quality video production in the future.

For more detailed information, please visit:https://t.co/uuP3d1iTis

Recently, I thought about the choice of mechanism regarding attention. When using chunked attention (chunked attention), the sensory field does not expand with increasing depth, which means that the 8192nd token cannot communicate with the 8193rd token.

This reminds me of some previous research that compared a variety of attention mechanisms, including local attention, which is actually chunked attention. This kind of chunked attention makes it easier to write parallelized code when performing contextually parallel processing.

Llama 4 The latest version of iRoPE has made significant progress in this area. Its industry-leading 10M+ multimodal context length (over 20 hours of video) is a major breakthrough. iRoPE architecture ("i" stands for interleaved layers, infinite) plays an important role in realizing the goal of infinite context over the long term.

The architecture solves the long context problem in the following way:

• Locally parallelizable chunked attention models only short contexts (e.g., 8K).
• The global attention layer, on the other hand, is responsible for modeling long contexts (e.g., >8K) without the use of positional embedding, thus improving extrapolation. The maximum training length is 256K.

To compensate for the flattening of the attention weights as the context increases, we apply temperature scaling of the inference at the global layer to enhance long-range inference while maintaining performance for short contexts (e.g., 8K):

xq *= 1 + log(floor(i / α) + 1) * β # i = position index

We believe in the importance of open research and will be sharing more technical details via podcast in the near future.

Heavy news: Veo 3 is now rolling out to Gemini Pro users in the US.

This update marks the further expansion of Veo 3 on the Gemini platform, providing more powerful features and a better user experience for professional users.Veo 3 is a tool designed for professional users to enhance their capabilities in data processing, analysis and visualization.

Key features include:

• Enhanced data processing capabilities to support more complex data sets
• More advanced analytical tools to help users quickly discover key information in their data
• Intuitive visualization interface for clear and easy-to-understand data presentation

In addition, Veo 3 offers a range of customization options that allow users to personalize their settings to suit their needs.Gemini Pro users can check out the detailed changelog and installation guide on the Gemini website to ensure a smooth transition to the new version.

With the launch of Veo 3, Gemini continues to consolidate its leadership position in data analytics, offering more efficient and smarter solutions for professional users.

Veo 3's crazy street interview video reveals confusion about the real and virtual worlds.

With the advancement of technology, especially artificial intelligence and deep faking, it has become difficult for humans to distinguish the real from the virtual. This phenomenon has sparked widespread discussion, and many have begun to wonder if the world we live in is also a highly simulated environment.

The human world has been highly simulated.There is reason to believe that we may be living in a multilevel simulation. This "nesting" theory of simulation suggests that the world we live in may be the result of simulations of higher-level intelligences in which we further create new virtual worlds.

This idea has attracted widespread attention not only in science fiction, but also in philosophy and science. For example, Nick Bostrom's "simulation hypothesis" suggests that if a civilization were able to reach a high enough level of technology, they could create a large number of simulated worlds, making the probability of us living in a simulation very high.

Veo 3's street interview videos show how the general public views and reacts to these complex issues, provoking more thoughts on the nature of reality.

The latest version of the ChatGPT web app is now live, with newmodel selectorfeature to provide users with more flexibility. This update allows users to select different language models for custom GPTs.For GPTs with customized network operations, only the GPT-4 and GPT-4.1 models are currently supported. If the user's GPT does not have a custom operation, any available model can be selected.

The update aims to enhance the user experience with more customization options. By selecting different models, users can adjust the style and quality of the generated content according to their specific needs. For example, GPT-4 is suitable for tasks that require high precision and complex reasoning, while GPT-4.1 performs better in handling certain specific tasks.

This feature is particularly suited for developers and content creators who can use these advanced models to create smarter, more personalized apps and services.

Recently.MistralAIAnnouncing a compact and powerful coding modelDevstral, the model is able to run efficiently on 24GB of video memory.

This model has excellent code generation capabilities, solves real GitHub problems, and integrates seamlessly into code agent frameworks. In terms of performance.DevstralSignificantly outperforms other open source models, demonstrating superior processing power and stability.

• Excellent performance:Significantly outperforms comparable open source models.
• Widely used:For single RTX 4090 graphics cards or Macs with 32GB of RAM, ideal for local deployments and enterprise use.
• Open Source Open:Uses the Apache 2.0 license, allowing the community to build and customize freely.

This breakthrough technology not only provides developers with powerful tools, but also opens up new possibilities for the use of AI in software development.

As more and more businesses and developers begin to focus on automation and intelligent tools, theDevstralIt promises to be an important part of future software development.

The official release of Gemini 3 has generated a great deal of interest in the Gemini series, which has long been known for its superior natural language processing and multitasking capabilities.

The release of Gemini 3 not only offers significant performance improvements, but also introduces a number of new features. These include more powerful contextual understanding, a smoother conversation experience, and a wider range of application scenarios.Gemini 3 is able to better understand and respond to user needs and provide more personalized services.

Key highlights include:

• enhanced natural language understanding
• More efficient multitasking
• Improved dialog fluency
• More diverse application scenarios

The release of Gemini 3 marks another major leap in AI assistant technology, which is expected to play an important role in several industries and sectors.

Meta researchers have introduced a new architecture that enhances the performance of large language models (LLMs) by adding trainable memory layers. These memory layers can efficiently store and retrieve relevant factual information without a significant increase in computation.

The team designed the memory key structure as a combination of smaller "half-keys," which significantly increased memory capacity and kept the system efficient. In tests, LLMs equipped with this memory layer outperformed unmodified models in several quiz benchmarks, even though they used much less training data.

This innovative approach not only improves the efficiency of the model, but also extends its application scenarios, e.g., in knowledge-intensive tasks. This is of great significance for domains that need to deal with large amounts of factual information, such as natural language understanding and intelligent question and answer systems.

At Google I/O 2025, the Gemini API received a number of important updates that further enhance its leadership in the multimodal intelligence platform.

This update includes:

• The new model with upgraded voice capabilities supports more than 24 languages, greatly expanding its range of applications.
• The real-time music generation feature opens up new possibilities in the field of music creation and entertainment.
• Multimodal input enhancements, especially in video understanding, make Gemini more efficient in processing complex visual information.
• Developer tools and API structure updates such as thought summaries, browser controls, and asynchronous function calls significantly enhance the developer experience.
• Providing developers with more efficient and cost-effective access to models, such as batch APIs, reduces the cost of large-scale deployments.

The Gemini API has gradually evolved into a complete multimodal intelligence platform for a wide range of scenarios, from code generation to audio dialog, from web page information extraction to browser operation control. Together with Google AI Studio, developers can rapidly prototype and test, build commercially available intelligent agents, and integrate multiple input and output methods such as voice, video, and image.

application scenarioIncluding but not limited to:

• Intelligent customer service system for globalized customer service through multi-language support and efficient voice recognition.
• Creative content generation, utilizing real-time music generation and multimodal input processing, provides users with a rich multimedia experience.
• Automated workflows, combined with tool scheduling and automated control interfaces, improve the operational efficiency of your organization.

These updates not only enhance the Gemini API, but also provide developers with more opportunities to innovate.

II-Agent's goal is to build what is probably the best intelligent agent in the world and open source it. This openness will not only help advance AI technology, but will also allow more developers and researchers to get involved and work together to advance human intelligence as a whole.

II's mission is to enhance human intelligence, and II-Agent, as part of its latest build, is designed to be a key component of that goal. It will meet or exceed existing intelligent agent benchmarks through top-notch technology and innovative features.

Here are some improvements and new features suggested for II-Agent:

• Enhanced natural language processing capabilities for smoother and more natural conversations.
• Integrate more practical application scenarios, such as customer service and educational counseling.
• Optimize self-learning and adaptive mechanisms to excel in different environments and tasks.

Through these improvements, II-Agent will not only serve users better, but also bring great value to the whole society. Ultimately, everyone will benefit.