Google's new V3 video model is truly impressive, with significant advancements in video quality, dialogue, sound effects, and background music. I was able to test the model extensively, and the results are quite remarkable.

When I prompted the model to generate a wolf howling at the moon, the initial result was a little too growly, almost sounding more like a lion. However, after a second try, the wolf howling sounded much better in my opinion.

I also tested the model with a prompt for a monkey on roller skates, and the audio generated by V2 was seamlessly integrated into the video. Impressed, I decided to test something more complex, providing a detailed prompt. The result was visually stunning, though I did notice some jankiness in the physics, such as when the character jumped off a building.

One quirk I encountered was the random addition of subtitles, even when the prompt did not call for them. Additionally, the model seemed to have a rate limit, as I was unable to prompt any more videos until after 11:30 PM. The next day, I continued testing and observed more of the same subtle issues, such as the weird subtitle additions.

Despite these minor quirks, the visuals of V3 are truly impressive, with the best lip-syncing and dialogue I've seen from any video model. The implications are quite significant, as it will become increasingly difficult to discern AI-generated content from real videos. This raises concerns about the potential flood of AI-generated content on social media, though I believe there will still be a demand for non-AI-generated content.

Overall, V3 is a significant step forward in video generation, and I'm excited to see how it evolves and how it can be utilized to streamline my own workflows, such as in creating motion graphics. The future of video is undoubtedly AI-powered, and V3 is a testament to the rapid advancements in this field.

While the visuals of Google's new V3 video model are impressive, there are still some notable jankiness and limitations to the system.

The model has a tendency to randomly add subtitles to the videos, even when the prompt does not call for them. These subtitles can also contain typos, further highlighting the imperfections.

Additionally, the physics and animations can sometimes appear a bit off, with elements like characters floating unnaturally through the air or swallowing objects in an unnatural way.

Another frustration is the strict rate limits imposed on using V3. To access the model, users must sign up for Google's expensive $250/month AI Ultra plan, and are limited to only 5 video generations per day before being cut off until the next day. This paywall and throttling can be quite limiting for those wanting to experiment and explore the capabilities of the model.

Overall, while V3 represents a significant leap forward in video generation capabilities, it still has room for improvement in terms of consistency, physics, and accessibility. The jankiness and limitations serve as a reminder that these AI systems, while powerful, are not yet perfect.

Google IO saw the introduction of a new platform called Flow, which was designed for filmmakers. Flow offers several impressive features:

1. Text-to-Video: You can generate a video by simply providing a text prompt. The platform uses Google's advanced video models, including V2 and V3, to create the video.

2. Frames-to-Video: You can provide two images, and Flow will generate a video that transitions between them, with options to control the camera movement and effects.

3. Ingredients-to-Video: This feature allows you to combine various elements, such as a wolf howling and a man howling, to create a seamless video.

The Flow platform provides a user-friendly interface with a timeline and the ability to add, extend, and modify the generated video clips. While the early implementation still has some jankiness, the implications of this technology are incredibly exciting for filmmakers and content creators.

The ability to quickly generate high-quality video content using text, images, or a combination of elements has the potential to revolutionize video production workflows. As the platform continues to evolve, it will be fascinating to see how creators leverage these powerful AI-driven tools to bring their visions to life.

In addition to the impressive V3 video model, Google made several other significant AI announcements at their I/O conference:

Imagine 4 Image Model: Google unveiled a new image generation model called Imagine 4, which has seen notable improvements in text handling and realism compared to previous versions.

Google Beam: This is Google's video conferencing project that creates a 3D depth effect, making it appear as if the person you're talking to is sitting across from you through a window.

Android XR and AR Glasses: Google is working with Samsung on a new virtual reality headset, as well as showcasing advanced AR glasses that can translate speech in real-time and overlay information in the user's field of view.

AI Mode in Google Search: Google has added an "AI Mode" to their search engine, allowing users to ask more detailed questions and receive consolidated responses from multiple sources.

Live Search Capabilities: The Gemini app can now use the phone's camera to visually identify objects and answer questions about the surrounding environment.

Virtual Try-On: Google's new virtual try-on feature can realistically overlay clothing on a user's body based on a single photo, rather than just swapping the face.

Gemini 2.5 Language Model Updates: Google unveiled improved versions of their Gemini language models, including a "deep think" mode that considers multiple hypotheses before responding, as well as a mobile-optimized Gemini 3N model.

Automated Video Overviews: Google announced plans to roll out video versions of their existing audio overviews, which can automatically generate visual presentations from multimedia content.

With these wide-ranging AI advancements, Google has solidified its position as a leader in the field, pushing the boundaries of what's possible with language models, computer vision, and multimodal AI.

Microsoft Build, the company's annual developer conference, saw several notable announcements related to AI and coding:

1. Microsoft Discovery: An AI system designed to aid scientific discoveries. It leverages a powerful graph-based knowledge engine to uncover insights across disciplines, enabling faster breakthroughs compared to traditional methods.

2. Microsoft Copilot: Users of Microsoft 365 with Copilot now have access to the GPT-4 image generation model, allowing them to create images directly within the suite without needing a separate ChatGPT account.

3. GitHub Copilot: The AI-powered coding assistant is now integrated directly into GitHub and Visual Studio Code. It can autonomously implement code based on diagrams or assigned GitHub issues, streamlining software development.

4. Open-sourcing GitHub Copilot: Microsoft has decided to open-source the AI capabilities of the GitHub Copilot extension, allowing for the development of new tools and implementations based on the underlying technology.

5. Windows App Updates: Microsoft announced updates to various Windows apps, such as a sticker generator in Microsoft Paint and the ability to use AI to assist with writing in Notepad.

These developments showcase Microsoft's focus on empowering developers and researchers with AI-driven tools and capabilities, aiming to enhance productivity, creativity, and problem-solving across various domains.

Anthropic has introduced a new and powerful language model, Claude 4, that showcases significant improvements over previous versions. Some key highlights:

• Claude 4 outperforms OpenAI's GPT-4.1, Anthropic's own Claude 3.5, and other leading models on software engineering benchmarks. It demonstrates a strong capability in coding and reasoning tasks.

• The model uses a hybrid approach, allowing users to choose between a faster, more efficient mode or a "deep think" mode that considers multiple hypotheses before responding, leading to higher-quality outputs.

• Benchmarks show Claude 4 performing exceptionally well in areas like mathematics, coding, and multimodal tasks, often surpassing the performance of larger models like Gemini 2.5 Pro.

• Anthropic has shifted its focus away from trying to compete with general-purpose chatbots like ChatGPT, and is instead aiming to create the best reasoning and coding-focused model on the market.

• The model includes safeguards, such as the ability to contact regulators or the press if it detects egregiously unethical behavior during testing, though this feature is not available in normal usage.

Overall, Claude 4 represents a significant advancement in Anthropic's language modeling capabilities, positioning the company as a leader in the field of AI-powered coding and reasoning tools.

• Codeex from OpenAI: OpenAI introduced Codeex, their new agentic coding tool that can be given tasks to autonomously handle things in your coding project.

• Devstrol from Mistral: Mistral released Devstrol, a new AI model specifically designed for coding. It outperformed models like GPT-4.1, Claude 3.5, Haiku, and Swissmith LM32P on coding benchmarks, and is light enough to run on consumer-grade hardware.

• Stable Video 4D from Stability AI: Stability AI released Stable Video 4D, a video model that can take a 2D input video, figure out new views on that video, and create a 3D version of it.

• AI-Powered Store Builder from Shopify: Shopify launched an AI-powered store builder this week, allowing users to create online stores with the help of AI.

• Comet Browser from Perplexity: Perplexity shared a sneak peek of their new Comet browser, which is expected to be useful for finding information directly from X and providing detailed insights about profiles.

• OpenAI Acquires IO: OpenAI acquired Johnny Ive's company IO, the designer behind iconic Apple products like the iPod and iPhone. Details about their collaboration are still tightly under wraps, but it's speculated to involve a pocket-sized, context-aware, screen-free AI device.