While the AI world was busy playing catch-up, Meta quietly made history with the release of Llama 4 Scout. This model can hold an astounding 10 million tokens in memory, which is the equivalent of uploading 94 novels all at once. In comparison, Google's Gemini 2.5 model, previously considered a big deal, could only handle 2 million tokens - a mere fraction of Scout's capabilities.

The significance of this memory capacity cannot be overstated. Llama 4 Scout can not only hold this vast amount of information but can also remember and answer questions based on it, without the need for scrolling or summarizing. This is a massive leap forward, as it allows for instant, crystal-clear answers to queries, even when the relevant information is buried within a vast amount of text.

But Llama 4 is not just about size. Meta also introduced Llama 4 Maverick, a larger model in terms of training data, but with a smaller memory capacity of 1 million tokens. This is still enough to hold the entire Harry Potter series, making it a formidable model in its own right.

The most impressive model, however, is the yet-to-be-released Llama 4 Behemoth. This model is said to be trained with a staggering two trillion parameters, making it the largest model we've ever heard of, at least publicly. It's clear that Meta is pushing the boundaries of what's possible with large language models, and the implications of these advancements are yet to be fully realized.

While the AI world was busy playing catch-up, Meta quietly made a significant leap with its Llama 4 models. One of these models, Llama 4 Scout, has made history by being able to hold an astounding 10 million tokens in memory. To put this into perspective, that's the equivalent of uploading 94 novels all at once and being able to remember and answer questions based on all of them, without the need for scrolling or summarizing.

This is a massive leap compared to Google's Gemini 2.5 model, which could only handle 2 million tokens. Llama 4 Scout's ability to pass the "needle in a haystack" test, where researchers bury facts inside huge texts, and its crushing performance on the benchmark chart, further solidifies its revolutionary capabilities.

But Llama 4 Scout is not the only impressive model in the Llama 4 lineup. Llama 4 Maverick, a larger model in terms of training data, can still handle 1 million tokens, which is enough to load the entire Harry Potter series. And then there's the highly anticipated Llama 4 Behemoth, which is said to be trained with a staggering two trillion parameters, making it the largest model we've heard of publicly.

However, the launch of Llama 4 has not been without its share of drama. Allegations of Meta training the models directly on benchmark tests, essentially teaching them the answers before the tests, have sparked a wildfire of controversy. While a Meta engineer has denied these claims, the performance of the public models has been called into question, with Llama 4 Maverick and Scout dropping in rankings on the LM Marina leaderboard.

Despite the ongoing drama, the sheer capabilities of Llama 4 Scout's long-term memory and the potential of the Llama 4 lineup as a whole are undeniable. As the AI landscape continues to evolve, the race for the most powerful and versatile models is heating up, and Meta's Llama 4 series is undoubtedly pushing the boundaries of what's possible.

While Meta's Llama 4 models made headlines, the launch was not without controversy. One of the Llama 4 models, Llama 4 Scout, made history by being able to hold 10 million tokens in memory, a massive leap from Google's Gemini 2.5 model which could only handle 2 million tokens. However, soon after the launch, someone claiming to be a former Meta employee alleged that the Llama 4 models' performance was inflated, as they were allegedly trained directly on benchmark tests, essentially teaching the models the answers before the tests.

This claim sparked a wildfire, with another Meta engineer denying the allegations and claiming that the public models were underperforming due to poor integration from third-party tools. However, the drama continued when the Llama 4 Maverick model, which initially ranked just behind Gemini 2.5 Pro on the LM Marina leaderboard, later dropped to the 32nd position, and the Llama 4 Scout model fell out of the top 30 altogether. LM Marina stated that Meta had submitted a fine-tuned version of Maverick, optimized for human preference tests, but that was not the version regular users were getting.

This incident highlights the ongoing issues of transparency and trust in the AI industry. As the race for AI dominance continues, it is crucial that companies are transparent about their models' capabilities and the processes used to train them. The Llama 4 controversy serves as a reminder that the AI community must prioritize trust and accountability to ensure the responsible development and deployment of these powerful technologies.

Microsoft celebrated its 50th anniversary in Seattle, bringing together industry legends like Bill Gates, Steve Ballmer, and Satya Nadella. Alongside the celebrations, Microsoft also unveiled some significant AI updates.

Microsoft's main AI assistant, Co-Pilot, now has the ability to remember personal details you share, such as your dog's name, fitness goals, or even information about a stressful project you're working on. This added memory feature enhances Co-Pilot's capabilities and makes it more personalized to the user's needs.

Furthermore, Microsoft showcased something truly remarkable - an AI-generated version of the classic game Quake. Every frame of the game is created in real-time using Microsoft's Muse model, which can be experienced at the website copilot.microsoft.com/wam. While some developers expressed distaste for this technology, the co-creator of the original Quake, John Carmack, defended it, stating that AI tools have always replaced old methods, and this is just the beginning of a new chapter for creators.

These advancements from Microsoft, including the enhanced Co-Pilot and the AI-generated Quake, demonstrate the company's continued efforts to push the boundaries of AI and its applications in various domains, from personal assistants to creative tools.

While the AI world was focused on the drama surrounding Meta's Llama 4 models and Microsoft's 50th anniversary celebrations, Google quietly made some significant advancements in their AI capabilities. These upgrades span across their search, cloud, and workspace offerings, showcasing Google's strategic and methodical approach to shaping the future of AI.

In Google Search, the AI-powered experience has received a major upgrade, becoming better at handling comparisons, how-to guides, and answering more open-ended, creative questions. The most remarkable feature is the ability to recognize book titles from a photo of a bookshelf and provide personalized recommendations, demonstrating the power of multimodal search.

Moving to Google Cloud, the announcements at their flagship event, Google Cloud Next 25, were equally impressive. One of the most futuristic developments is the introduction of the agent-to-agent (A2A) protocol, which allows AI agents to communicate with each other and execute tasks independently. This means that a user can simply request a task, such as finding flights under $300 for a weekend trip, and the AI agents will handle the entire process without the user's direct involvement.

Google has also made significant improvements to their Workspace offerings. Within Google Docs, users can now access audio summaries, a "help me refine" tool to polish their writing, and smarter Sheets with new AI analysis tools. Additionally, the Gemini feature in Google Meet can now summarize meetings and allow users to ask questions about them afterward, saving time and increasing focus.

Beyond these user-facing features, Google has also expanded their Vertex AI platform, introducing Chirp 3 for audio generation, Imagen 3 for text-to-image generation, and LIA for generating music from text prompts. These advancements showcase Google's commitment to building a comprehensive AI ecosystem, spanning business workflows, creative tools, and more.

While others in the AI space may be grabbing headlines with dramatic leaderboard battles or high-profile announcements, Google's approach is more subtle yet powerful. They are steadily and strategically building out their AI capabilities, transforming the way we interact with search, collaborate in the cloud, and enhance our productivity in the workspace.

Google's quiet but strategic AI advancements have set the stage for the next phase of how we work and create. From smarter search capabilities that understand multimodal inputs to AI agents that can collaborate autonomously, Google is building a comprehensive AI ecosystem.

While others have been making headlines, Google has been steadily rolling out powerful upgrades across its products and platforms. The new audio summaries in Google Docs, the advanced analysis tools in Sheets, and the generative AI capabilities in Vertex AI all demonstrate Google's commitment to delivering practical, user-focused AI solutions.

As the AI landscape continues to evolve rapidly, now is the time to get ahead of the curve. The 5-day AI challenge provides a valuable opportunity to dive into the latest tools, workflows, and real-world use cases, equipping you with the knowledge and skills to thrive in this transformative era. With the $1,500 giveaway closing in just 2 days, there's no better time to join and position yourself for success in the AI-powered future.