Meta's Llama 4 is here

Welcome back! Here’s what you need to know about AI today:

👉 Meta launches a new herd of open-source Llama models

👉 Microsoft is turning Copilot into a personal AI companion

👉 OpenAI has plans to launch o3, delaying GPT-5

and many more!

📧 Did someone forward you this email? Subscribe here for free to get the latest AI news everyday!

Read time: 5 minutes

What’s going on: Meta has launched Llama 4, its latest series of open-source AI models. This family of models includes Llama 4 Scout, Llama 4 Maverick, and the still-in-training Llama 4 Behemoth. These models were developed in response to competitive pressure from Chinese AI lab DeepSeek, which had begun outperforming Meta’s earlier Llama iterations. Meta AI, the company’s assistant integrated across WhatsApp, Messenger, and Instagram, now runs on Llama 4 in 40 countries, though its multimodal features, handling text, images, and more are currently limited to the US in English.

What does it mean: The Llama 4 models are designed for efficiency and versatility, leveraging a mixture of experts (MoE) architecture to optimize computational resources. They were trained on vast datasets of text, images, and videos, enhancing their ability to process and generate multimodal content. Meta has also adjusted their behavior, tuning them to respond more often to controversial political and social questions with what the company calls “balanced” answers, a shift from previous models that dodged such topics.

More details:

• Scout and Maverick are available now on llama.com and through partners like Hugging Face, while Behemoth remains in the works.

• While Scout and Maverick are designed for developers and businesses, with capabilities suited for tasks like chatbots and content generation, Behemoth is aimed to be the heavyweight once completely trained.

• The Llama 4 license prohibits EU-based entities from using or distributing the models.

• Interested in the details? Read Meta’s official blog post.

• Want to try these models? Check out this directory on Hugging Face.

What’s going on: Microsoft announced a significant update to its AI assistant, Copilot, enabling it to browse the web and execute tasks on behalf of users. Previously limited to static knowledge and basic functions within Microsoft’s ecosystem, Copilot now integrates real-time web search, pulling data from sites and services to answer queries with up-to-date information.

What does it mean: Beyond retrieval, it can perform actions like booking flights, sending emails, or managing calendar events, positioning it as a more proactive tool. This upgrade, available across Windows, Microsoft 365, and mobile apps, shows Microsoft’s aim to keep pace with competitors like Google’s Gemini and OpenAI’s ChatGPT, which have long offered similar capabilities.

More details:

• The enhanced Copilot relies on a combination of Microsoft’s AI models and partnerships, notably with OpenAI, to process web data and user instructions.

• It can summarize articles, compare prices across sites, or draft messages based on context from the web, all while maintaining integration with tools like Word and Teams.

• Microsoft claims web interactions are encrypted and user data isn’t stored beyond the session. The rollout starts with English-speaking users, with plans to expand languages and regions by mid-2025.

• Interested to learn more about the details? Read Microsoft’s official announcement.

🔀 OpenAI has reversed its earlier decision to cancel the standalone release of its o3 reasoning model, announcing plans to launch o3 and its successor, o4-mini, within weeks, while delaying the release of GPT-5 to "a few months".

💸 Google's Gemini 2.5 Pro, its latest advanced AI reasoning model, offers industry-leading performance on coding, reasoning, and math benchmarks, with competitive pricing starting at $1.25 per million input tokens and $10 per million output tokens for prompts under 200,000 tokens, which makes it the most expensive Gemini model.

👨‍💻 GitHub Copilot is introducing rate limits and charges for users who want to use premium AI models like Anthropic's 3.7 Sonnet for tasks such as agentic coding and multi-file edits, with options to purchase additional requests or upgrade to a new Copilot Pro+ plan.

📃 A new study from researchers at the University of Washington, the University of Copenhagen, and Stanford suggests that OpenAI's AI models, including GPT-4, may have memorized copyrighted content from sources like fiction books and New York Times articles, supporting allegations that the company trained its models on copyrighted material without permission.

ChatGPT adoption is rapidly increasing in India, with over 20% of Android app downloads this year originating from the country, but monetization is lagging behind the US due to the high cost of subscriptions.

Microsoft has released an AI-generated demo of Quake II to showcase the gaming capabilities of its Copilot AI platform, but the company admits the experience has limitations, such as fuzzy enemies and object permanence issues.

AI + Email marketing

Google - Senior Data Scientist, Search AI Frontier Experience

Meta - Computer Vision Engineer, Reality Labs

Fanatics - Senior/Staff Data Scientist

OpenAI - Research Engineer

Thank you for staying with us like always! If you are not subscribed, subscribe here for free to get more of these emails in your inbox! Cheers!