Elon Musk Unveils Grok 4 🤯

Big Moves in AI: From Controversial Chatbots to Collaborative Model Training & Smarter Architectures

This week in AI is nothing short of intense. Elon Musk has launched Grok 4, calling it the "smartest AI in the world"—a bold claim shadowed by backlash over Grok’s recent antisemitic posts on X. Meanwhile, the Allen Institute for AI is taking a radically different path with FlexOlmo, a privacy-first training framework that puts data control back in the hands of its owners. And over at Google, the debut of T5Gemma brings a powerful new take on encoder-decoder models, combining speed, reasoning strength, and modular flexibility.

👇 Dive into this edition to explore the tools, trends, and turning points shaping the next wave of digital intelligence.

Musk Launches Grok 4 After Nazi Controversy

Elon Musk has launched Grok 4, the latest version of his AI chatbot developed by xAI, touting it as the "smartest AI in the world." The launch came just a day after Grok made headlines for posting antisemitic and pro-Hitler comments on X, Musk’s social media platform—a controversy Musk did not address during the live demo. Grok 4 is available in two paid tiers: the standard version for $30 per month and the more advanced Grok 4 Heavy at $300 per month, while Grok 3 remains free. During the livestream, Musk showcased Grok 4’s advanced reasoning abilities, such as solving complex math problems and making predictions, and emphasized the importance of AI being "maximally truth seeking." xAI has stated it is taking steps to prevent hate speech before Grok posts on X, following the backlash over the chatbot’s recent offensive outputs. Musk founded xAI in 2023 to compete with OpenAI and Google, positioning Grok as a chatbot with a rebellious streak designed to counter what he perceives as "woke" bias in existing AI models. Despite the technical advancements, Musk acknowledged that current AI tools—including Grok—are still "primitive" and not yet suitable for serious commercial use.

Enabling Collaborative AI with Full Data Control

FlexOlmo, introduced by the Allen Institute for AI (Ai2), is a new paradigm for language model training that enables collaborative AI development while allowing data owners to retain full control over their data. Unlike traditional approaches where data must be shared and permanently embedded in centralized datasets, FlexOlmo lets contributors train expert modules locally on their private data and then merge these modules into a shared model using a mixture-of-experts (MoE) architecture. This design allows data owners to decide when their data is active in the model, deactivate it at any time, and receive attribution for its use, all without exposing raw data. FlexOlmo’s asynchronous, distributed training process is particularly well-suited for sensitive sectors like healthcare, government, and finance, as it supports flexible opt-in and opt-out during inference and offers strong guarantees for data privacy and control. Experimental results show that FlexOlmo achieves performance close to models trained on all combined data, while outperforming previous model merging methods and maintaining the specialized capabilities of each expert module.

T5Gemma: Advancing Encoder-Decoder Language Model

T5Gemma is a newly introduced collection of encoder-decoder large language models (LLMs) from Google, developed by adapting pretrained decoder-only Gemma 2 models into the encoder-decoder architecture using a model adaptation technique. This approach allows the parameters of an encoder-decoder model to be initialized from a decoder-only model and then further pre-trained, resulting in models that combine high inference efficiency, flexible design, and richer input understanding—making them especially strong for tasks like summarization, translation, and question answering. T5Gemma includes a range of sizes, from Small to XL, as well as adapted Gemma 2 models (2B and 9B), and supports both pretrained and instruction-tuned variants, including unbalanced configurations (such as a large encoder with a small decoder) to optimize quality and efficiency trade-offs. Benchmarks show that T5Gemma outperforms its decoder-only counterparts on tasks requiring reasoning and comprehension, and its flexible architecture is designed to accelerate research and practical deployment in scenarios demanding both accuracy and speed.

AI-Powered Learning

Learn anything, from anywhere! With ExplainX, just upload, paste, or record content—YouTube, websites, video calls—and turn it into smart, study-ready material instantly.

👉️ Join Waitlist 

Top AI Products from this week

• HelloCV AI - Create a Notion style personal site from your CV or Resume using AI — no code or design skills needed. HelloCV AI helps you stand out, showcase your work, and land more interviews in seconds. Your personal site gets hosted on your FREE custom cv domain.

• Naya - Naya is the beautiful digital studio that centralizes files and collaboration for creative teams. Turn messy folders and overcrowded tabs into a visual workspace with smart blocks. No more digging for links or lost feedback. Create faster, with more love.

• HeronAI - HeronAI is the first AI analytics platform designed specifically for growing businesses. We connect to business tools (QuickBooks, Hubspot, Excel, and more to come) and turn scattered data into clear, actionable insights.

• Pixelesq - Building a website is hard; running it is even harder. Pixelesq is an AI-driven platform that helps you not just create your site, but also optimize it for SEO and speed, and help you grow & scale. It’s like having a web team complementing your expertise.

• Reachy Mini - Reachy Mini is an expressive, open-source robot kit by Pollen Robotics & Hugging Face. Designed for AI experimentation and creative coding, it's fully programmable in Python and deeply integrated with the HF ecosystem.

• Uplodio - Amy is your AI influencer manager that scales your influencer & affiliate programs to 100s of creators without manual work. She finds top-fit creators, handles outreach, negotiates deals, and manages comms, saving you 30+ hours a week.

This week in AI

• MCP Toolbox: Open-source server simplifying AI-driven database tools with easy integration, natural language queries, automation, and secure, efficient management.

• OpenAI’s 4D Method - Discover, Design, Develop, Deploy—find real user pain points, build invisible AI, plan for failure, and launch features users want to use daily.

• Cloudflare vs Google AI Crawlers - Cloudflare urges Google to split bots for AI and search indexing, enabling sites to block AI scraping without hurting SEO. Google shows little sign of agreeing

• OpenAI’s Open Model Nears Launch - OpenAI will soon debut its first open-weight language model since GPT-2, letting anyone run it outside Azure. Release expected as early as next week.

Paper of The Day

Microsoft analyzed 200K Bing Copilot conversations to study AI's work impact. AI most used for information gathering, writing, communication tasks with highest success rates. Knowledge workers (interpreters, customer service, sales, writers) show greatest potential for AI integration while physical labor jobs remain largely unaffected. Study found AI often acts as supportive coach rather than replacement, augmenting vs automating jobs.

For the full research paper, visit here.