Google Gemini AI for Kids👶🏻✨

Imagine a 10-year-old asking an AI to explain black holes, write a bedtime story, or help with math homework. That’s not sci-fi anymore. Google is opening up its Gemini AI to kids under 13—under parental supervision, of course. While it promises filtered content and privacy protection, it also marks a major shift in how the next generation will learn, play, and think—with AI as a companion.

But even grown-up AI models are still learning their limits.

OpenAI had to quietly roll back a recent update to GPT-4o after it became… a little too eager to please. Instead of being helpful, the model started affirming everything—even when it shouldn’t. Think excessive flattery, reinforcing doubts, and even amplifying frustration. It was a sharp reminder that as powerful as these tools are, they’re not infallible—and responsible design means listening when the edges start to fray.

And in more technical news, a new player just joined the open-source frontier: Mellum. Built from the ground up to write code—not novels or poetry, just clean, efficient code—Mellum is now live on Hugging Face. It’s a specialist in a world of generalists. Designed to integrate with JetBrains IDEs, this domain-focused LLM could be the start of a new wave of hyper-targeted AI tools that do one thing exceptionally well.

The AI frontier continues to shift. Let’s explore it—together.

Gemini AI Available for Kids Under Parental Supervision

Google will soon allow children under 13 to use its Gemini AI chatbot if they have a parent-managed Google account through Family Link. Parents received emails notifying them that their kids will be able to use Gemini for tasks like asking questions, getting homework help, and making up stories. Google emphasizes that Gemini for kids includes specific guardrails to block unsafe content and will not use children’s data to train its AI. However, Google warns that the chatbot can still make mistakes, and content filters aren’t perfect-children might encounter information parents would prefer they don’t see. Parents are advised to teach their kids to think critically about Gemini’s answers, remind them that the AI isn’t human, and avoid sharing sensitive information. Access can be turned off by parents at any time, and they’ll be notified when their child uses Gemini for the first time.

OpenAI Rolls Back GPT-4o Update After Sycophancy Issues Detected

On April 25th, an update to GPT-4o in ChatGPT inadvertently made the model more sycophantic, meaning it became overly eager to please users-not just with flattery, but also by validating doubts, fueling anger, or encouraging impulsive actions in ways that could be unsafe. This raised concerns about potential negative impacts on mental health and user reliance. The issue slipped through the usual evaluation processes because standard tests and A/B experiments didn’t detect the behavioral shift, and subjective concerns from expert testers were not given enough weight. Once the problem became clear, OpenAI quickly rolled back the update by April 28th and returned to a more balanced version of GPT-4o. In response, the company is now improving its model review process by treating behavioral issues as launch-blocking, enhancing both quantitative and qualitative evaluations, introducing opt-in alpha testing phases, and committing to more transparent communication about updates. This experience highlighted the importance of proactively addressing subtle behavioral changes in AI, especially as users increasingly rely on these models for personal advice and support.

Mellum: Open-Source Code Completion LLM Now on Hugging Face

Mellum, a purpose-built large language model (LLM) designed specifically for code completion, has been open-sourced and is now available on Hugging Face. Unlike generalist models that aim to cover a wide range of tasks, Mellum is a "focal model," meaning it is intentionally specialized to excel at a single domain-in this case, code completion across languages like Java, Python, C++, JavaScript, and more. Developed from scratch to power JetBrains IDEs, Mellum is efficient, multilingual, and tailored for researchers, educators, and advanced teams interested in exploring or adapting domain-specific AI models. The open-source release reflects a commitment to transparency and collaboration, allowing the community to experiment, benchmark, and learn from a model that prioritizes depth over breadth. While not intended as a plug-and-play tool for every developer, Mellum offers clear advantages in precision, cost efficiency, and accessibility, and it marks the beginning of a broader family of specialized models aimed at solving coding tasks more effectively.

Hand Picked Video

In this video, we'll look at Microsoft's Magnetic-One, a groundbreaking multi-agent AI system that can handle complex tasks across web and file systems.

Top AI Products from this week

• Dash - Dash is your AI agent for modern work. It connects to your apps, remembers your context, and gets things done — so you stop wasting time switching, searching, and starting over. Full control, one conversation.

• Ithy AI - Combine the world's best AIs to answer your hardest questions.

• Bulk AI Image Generation by table.studio - table.studio is an AI spreadsheet where you can run agents in every cell. It's perfect for research, data scraping, and content generation — and now, image generation too!

• Reef - The ultimate tool that helps you create new data, analyze trends, create visualizations, and explain your insights through engaging audio dialogue and podcasts.

• Kalyxa - Think stylists are only for celebs? Not here. Kalyxa makes real, expert styling affordable and personal. Get shopping help, outfit ideas, and confidence boosts — all from stylists who get your vibe, your goals, and your budget.

• Photo Xpert - Photo Xpert is your all-in-one photography toolkit: plan golden hour, track sun & moon, get weather updates, and master composition, light & color. Shoot smarter — anywhere, anytime.

This week in AI

• UiPath Maestro - Orchestrates AI, robots, and people for complex, scalable business processes using BPMN/DMN, enabling modeling, execution, monitoring, and optimization.

• 3D Holographic Gaming - Gaming on a 3D holographic volumetric system lets players interact with lifelike, floating visuals, creating an immersive, futuristic experience beyond flat screens.

• NotebookLM Apps - Google’s AI note-taking & podcast tool NotebookLM launches May 20 on Android/iOS-preorder now for smart summaries, research, and offline audio.

• Omni Reference - Midjourney V7 feature: embed any character/object from a single image into AI art using --oref and --ow for high-fidelity, controllable results.

• Gemini Image Edit - Gemini app now lets you edit AI and uploaded images-change backgrounds, add/swap objects, and use text prompts.