Subscribe, if this was forwarded to you

Hey there,

This week’s edition is packed with resources, product updates, and a few standout releases. We’re featuring a curated collection of generative AI learning materials, a brilliant offline-first meeting notepad, and Google’s latest image editing model—yes, “Nano Banana” is real (and surprisingly good). Plus, Microsoft is doubling down on voice and agent tech, while new tools like EmbeddingGemma and HumanLayer continue to reshape how we build with LLMs.

Let’s dive in!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Master Generative AI with This Curated GitHub Guide (Courses, Notebooks & More)

🛠️ Tool Spotlight: Hyprnote – An Offline AI Notepad That Understands Your Meetings

📰 Top News: Google’s “Nano Banana” Update Makes Image Editing Smarter and More Consistent

💡 Tip: HumanLayer – Add Human-in-the-Loop Checks to Your AI Agents

Let's get started!

Find this Newsletter helpful?
Please forward it to your colleagues and friends - it helps us tremendously.

Tutorial of the week

Your Ultimate Generative AI Toolbox in One Repo

This week, we're spotlighting a powerful resource—the Awesome Generative AI Guide by Aishwarya Naresh Reganti. This GitHub repository is a goldmine for anyone diving into generative AI, offering everything from cutting-edge research summaries to ready-to-use notebooks and interview prep materials.

What You’ll Find in the Repo:

• Curated Research Updates: Monthly paper summaries keep you in sync with trends like Multimodal LLMs and RAG.

• Interview Prep: A set of frequently asked questions to help you nail AI and LLM-related interviews.

• Learning Paths: Structured courses like "Applied LLMs Mastery 2024" and "Generative AI Genius 2024"—complete with lesson plans and roadmaps.

• Free Courses & Interactive Notebooks: Compile your own learning path using 70+ open course links and notebooks for AI experiments.

This guide isn't just for learners—it’s continuously updated, serves both professionals and newcomers, and is open-source under MIT license. Consider it your go-to launchpad for structured learning, hands-on tutorials, and staying current with the AI frontier.

Tool of the week

Hyprnote — Your Privacy-First AI Notepad for Meetings

When meetings pile up, keeping track of key points without losing privacy can be a challenge. Meet Hyprnote, an open-source, offline-first AI notetaker that transcribes, structures, and summarizes your meetings right on your device—no data leaves your laptop or server.

Why Hyprnote Stands Out:

• Local-first & Secure: All speech-to-text transcription and note summarization happen offline—perfect for sensitive environments or air-gapped teams.

• Real-time Assisted Workflow: As you take notes, Hyprnote listens and enhances them with summaries, action items, and follow-up prompts, giving your meeting transcripts context and clarity.

• Flexible Model Support: Swap in your preferred LLM—whether local via Ollama, or APIs like Gemini, Claude, or GPT—Hyprnote adapts to whatever fits your privacy policy.

• Templates & AI Chat Built-in: Use structured templates for summaries, ask questions like "What were the action items?" directly in the notes, or translate on the fly.

With 6,000+ stars on GitHub and backing from privacy-conscious teams, Hyprnote is becoming a go-to tool for professionals who want AI-powered productivity without compromising confidentiality.

Top News of the week

Google's “Nano Banana” Makes AI Image Editing Smarter and More Consistent

Google has released Gemini 2.5 Flash Image, also known by its internal nickname “Nano Banana”, and it’s turning heads for all the right reasons. This new model delivers impressive improvements in image editing, particularly around character consistency—ensuring faces, pets, and objects remain stable across multiple prompts and edits. It also supports multi-step instructions, meaning you can layer edits (e.g., change background, add sunglasses) without the model losing track.

Google also shared tips on X for getting the best results from Nano Banana, including prompt strategies for more controlled editing and character coherence. The model is now available to all Gemini Flash users inside the Gemini app.

🔗 Read the official announcement

Microsoft Unveils In-House AI Models, Marking Shift from OpenAI Dependency

Microsoft has introduced two new models under its MAI (Microsoft AI) initiative: MAI‑Voice‑1, a fast, expressive speech model already powering Copilot Daily and Labs; and MAI‑1‑Preview, a Mixture-of-Experts text model trained on 15,000 H100 GPUs, now live on LM Arena. With these models, Microsoft takes a clear step toward AI independence, signaling reduced reliance on OpenAI for core infrastructure.

🔗 Read the full announcement

Google Releases EmbeddingGemma – Compact Embedding Model for On-Device RAG

Google has introduced EmbeddingGemma, a 308M parameter lightweight embedding model optimized for on-device retrieval-augmented generation (RAG) and semantic search. The model uses matryoshka representation learning, which allows dynamic control over embedding size depending on use case. EmbeddingGemma is released under the Apache 2.0 license and is available on Hugging Face and Kaggle.

🔗 Read the full announcement

Bring AI into Excel — Literally

Microsoft has quietly added the COPILOT function to Excel (Windows & Mac with Insider/Beta access). Now, you can use formulas like =COPILOT("Summarize this feedback", A2:A20) right in your spreadsheet. It works seamlessly within Excel’s calculation engine, so results update live when cells change and can be embedded in IF, LAMBDA, or other logic—but Microsoft warns against using it for critical tasks where accuracy is essential.

🔗 Read the full blog post here

Microsoft Unveils VibeVoice—a Leap in Open-Source Long‑Form TTS

Microsoft has released VibeVoice-1.5B, an open-source TTS model capable of generating up to 90 minutes of expressive audio with up to four distinct speakers, ideal for podcasts, interviews, or multimedia narration. Its advanced architecture uses continuous acoustic and semantic tokenizers at just 7.5 Hz, enabling efficient handling of long sequences while preserving fidelity. Built atop Qwen2.5-1.5B, VibeVoice employs a diffusion-based decoder for high-quality speech and includes safety features like audible disclaimers and embedded watermarks. Released under an MIT license, this model is now available for experimentation via GitHub and Hugging Face.

🔗 Explore VibeVoice on the project page

Tip of the week

Human-in-the-Loop Control with HumanLayer

When your AI agent performs critical tasks—like modifying data, sending emails, or triggering customer actions—mistakes can be costly. HumanLayer offers a flexible human-in-the-loop SDK that ensures key actions are always supervised by a real person.

• Use @hl.require_approval() to pause tool or function calls until a designated user explicitly approves or rejects them.

• Or call hl.human_as_tool() when you need human insight mid-task—whether it’s debugging, guidance, or creative feedback, sent via Slack, email, or soon Discord.

• It works with your favorite frameworks—such as LangChain, OpenAI, CrewAI—and supports flexible approval workflows with escalation rules, timeouts, and audit logging.

If you're moving toward “autonomous agents,” HumanLayer provides the safety net that lets them act confidently—but not unconstrained.

Explore the features and getting-started guides here:

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help