Hello AI Enthusiast,

This week, Voice AI is getting better, AI music is getting longer, and a well-funded coding startup is learning that open source is fine but hiding it is not. A busy week, mostly from Google, with a side of drama from Silicon Valley. Let’s dive in.

The Big Picture 🔊

Google launches Gemini 3.1 Flash Live for more natural voice AI

Google released Gemini 3.1 Flash Live, its most advanced voice model yet. It handles complex tasks more reliably, follows longer conversations, and detects emotional cues like frustration or confusion. It now powers Gemini Live and Search Live, which expanded to 200+ countries. All audio is watermarked with SynthID to flag AI-generated content.

Helin Yontar Helin Yontar CPO and Polyglot

Voice AI is improving, but most of us still find it faster to read a response than listen to one. When it works naturally, this will be a big shift, since speaking is how we actually communicate.

We're not there yet, but Google rolling this out to 200+ countries is a serious push in that direction.

Google's Lyria 3 Pro generates full songs with structure

Google launched Lyria 3 Pro, an upgrade to its music generation model that creates tracks up to 3 minutes long with control over song structure, including verses, choruses, and bridges. It's now available across Gemini, Google Vids, Vertex AI, and Google AI Studio. All outputs are watermarked with SynthID, and the model avoids mimicking specific artists.

Biljana Prlichkova Biljana Prlichkova PRM and Outdoor Enthusiast

Going from 30 seconds to 3 minutes means you can now generate an actual song. Content creators who need background music without dealing with licenses and royalties will find this immediately useful.

The economics already make sense, and the quality is only going to improve from here.

Cursor's new Coding model was built on a Chinese open-source base

Cursor launched Composer 2, claiming it offered top-tier coding performance, but didn't mention it was built on Kimi 2.5, an open-source model from Chinese company Moonshot AI. An X user spotted the connection in the code. Cursor confirmed it, saying 75% of training was their own work. The co-founder admitted they should have disclosed the base model from the start.

Andrea Mattiello Andrea Mattiello CM and Board Lover

Using an open-source model as a base is technically fine, and Kimi even congratulated Cursor publicly.

But when you're valued at $29 billion and market a "new model" without mentioning it's built on someone else's foundation, that's a problem, especially for investors who probably funded you to build something original.

Every company is adopting AI at a different pace and for different reasons. Our corporate training is built around your specific context, so your team learns what is actually relevant to their work, not just what is trending.

Bits and Bobs 🗞️

  • Claude's new "computer use" feature on macOS allows the AI to control apps and interact with the screen to compile and test code directly on the user's machine.

  • Gemini now lets users easily import memories, preferences, and chat history from other AI apps for a personalized and seamless assistant experience.

  • Wikipedia has banned AI-generated content creation but allows AI for translations and minor copy edits with human review to prevent misinformation.

  • Microsoft 365 Copilot's new Copilot Cowork feature uses AI to plan, delegate, and track multi-step workflows across tools, enhancing productivity and teamwork.

  • A federal judge blocked the Trump administration from blacklisting AI company Anthropic as a national security threat, citing illegal retaliation for the company's government criticism.

That's a wrap on our newsletter! Before you go, here’s a quick recap of our offerings:

Catch you next week! 👋

Keep Reading