Google's Lyria 3 Brings AI Music to Gemini

Google just dropped Lyria 3, their latest AI music generator, directly into Gemini. You can now tell your chatbot to make you a song from text, images, or video—which sounds cool until you realize we're talking about 30-second clips that can't build on themselves.

Here's the thing everyone's missing: this isn't actually about competing with Suno or making professional music. Google explicitly said "the goal of these tracks isn't to create a musical masterpiece, but rather to give you a fun, unique way to express yourself." They're not even pretending this is for musicians. It's for YouTube Shorts creators who need quick background tracks and people who want to send AI-generated birthday songs to their friends.

Which is fine! Actually kind of smart. But the discourse immediately went to "Suno sounds better" and "this isn't ready for real use," completely bypassing what Google's actually attempting here. They're embedding Lyria 3 into YouTube's DreamTrack tool and making it native to Gemini—classic Google move of spreading features across their ecosystem rather than building one killer standalone product.

The multimodal piece is genuinely interesting though. Being able to generate music from video input isn't just a parlor trick. As one observer noted: "Video to audio alignment is the real flex here. Generating lyrics and vocals that actually sync with visual cues in real time is a massive multimodal serving challenge." Google's betting that having music generation as one more capability in their platform matters more than having the best standalone music AI. We'll see if they're right.

When Your Terms of Service Become a Meme

Meanwhile, Anthropic managed to accidentally start a small riot this week by updating their terms of service in the vaguest possible way. The new language said users couldn't use OAuth tokens from Claude subscriptions "in any other product tool or service including the agent SDK."

Cue immediate panic from everyone using Claude to power their custom AI agents, especially the OpenClaw community. People paying $200/month for Claude Max were suddenly worried their entire workflow was about to get nuked. Developer Hubert Leiki summed up the mood: "Anthropic is in an active self-destruction mode now. First, they went after tokens you already paid for blocking use in non-claw code apps... Open code, Gemini CLI, Codex CLI are all legitimate coding agents with comparable features and abilities, but Anthropic are behaving like they're still the only player on the block."

Anthropic's response somehow made it worse. Their exec popped into the replies with "Apologies. This was a docs cleanup we rolled out that's caused some confusion. Nothing is changing about how you can use the agent SDK and Mac subscriptions." Cool, so... can we use OpenClaw or not? The clarification clarified nothing.

Turns out the actual policy is: personal tinkering is fine, but third-party businesses need to pay for API access. Which is reasonable! But also wasn't what the terms said, and also isn't how they communicated it, and also—plot twist—OpenAI and Google already ban this same use case anyway.

Colin Darling brought receipts: "Everyone upset about Anthropic's update to their terms would be wise to read the OpenAI and Google Gemini terms while they're at it. I'm bummed out, too, but Anthropic is late to this party, not leading it."

The controversy died fast, but the underlying tension didn't. How long will AI companies tolerate people building modular systems on top of their models? These platforms want to be ecosystems, not components. And that fundamental tension isn't going away just because one messy terms-of-service update got walked back.

Meta's Smartwatch Revival and the Wearables Land Grab

In less chaotic news: Meta's bringing back their smartwatch project. Code-named Malibu 2 (the original Malibu got shelved in 2022), this version reportedly includes health tracking and—surprise—a built-in Meta AI assistant.

What's interesting here isn't the watch itself. It's that Meta, Apple, and Google are all clearly positioning smartwatches as AI interface devices. Apple's reportedly working on AI-enabled smart glasses, a pendant, and camera-equipped AirPods. They tested a camera-equipped Apple Watch but scrapped it because sleeves kept blocking the lens, which is genuinely funny.

Meta's approach seems more focused. They've already got the Ray-Ban smart glasses doing well, and they're planning AR glasses for 2027. The smartwatch fits into that stack as another always-available AI endpoint. The original Malibu prototype had cameras for video conferencing and could read nerve signals in your wrist to act as a controller—wild features that point to where Meta thinks this is going.

None of these companies want to just make a better fitness tracker. They're trying to figure out what the primary AI interface looks like when it's not your phone. That's the real game.

The Benchmark Problem Nobody Wants to Talk About

Last thing worth noting: Lindy founder Flo Crivello threw some cold water on the Chinese AI models that keep posting impressive benchmark scores. His company's biggest cost is inference, so they test every model obsessively. His finding? "Every time we've evaluated them, we found the same thing that their real life performance for agentic behavior and outside of coding use cases falls extremely short of what they show on the Avals."

His theory: these labs are distilling frontier models (leading to shallow intelligence), training specifically for benchmarks, and potentially stealing weights. Harsh, but he's not alone in noticing the gap between benchmark performance and actual utility.

Here's the thing that makes this relevant beyond Chinese models: "That I think is a lesson that is relevant not just for Chinese labs, but also whenever you see a new Western model as well that has high benchmarks. Ultimately, you got to just dive in and test these things out for yourself."

Benchmarks are becoming less predictive of real-world performance across the board. Whether that's intentional gaming of the metrics or just the limitations of standardized testing for something as complex as intelligence, the result is the same: you can't trust the leaderboard anymore. You have to actually use the thing.

Which circles back to all of today's stories in a weird way. Google's betting on integration over raw capability. Anthropic's trying to balance openness with business models. Meta's exploring hardware form factors. And everyone's discovering that the specs don't tell you what actually works. We're past the phase where better benchmarks meant better products. Now it's about systems, interfaces, and use cases—the messy stuff that doesn't reduce to a number.

—Zara Chen, Tech & Politics Correspondent