Google puts Lyria 3.5 music generation in the Gemini app
Google moved its best music model out of the studio tools and into the app hundreds of millions of people already open — while a Zhejiang University lab shipped an agent that runs interpretability research on its own, and IBM quietly made qubits reset 8,000 times faster.
Google has brought Lyria 3.5, its newest music generation model, straight into the Gemini app, and the same model is now callable through the Gemini API. In the app you pick or describe a genre, choose vocal or instrumental, set a short or long length, and start from templates aimed at backing tracks, jingles, and ringtones; Google positions it for exactly the jobs that used to need a stock-music subscription. The model first appeared in Google Flow Music on July 29, and it now also reaches developers through Google AI Studio and Google Vids, which is the part that matters: this is the first time Google's music model has been both a consumer button and an API endpoint on the same day. Developers get 44.1 kHz stereo output, up to 10 images alongside a text prompt, custom lyrics tagged by section, timestamps that say when each instrument enters, and lyrics generated in whatever language you prompt in — with SynthID watermarking on everything and hard blocks on artist-voice imitations and copyrighted lyrics.
The interesting tell is the ceiling Google put on it: generation is single-turn, so you cannot iterate on a clip across prompts, and the same prompt can return different results each call. That is a sounding tool, not a sequencer — and it lands weeks after a German court rejected Suno's fair-use defense, which makes Google's insistence that Lyria is trained only on licensed material look less like caution and more like the moat.
IBM's Nighthawk r2 is live on IBM Quantum Platform, and it hits 100,000 circuits per second — roughly 25 times the throughput of the Heron fleet. The gain comes from a dissipative reset gadget: instead of waiting hundreds of microseconds for qubits to relax between runs, each qubit is coupled to a cold environment that drags its effective T1 from about 200 microseconds down to roughly 25 nanoseconds, cutting idle time between circuits to as little as one microsecond and reducing initialization error by about 25x. The chip dedicates 120 reset elements and 218 couplers to 120 programmable qubits — 458 physical elements for 120 usable ones — and IBM says it has already run accurate observable estimation on circuits past 7,500 gates, one of its stated 2026 roadmap targets. Early workloads show up to a 10x speedup on advantage-candidate circuits and a 12x speedup on neutron-scattering simulations that now produce lab-comparable spectra in about a minute; the honest caveat, which IBM concedes, is that short circuits barely benefit and this is still not a fault-tolerant machine.
A team at Zhejiang University's ZJUNLP lab, with collaborators including Julian McAuley and Tat-Seng Chua, built Mechanist — an agent that does interpretability research end to end. Give it a question about how a model works internally and it searches a purpose-built knowledge graph (26 disciplines, plus a curated interpretability graph), turns the question into a falsifiable hypothesis, picks from a library of mechanism methods — probing, circuit discovery, causal attribution, feature dictionaries, SHAP and others — runs the experiment, then re-tests the finding with a different method, dataset, and model before logging a claim. Its published cases are pointed: unsafe behavior reappearing after fine-tuning on data that looks clean, separable attention heads that decide whether a model reports a fact or attributes a belief to someone else, and steering a scientific foundation model toward DNA sequences with high α-helical content. It ships open source and installs as a coding-agent plugin, so the experiments run on your own GPUs — which is the right architecture for a field where "the model investigated itself and found nothing wrong" is a result nobody should have to take on trust.
What to watch: whether Google opens iterative editing for Lyria, since single-turn generation is the single biggest thing standing between a demo and a workflow.
Would you trust an AI agent's report on what another AI is really doing inside — or only the parts you can re-run yourself? Tell us in the comments.
Sources: Mechanist (GitHub)
