AI in everyday software | AI Field Notes #24

AnalysisGoogle's music model Lyria 3 Pro just got a distribution channel. Believe, the indie music group whose TuneCore arm distributes hundreds of thousands of independent releases, will offer its artists Flow Music, the Google Labs tool wrapping Lyria 3 (formerly ProducerAI). The model takes prompts for intros, verses, choruses, and bridges, and emits tracks up to three minutes. Google says it does not claim ownership of generated output and that Lyria 3 was trained on 'materials YouTube and Google has a right to use,' the same defense Suno and Udio (rival generative-music startups facing label copyright suits) are running. Selected artists become weekly 'ambassadors' meeting the product team, free user research from working musicians.

AnalysisAnthropic pushed Claude into Microsoft 365 on May 11, integrating the model across Excel, Word, PowerPoint, and Outlook. The distinguishing design choice is cross-app context persistence: Claude can hold what it learned in a Word document while working in an adjacent Excel sheet, without the user re-explaining the context. Most enterprise AI integrations treat each app as a separate conversation. The cross-app approach reflects how knowledge workers actually move between tools, and for developers building M365-connected workflows, it changes what Claude can see and respond to within a standard enterprise session without additional context wiring.

AnalysisThinking Machines Lab, a new research group, released a preview on May 12 of what it calls interaction models: AI systems that process audio, video, and text in parallel streams rather than sequentially. The design lets a model respond mid-sentence and adjust based on real-time visual or audio feedback, eliminating the turn-based rhythm that makes current voice assistants feel slow. The lab trained the system from scratch on the multi-stream architecture rather than adapting an existing text model. No production API, pricing, or timeline was announced; this is a research preview.

AnalysisCowboy Space Corporation announced on May 13 it raised $275 million to build rockets designed to deploy AI data centers in orbit, with each satellite hosting approximately 800 GPUs and first deployment targeted by end of 2028. Orbital positioning sidesteps two of the ground-based buildout's biggest friction points: regulatory land-use approvals and community opposition over water and power. The trade is power density: satellites run on solar panels with constrained wattage, and 800 GPUs per satellite is a small fraction of what a modern ground-based facility runs. At $275 million the company is a bet on the permitting bottleneck outlasting current timelines.

AnalysisFormal community opposition to AI data center construction is emerging as a concrete bottleneck in Utah, Virginia, Arizona, Georgia, and Texas, where residents are raising objections over water consumption, power demands, and grid load. The backlash is slowing permit approvals for projects that were otherwise funded and ready to break ground. The Big Four hyperscalers (Microsoft, Amazon, Alphabet, and Meta) collectively expect to spend $725 billion on AI infrastructure in 2026, but permit delays could compress that timeline. Kevin O'Leary's $1 billion Stratos data center project drew attention after local opposition became a direct obstacle to construction.