0911 | Local AI, Label-Backed Music, and a Foldable iPhone: This Week's Launches

Mia: Welcome back to the show, everyone. I'm Mia.

Milo: And I'm Milo. Today's episode is basically a tour of how intelligence is showing up everywhere right now — in the music you stream, in the apps running quietly on your laptop, in your code editor, even in the car next to you on the highway.

Mia: Yeah, and we're not just listing launches. For each one we're going to ask the same questions: who is this actually for, what does it change compared to what people already use, what's the strongest evidence it works, and what's still unknown.

Milo: And we'll be honest about that last part, because with a lot of these, the marketing is a claim, not a proven result.

Mia: Exactly. So let's start with what might be the biggest one, at least symbolically: AI music getting the stamp of approval from the major labels.

Milo: Okay, so Suno has released v6, and the headline here isn't just the model — it's who's involved. They built it in partnership with Warner, BMG, and Believe. Those are major players in music licensing, and that's genuinely a landmark deal for AI music, because until now this space has mostly been defined by lawsuits and ambiguity.

Mia: Right, and that licensing angle is why this matters commercially. If the major labels are at the table as partners rather than plaintiffs, AI-generated music suddenly has a legitimate path into the actual industry. That changes how artists, labels, and platforms can treat it.

Milo: Now, on the product side, v6 ships in three versions: there's the standard v6, then v6-wild — which as the name suggests is presumably the more experimental, riskier one — and v6-mini, which sounds like a lighter, faster option. So they're covering different needs rather than one model for everybody.

Mia: And there's a real editing capability now, not just "generate and hope." You can edit by section and by lyric. That's the difference between a toy and something you can actually finish a song with — you can go in and rework a chorus or rewrite a specific line instead of rerolling the whole track.

Milo: Who's this for, though? I'd say two audiences. Hobbyists who just want to make music, obviously. But also content creators — and that connects to some other things we're seeing. There's this wave of creator tools right now where the browser is becoming the medium. There's a product called Whip, for example, which is a YouTube-style feed of interactive experiences — games, 3D, generative art — all playable right in the browser, with channels for creators.

Mia: That's a good parallel, because music and interactive content are both categories where the tools are getting dramatically easier, and the distribution is shifting to things you experience directly rather than just watch or listen to.

Milo: And on the video side of creation, there's Mock Magic — it takes screen recordings and turns them into polished videos with a device frame and your branding, all in the browser. Images are free, video is the Pro tier.

Mia: So the pattern across Suno, Whip, and Mock Magic is the same: take something that used to require a studio, a dev team, or a motion designer, and make it a browser tab. Suno is just the one where the rights question was sharpest — and they've answered it, at least partially.

Milo: What's still unknown there? The exact licensing terms aren't public as far as we know. What the labels got, what Suno can and can't generate, whether artists can opt in or out — all of that matters enormously for what "legitimate" actually means.

Mia: And the open question going forward is how these tiers actually perform with real users — whether v6-wild produces anything genuinely interesting or just noisier output, and whether the section and lyric editing is granular enough for serious work.

Milo: One more honest caveat: everything we just described about quality is the maker's claim. Nobody's independently benchmarked these tiers yet. So treat "legitimized by Warner and BMG" as fact, and treat "the music is great" as a claim to be tested.

Mia: Now, from that — Suno is very much a cloud service. All the heavy lifting happens on their servers. Let's go in the completely opposite direction: AI that runs entirely on your device.

Milo: This is probably the clearest theme of the moment. There's a company called Desert Ant Labs building small AI models that run on-device — voice, text, and vision — and they distribute them through an SDK. So if you're a developer, you can embed local AI into your app without training anything yourself.

Mia: And the pricing model is interesting: it's free up to 100,000 monthly active devices. That's a very low barrier. It means a solo developer or small team can ship a genuinely local AI app and not pay a cent until they have real scale. That's the kind of thing that lowers the barrier for local AI apps dramatically.

Milo: The natural question is: why bother with local at all, when cloud models are bigger and better? And the honest answer involves trade-offs we don't fully know yet — on-device models are smaller, so the quality versus cloud is an open question. But what you get in exchange is privacy, no latency, and no dependence on someone else's server bill.

Mia: And the best showcase of why privacy matters is Speechmark. It's meeting notes for Mac, and it's 100% on-device. There's no bot joining your call, no account required. It keeps the original audio, so you can go back and verify. And the price is a one-time $49 — it'll go up to $79 later, but it's a license, not a subscription.

Milo: That no-bot part is worth dwelling on, because it's become weirdly normalized to have an AI participant listening to every meeting. Speechmark's pitch is that your meeting notes don't require anyone else's server hearing the conversation at all. For anyone in legal, HR, medicine, or honestly any sensitive business conversation, that's a meaningful difference, not a feature checkbox.

Mia: Compare that to the Suno model and you see the two poles of the industry. One says: put everything in the cloud, partner with the biggest players, scale up. The other says: shrink the model until it fits in your pocket and nobody else needs to see your data.

Milo: And there are more examples on the local side. Wealthfolio is personal finance that's open-source and local-first — investments, net worth, expenses — with an add-on called Connect that syncs with brokers, and that sync is encrypted. So even when it does touch the cloud, the design principle is that your financial data stays yours.

Mia: On the Mac utility side, there's Gojo — local on-device dictation, plus clipboard management, window management, file handling, media tools, all living in the notch. Three-day trial, then a lifetime license at $9.99. So local dictation has gotten good enough to be a $10 impulse purchase, which says something about how mature small models have gotten.

Milo: And on the more ambitious end, Viso Now — computer vision platforms generated from a prompt, with live dashboards, no model training or annotation needed. The direction is the same even if it's not necessarily fully local: the tooling for building intelligent apps is being collapsed down to natural language.

Mia: So what's next for this theme? More apps going local, almost certainly — the SDK economics from Desert Ant make that inevitable. What's unknown is whether on-device quality closes the gap with cloud for anything beyond these focused use cases like dictation and meeting notes.

Milo: Which sets up our next topic perfectly, because these local models and these agent tools are increasingly intertwined. Agents — AI systems that do work, not just chat — are getting real infrastructure.

Mia: Let's start with developers, since that's where agents are furthest along. There's a tool called hob: a local, professional workspace for code agents — it supports Claude Code, Codex, and OpenCode. It gives you worktrees, pull requests, and encrypted remote access.

Milo: Why does that matter? Because running an AI coding agent in a terminal window is fine for a demo, but professionals need the same things human engineers need: isolated workspaces so parallel work doesn't collide, a review process, and secure access when you're not at your desk. hob is saying agents deserve professional tooling, not just a chat prompt.

Mia: Right — the pitch is "agents are getting infrastructure, not just chat," and I think that framing is exactly right. It's the same story with Modeinspect, which is an AI design canvas that sits on top of your real codebase. It uses your actual components and design tokens, and the changes it produces can be published as type-safe pull requests.

Milo: That's a notable difference from the usual "AI generates a mockup" tools, where the design drifts away from the code and someone has to reimplement it by hand. Modeinspect keeps design and code in the same world, and type-safe PRs mean the changes should at least compile against your real types.

Mia: Both of those assume your project is in decent shape to begin with — and there's a tool for that too. Thousand is documentation as git repositories in Markdown, with access organized by folder, and permissions per reader. Humans and agents read the same files.

Milo: That last part is the quietly important bit. If your docs live in a proprietary wiki, your coding agent can't see them. If they're Markdown in a git repo, both your teammate and your agent read the same source of truth. Same files, same permissions, both audiences.

Mia: Now, agents aren't just for developers. Typewise Nova is an AI operator that creates and improves customer support agents without needing developers. You test it against your past tickets, and you get a thousand resolutions free.

Milo: That "test on past tickets" approach is the strongest part of the pitch, honestly. Instead of trusting a demo, you can check how the agent would have handled conversations you already know the outcome of. A thousand free resolutions gives you room to do that seriously.

Mia: But let's apply our standard skepticism here. For all three of these — hob, Modeinspect, Typewise Nova — the big unknown is the same: how reliable are the generated changes in production? A type-safe PR is type-safe, but that doesn't mean it's correct. A support agent that passed on past tickets still has to handle the weird new ticket.

Milo: And the thing to watch next is consolidation. Right now you've got separate tools for agent workspaces, agent design, agent docs, and agent support. Those workflows are going to merge or get absorbed — that's the natural next step.

Mia: And observability is part of that consolidation story — there's OpenObserve's AI Observability, which traces OpenTelemetry spans from agents and LLMs, gives you cost per session, detects loops, and does inline evals right alongside your logs and infrastructure. That's the "treat your AI like production software" mindset arriving in monitoring tools.

Milo: Which is basically the prerequisite for consolidation — you can't merge agent workflows until you can see what all your agents are actually doing.

Mia: Okay, so we've done cloud AI, local AI, agent tooling. Let's shift to hardware, and specifically Apple's next bets.

Milo: The big one is the iPhone Duo — Apple's first foldable. The numbers: the screen is 50% larger than the iPhone 18 Pro Max when unfolded, and the exterior display covers roughly 90% of the 18 Pro's screen area.

Mia: That exterior display number is the interesting one. A lot of foldables treat the outer screen as an afterthought — a tiny strip. If the cover display is nearly a full-sized phone screen on its own, that suggests Apple is positioning this as two phones in one, not a phone with a tablet attachment.

Milo: The question of who it's for is still genuinely open, though. Foldables so far have appealed to a niche — people who want the bigger canvas. Whether Apple's take broadens that or just serves the same niche more expensively, we don't know.

Mia: Alongside it, AirPods 5. Active noise cancellation is 1.5x stronger, in an open-ear design — which is a real technical challenge, because open-ear buds don't seal, and ANC without a seal is hard. They also get conversational Siri and live translation. They ship September 18, 2026.

Milo: The open-ear plus strong ANC combination is the bet worth watching. If Apple pulls that off, it changes the calculus for people who hate the sealed feeling of in-ear buds but still want noise cancellation. And live translation on your ears is the kind of feature that sounds incremental until you're standing in a conversation in another language.

Mia: And notice the ecosystem angle — Gojo, the Mac utility we mentioned, shows why hardware like this keeps people in Apple's world. The dictation and the notch utilities only work because Apple gives developers deep local access to the OS. Hardware and software reinforce each other.

Milo: What's unknown on the Duo: battery and weight. Folding always costs you something — you're either carrying a thicker phone or a smaller battery, usually both. Apple hasn't told us how they handled that. And going forward, we're waiting on pricing and real-world durability, which for a first foldable is everything.

Mia: So hardware is one half of the story. The other half is software and services — and specifically, tools solving very concrete, practical problems. Let's close the loop on that.

Milo: Two launches here that I think are genuinely useful rather than just impressive. First, FreeScan: a free audit of any URL covering SEO, AEO — that's answer engine optimization, being visible in AI answers — plus security and accessibility. And crucially, it gives you evidence and fixes, not just a score. The Pro tier adds sitewide audits and MCP.

Mia: The evidence-and-fixes part is what separates it from the pile of "run a scan, get a number" tools. If you run a small business site, you don't need a grade, you need to know what to change on Tuesday morning. Free audit for a single URL, pay for the whole site — that's a sensible ladder.

Milo: Second is Drive, an iPhone app that turns your phone into a telemetry logger — it captures g-force and braking — and it warns you about LPR cameras using OpenStreetMap data.

Mia: So the same device in your pocket that AirPods complement and Speechmark's cousin apps run on becomes a driving analytics tool. It's a nice example of mobile tech applied to a concrete problem: understanding how you actually drive, and knowing where the cameras are.

Milo: Who's it for? Enthusiast drivers who want data, and commuters who want camera warnings. What's unknown — and this is the meaningful limitation — is the accuracy of that camera data. It's crowdsourced through OpenStreetMap, which means it's only as complete and current as the contributors have made it. If a camera is new or unlogged, Drive doesn't know about it.

Mia: That's a fair caveat for any map-based safety claim: treat the warnings as helpful, not guaranteed. Same skepticism applies to FreeScan's fixes — the audit gives you evidence, but you're still the one verifying the fix works.

Milo: And a couple more practical tools round out the picture. Athenic AI does stock analysis with over 90 licensed institutional datasets — datasets that normally come with a $50,000 license — plus AI for screening and research in natural language. The value there is access: research that was paywalled for institutions, now queried in plain English.

Mia: And Live Captions by Subanana handles the live-event problem: multilingual captions delivered via a QR code straight to the audience's phones, up to five languages, with a layer that works with OBS and vMix for streamers.

Milo: So across FreeScan, Drive, Athenic, Live Captions, and OpenObserve's AI observability, the thread is measurement and access. Measure your site, your driving, your portfolio, your event audience, your AI systems.

Mia: "Measure everything" is a good place to land, actually. Because that's the throughline of this whole episode — whether it's Suno licensing with the majors or a foldable iPhone, the tools getting interesting are the ones that give people information they didn't have before and act on it.

Milo: And the honest summary is: a lot of claims, a lot of genuine progress, and a lot still unproven. The licensing terms are unknown, the on-device quality gap is unknown, the production reliability of agent-generated work is unknown, the foldable's battery is unknown, and the camera data accuracy is unknown.

Mia: Which is exactly why we do it this way — who it's for, what changes, what's the evidence, what's still open. That's the episode. Thanks for listening, everyone.

Milo: We'll see you next time.