Harvey, the legal-AI outfit valued around $11 billion, spent years building on top of other people's models — OpenAI, Anthropic, whoever had the best contract-drafting chops that week. On Tuesday it introduced Tenet, its first in-house, proprietary model built specifically for legal work. That's a bigger deal than a single product launch; it's a company quietly telling its suppliers, "thanks for the ride, we'll take it from here."
The supplier-competitor problem nobody wants to talk about
The awkward part of Harvey's business was always the same one every vertical AI startup faces: you build your moat on rented brains. Every time a lawyer fires up a model through Harvey, someone upstream gets paid per call. Usage grows, the tab grows, margins shrink. Anthropic has been angling at lawyers with document-review plugins, OpenAI just hired the Ironclad founder to run its legal push, and Google and Meta are circling. When your model supplier decides it wants your customers too, you're in a weird spot — you're paying the person who's about to eat your lunch.
So Harvey built Tenet to route more of the work through its own engine, cut those per-call fees, and stop being hostage to roadmap decisions it doesn't control. From my perspective, this is the play every successful vertical AI company has to run eventually. You can't stay a thin wrapper over the giants forever and keep pretending you own the relationship with your users. The real question is whether a specialist model can actually match the frontier generalists on messy, real-world legal tasks — contracts are full of jurisdiction quirks and precedents that a general-purpose model trained on the whole internet handles about as gracefully as you'd expect.
To be fair, Harvey isn't ditching third-party models entirely; it's adding a cheaper, targeted option into the routing. That's the pragmatic version of vertical integration. It just means the era of "your supplier is also your rival" is officially open for business, and legal is only the first industry where it gets ugly.
Japan wants to know what went into your training set
Across the Pacific, Japan is moving to require AI firms to disclose what data their models were trained on. The proposed rules are aimed at transparency — knowing whether a model was trained on copyrighted material, where it came from, how it was licensed. This is the conversation that's been bubbling for two years now, and it's finally starting to harden into actual regulation rather than angry blog posts.
I'm genuinely torn on this one. On one hand, the opacity of training data is absurd for an industry asking the world to trust it with everything from medical advice to legal filings. You literally cannot verify claims about bias, provenance, or contamination if nobody will say what went into the pot. On the other hand, "disclose your training data" runs headfirst into real trade secrets — the exact composition of a training corpus is where a lot of model quality actually lives. Nobody wants to hand their competitor the recipe for free.
The interesting bit is that Japan is picking a middle path: not full public disclosure, but disclosure to regulators, with carve-outs for genuinely proprietary details. That's the kind of compromise that could actually survive contact with the industry. Keep in mind, this is early days — the rules aren't final, and enforcement is a whole separate fight. But it's the first major jurisdiction to make transparency a formal requirement rather than a PR talking point.
Meta finally lands on the Mac — about time
Meta's AI assistant has been living on phones and in WhatsApp for a while, but it took until now to get a proper Mac app. It's in beta, aimed squarely at businesses and creators, with screen sharing during sessions and system-wide dictation. It hooks into Facebook and Instagram for performance analytics and connects to Google Workspace — though you'll need a professional account for that part.
What actually grabbed me about this one: the app can look at a window you share and give advice on what you're working on. That's genuinely useful for people staring at dashboards. But Meta's also pushing the assistant to "complete recurring tasks and reminders and generate decks, docs, spreadsheets," which is the same vague agent promise every company is making this year. The dictation being system-wide is nice. The rest, honestly, is catching up to what the other assistants have been doing on desktop for a while.
And a quick add-on note: Meta's privacy policy makes it clear that AI interactions can feed back into how the systems learn. For business users, that's a real decision to make, not a checkbox to ignore. Creators running client work through a shared window should think hard about what's in that window before they hand Meta a live view of it.
The hardware company that's done waiting
PINE64 — the open-source hardware people behind the PinePhone and PineBook — has decided to stop producing Linux hardware until what it calls the AI bubble bursts. That's a striking thing to read from a company that's spent years shipping devices nobody else would build. Their read: component pricing and supply have been warped by the AI buildout, and the economics of small-batch, enthusiast hardware no longer work while everyone's hoarding silicon and jacking up prices.
I have mixed feelings. On one level it's a legitimate grievance — the AI boom has genuinely distorted the market for memory, storage, and compute, and small hardware makers are getting squeezed by buyers with effectively unlimited budgets. On another level, "we're stopping until the bubble bursts" reads like a company that's exhausted, not one that has a plan. Timing the collapse of an AI bubble is a hobby, not a business strategy. But it's a useful signal about how badly the boom's side effects are hammering everyone who isn't selling shovels.
Gemini goes to school
Google is pushing Gemini into education in a serious way this month — expanding student access globally and reportedly offering a free year of Gemini Pro plus study tools, a direct jab at ChatGPT's education play. The governance questions are real: schools adopting AI assistants need to think about privacy, what happens to student data, and whether an assistant optimized for homework completion is actually teaching anything.
That last part is where I get skeptical. I've watched students use AI to skip the thinking step entirely, and a free year of Pro is going to supercharge that. The tools are genuinely good for summarizing dense material and checking work — I use them myself for that. But the difference between "AI helps you understand" and "AI answers for you" is entirely about the habits you build, and a marketing push aimed at students is going to pull hard in one direction. Schools that adopt this without a policy for how to use it are basically deciding to wing it.
The takeaway
Vertical models are going to eat the middle of the AI stack, regulators are finally moving from complaining to rulemaking, and the consumer assistant wars are settling into "me too" territory on desktop. None of these are clean wins — every one of them has a real trade-off buried in it. That's just where this industry lives now: every win comes with a string attached, and the job of paying attention is figuring out which string you're willing to pull.
By the way, I've been putting together rough cost comparisons on different AI setups and plans for a while — threw them into a small Decision Calculator page if you're in the market and don't feel like doing the math by hand.
Top comments (0)