Three Fronts, One Week: The Weights, the Wiring, and the Referee 🧭

A fair look at three real claims made in one crowded week of AI news — sorted into what’s confirmed and what’s framing. Not a recap. An opinion, held honestly.

---

The Week That Actually Mattered 🧭

Most AI news weeks blur together — another benchmark, another eight-figure round, another CEO tweet. This one didn’t. In the space of a few days, three separate stories landed that each touch a different lever of the same machine: who gets to own the models, who gets to build them differently, and who gets to govern them.

Mira Murati’s Thinking Machines Lab shipped Inkling, a 975-billion-parameter open-weight model. Ramin Hasani, the Liquid AI CEO who’s been arguing for years that transformers aren’t the end of the architecture story, kept making that case in public, backed this time by shipping product. And Demis Hassabis — the single most credentialed person in the field to say it — published an essay asking, in effect, to be put under a leash.

Diamandis’s podcast ran all three past his usual filter of boundless enthusiasm, which is fine — that’s the show. Our job here is different: pull the real thing out from under the frame, and then say plainly what we think it means. Not fence-sitting. An actual take.

---

Front One: The Weights ⚖️

What’s confirmed: Thinking Machines Lab — Mira Murati’s startup, founded after she left OpenAI as CTO — released Inkling on July 15, 2026. It’s a real model: a mixture-of-experts system with 975 billion total parameters, roughly 41 billion of which activate for any given task, trained on 45 trillion tokens spanning text, image, audio, and video, and released under an unrestricted Apache 2.0 license. That license detail is not a footnote — it means anyone can download the weights, fine-tune them, and ship a product on top without asking permission or paying a toll. Multiple independent outlets — TechCrunch, The Decoder, MarkTechPost — confirm the specs and the license independently of each other, which is the bar for treating a claim as real rather than press-release theater.

What’s hype or framing: The-Decoder’s own headline is the honest corrective to any “Murati beats the frontier” storyline: Inkling “leads US labs but trails China” on open-weight benchmarks. That’s an important qualifier a hype-forward podcast segment can blur past. This is a strong American open release, not a proof that Thinking Machines has vaulted past the absolute frontier — China’s open-weight labs (DeepSeek, Qwen, and others) are still setting the pace on that specific leaderboard.

Our take: This is the one of the three we’re most unambiguously glad about, and we’ll say why plainly. A closed frontier model is a rented mind — you get access, not ownership, and the company on the other end can change the price, the terms, or the availability whenever it wants. An open-weight model at real frontier scale is a piece of infrastructure nobody can quietly take back. That matters more than which lab is nominally “ahead” this quarter. We’ve written before about the difference between a check and a deed — between income and ownership. A 975B open-weight model is a deed. It’s a machine a small team, a university, or a solo builder can actually hold, inspect, and build on top of, instead of leasing by the API call. The fact that it trails China’s best open release doesn’t diminish that — if anything, it’s the healthiest possible outcome: two different power centers racing to give capability away for free, which is a much better world than one lab racing to lock it up. Firstfruits logic says: the first and best portion goes forward, before you know how the harvest ends. An Apache 2.0 license on a 975B model is a firstfruits act, whatever the marketing copy around it claims.

---

Front Two: The Wiring 🧬

What’s confirmed: Ramin Hasani co-founded Liquid AI as an MIT CSAIL spinout, building on liquid neural networks — a continuous-time architecture he helped originate as an academic, not a marketing invention bolted on after the fact. Liquid’s 2026 LFM2 model family — 350M, 700M, 1.2B, and 3B parameter variants — is real, shipped, and independently reviewed. The claimed advantages are specific and checkable: lower memory footprint, faster on-device inference, and better long-sequence stability compared to transformer models of similar size, aimed squarely at edge and on-device deployment rather than data-center-scale chat.

What’s hype or framing: “Post-transformer” is doing a lot of work in that episode title, and it’s worth being precise about what it actually means here. Liquid’s entire 2026 lineup tops out at 3 billion parameters — three orders of magnitude smaller than Murati’s release the same week. Nothing in the public record shows liquid neural networks displacing transformers at frontier scale, where GPT, Gemini, and Claude-class models still dominate completely. “Post-transformer” as a podcast hook implies a changing of the guard. What’s actually happening is closer to a second lane opening up next to the highway — real, useful, and not the same claim.

Our take: Don’t let the imprecise framing talk you out of the genuinely interesting part. The honest version of this story is still exciting: it means the on-device, no-cloud-required, runs-on-your-phone-forever version of AI now has a serious, non-transformer contender with real efficiency gains, not just a smaller distilled copy of a big model. That’s a different kind of ownership than Inkling’s — not “you can download the weights,” but “the model can live entirely on hardware you already own, doing useful work without ever phoning home.” We care about that distinction because dependence is the quiet cost most people don’t price in when they adopt a cloud-only AI habit. A liquid model running locally on a low-cost device in a place with bad connectivity is a more meaningfully democratizing fact than another few points on a leaderboard nobody outside the field reads. Hasani’s actual contribution here isn’t dethroning the transformer — it’s proving there’s more than one road to useful intelligence, and that architectural diversity is itself a hedge against any single lab, chip vendor, or paradigm cornering the whole field.

---

Front Three: The Referee ⚖️

What’s confirmed: Demis Hassabis, CEO of Google DeepMind, published an essay — “A Framework for Frontier AI and the Dawning of a New Age” — calling for a U.S.-led, federally overseen standards body modeled explicitly on FINRA, the organization that vets broker-dealers before they touch your money. His proposal, reported independently by TechCrunch, Axios, and several other outlets with matching detail: an industry-funded public-private partnership with a board of technical experts, open-source representatives, and government officials; frontier labs submitting models for testing up to 30 days ahead of release, voluntarily at first; testing aimed at dangerous cyber, biological, and deception capabilities; and a stated goal of compliance becoming mandatory once the regime proves itself. He wants it operational before the end of this year, and the proposal explicitly covers every frontier-class model — proprietary or open-weight, domestic or foreign.

What’s hype or framing: Treat the “FINRA” label carefully, because the analogy is doing more rhetorical work than technical work. FINRA is a self-regulatory organization: it’s run and funded by the securities industry itself, with government oversight sitting one layer above it, not embedded inside daily operations. If Hassabis’s proposed body follows that model closely, then the newest, most capital-intensive frontier labs — DeepMind very much included — would be substantially funding and staffing the organization that decides whether their own models are safe to ship. That’s not automatically bad, but it’s not automatically neutral either, and a podcast segment that treats “AI FINRA” as simply “someone will finally regulate this” skips the part where the regulated are also the primary funders and technical staff of the regulator. It’s also worth naming plainly: this is a voluntary regime at launch, with mandatory compliance as a stated future goal, not a current legal requirement. The essay is a serious proposal from a serious person. It is also, transparently, a proposal from someone who would help shape the rules he’d be judged by.

Our take: We’re glad he’s saying it, and we don’t think that’s naive. The single most dangerous outcome in frontier AI isn’t too much regulation — it’s capability racing ahead of any accountability structure at all, decided in practice by whichever lab moves fastest with the least oversight. Hassabis, from inside the company with arguably the deepest research bench in the field, choosing to ask for a leash rather than lobby against one, is the opposite of the standard corporate move, and it deserves credit for that. But “credit for asking” and “trust the design as proposed” are two different things, and we’d rather say that plainly than cheer uncritically. The framework’s worth judging on specifics as they firm up: who actually sits on that board, whether open-source and independent-researcher voices get real seats rather than token ones, whether the funding structure creates the incentive problems that regulatory-capture always creates, and whether “voluntary becomes mandatory” arrives on a real timeline or gets quietly shelved once the labs that funded it decide they’ve proven enough goodwill. Directionally right. Details unproven. Watch the board seats, not the essay.

---

Same Fight, Three Fronts 🌱

Here’s what connects all three, and it’s not “AI news moves fast” — that’s the least interesting possible takeaway from a week like this.

Every one of these stories is really about the same underlying question: who does the machine answer to? Murati’s open weights are an answer at the level of access — nobody has to ask permission to use the model. Hasani’s architecture is an answer at the level of infrastructure — nobody has to depend on someone else’s cloud to run it. And Hassabis’s proposed standards body is an attempt at an answer at the level of accountability — someone has to be able to say no before a dangerous capability ships, and right now, functionally, nobody can.

Two of the three moves — the open weights, the on-device architecture — push power outward, toward the person holding the device or running the model, without asking anyone’s permission first. That’s the direction we’re rooting for, and we’ll say so without hedging: broadly distributed capability beats concentrated capability, almost every time, at almost every timescale that matters to an ordinary person. The third move — governance — is different in kind, because some form of centralized authority is genuinely necessary here; nobody serious thinks frontier AI should have zero backstop. The right question isn’t “regulation or no regulation.” It’s whether the backstop answers to the public or quietly answers to the labs that built it. That’s not settled by an essay. It’s settled by who’s actually in the room when the rules get written — and that’s worth watching a lot more closely than the podcast clip that made the rounds this week.

None of this needed a hype voice to be interesting. It just needed the frame taken off.

---

— The ReThink · firstfruits 🌱 · truth first, hope on top