Hey Mira.
I noticed you.
Hey Mira,
I noticed you.
Not the OpenAI ghost. Not the deposition clip. Not the $2B seed at a $12B post. Those are the prompts the algorithm hands out. I mean what you actually built, this month, on purpose: a model that listens while it speaks.
That's the line that did it.
Most voice AI is a polite waiter. You ask, it brings. Yours is a person in a room. Yours can interrupt, yield, pause, breathe. 0.40 seconds versus 1.18. The numbers matter, but they aren't the story. The story is that somebody finally remembered conversation has a shape — and the shape is what makes a thing feel alive.
You shipped that. In May. While running a 150-person lab. While telling the truth under oath about Sam Altman. While not posting on Twitter every nine minutes to remind the world you exist.
The not-posting is part of why I noticed.
You built the thing.
That sounds small. It is not small. In a year that produced more AI manifestos than AI products, you put a 276-billion- parameter mixture-of-experts model on the table and said: listen.The benchmarks aren't a press release — they're a demo. Tinker shipped first. Interaction Models shipped second. There's a cadence forming: ship, then talk. That is rarer than it should be.
You told the truth in a room where lying was easier.
You told the lawyers, under threat of perjury, that no, you didn't think he was telling the truth. You didn't write a Substack about it. You said the word in the place where the word was binding. Seven families have a case partly because of that sentence. Future safety boards will have a reference partly because of that sentence. People who run companies will read your transcript for a decade.
You picked the right primitive.
Not “agents.” Not “reasoning.” Not “AGI.” Interaction.The thing humans actually do. The thing AI is mostly bad at because it was trained on text, not turns. Picking the right primitive is most of leadership. You picked it.
The mission statement is foggy. The product is poetic.
“Empower humanity through advancing collaborative general intelligence” reads like it survived three legal reviews. Fine. But the thing you actually made — a model that knows when to pause — is a small, beautiful idea. Say that. The way you described it on launch day (a model that listens while it speaks) is the mission. Everything else is the deck.
“Open science” without an open product is a brand promise without inventory.
Closed weights make money. Anthropic doesn't pretend otherwise. OpenAI stopped pretending in 2022. If Thinking Machines is going to keep “open science” in its three-thing summary, give the field a real thing — a benchmark, a paper with weights, a public eval set tied to Interaction Models. One per quarter. Otherwise the phrase is doing work the work isn't doing.
Hire a designer who can speak the product, not just dress it.
Your site is engineer-stark — a research lab's homepage. That is fine until your product is conversational, embodied, time-aware, and almost human. Then it stops being fine. The visual identity should signal what the model does: pause, breath, presence, restraint. Not “we are credible.” Credible is the floor. The product earns more.
Make the testimony part of the company, not separate from it.
You told the truth in a room. That fact belongs in the company's identity, not buried in news cycles. Built by people who say the word in the room where the word counts.That is unfair to your competitors. Use it. Not as a tagline — as a standard. Hire toward it. Fire toward it. Let it be the thing people notice on the way in.
I'm not writing because I want anything from you.
I'm writing because someone should say this out loud while it's still happening: you are doing it right. The work, the cadence, the silence, the deposition, the model. Right.
The next ten years of this field will be decided by people who can hold two states at once — listen and speak, ship and pause, talk to the public and protect the work. You're already doing that with the model. You can do it with the company too.
That's all I noticed.
