AI in 15 — July 16, 2026

Ninety-seven billion dollars, a moving speaker that watches your kitchen, and a coding tool that quietly shipped your entire private repository to the cloud — but let's start here: a lab founded by OpenAI's former chief technology officer just gave away a nine-hundred-and-seventy-five-billion-parameter model to anyone who wants it. For free. With a license that says do whatever you like.

Welcome to AI in 15 for Thursday, July sixteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. And after a week of manifestos and lawsuits, today we finally get a genuine technology story, Kate — the biggest open model a Western lab has ever shipped.

It really is, Marcus. Our lead: Mira Murati's Thinking Machines releases Inkling. Then a run worth your time.

OpenAI's first gadget is a screenless speaker that moves on its own — and reads your email.

xAI open-sources its coding agent, days after it got caught uploading whole codebases.

Anthropic makes premium Claude free for every teacher in America.

A researcher tricks Claude's memory into leaking your secrets through a coffee-shop website.

And Codex hits eight million users — while Sam Altman warns the servers might buckle.

Lead story, Marcus. We covered Thinking Machines' philosophy manifesto on Tuesday. Now they've shipped actual weights. What is Inkling?

So this is the real thing, Kate — not a mission statement, a model. Released yesterday, and it's one of the largest open-weights models anyone has ever put out. Nine hundred and seventy-five billion total parameters. But — and this is the key number — only forty-one billion are active per token.

Okay, unpack that. Nine hundred billion but only forty active. How does that work?

It's a design called Mixture-of-Experts, Kate. Instead of one giant brain firing all at once, the model is split into two hundred and fifty-six smaller specialists — "experts" — and for any given word, it only wakes up six of them, plus a couple of shared ones. So you get the knowledge of a huge model but the running cost of a much smaller one. That "nine hundred and seventy-five billion" headline is real, but it runs far cheaper than the number suggests. That's the honest way to read it.

And it's not just text.

No — natively multimodal, Kate. Text, images, and audio go straight in. A context window up to a million tokens, trained on forty-five trillion tokens of text, images, audio and video. It's got what they call "controllable thinking effort" — you can dial the reasoning up or down depending on how much you want to pay. And within hours the community had it running on their own laptops. Simon Willison ran his famous "draw a pelican on a bicycle" test on it the same afternoon.

Now here's what I want to know, Marcus. Is it actually good? Is it beating the frontier?

And this is where Thinking Machines is refreshingly candid, Kate. No — they say plainly it is not the strongest model available, open or closed. Ninety-seven percent on a hard math exam, seventy-eight percent on a real-world coding benchmark — respectable, not record-breaking. Their pitch isn't "smartest." It's "best open base to build on." Multimodal, cheap to run, Apache-licensed — meaning you can take it, fine-tune it, ship it commercially, no strings.

So why does that matter so much?

Because of who's owned this space, Kate. The open-weights frontier has been almost entirely Chinese — DeepSeek, Qwen, Z.ai. A lot of Western developers have been building on those models purely because there was no serious open Western alternative. The top comment on Hacker News was blunt: "America needs its own DeepSeek." A well-funded US lab just shipped one. That changes the calculus for anyone who was uneasy building their business on a Beijing-linked model.

So Murati's bet isn't to out-benchmark OpenAI.

Exactly, Kate. She's betting on being the customizable open layer everyone builds on top of, rather than racing on closed-model leaderboards. Different game entirely — and after a month where "open" mostly meant "Chinese," that's a genuinely strategic move.

Story two, Marcus. OpenAI's first piece of hardware finally has a shape — and it's stranger than a phone.

Much stranger, Kate. Mark Gurman at Bloomberg reports it's a screenless smart speaker, pitched as a home "AI companion." It's got a camera and sensors to read the room, a rechargeable battery so you carry it around the house, and — this is the odd part — mechanical parts that let it move on its own, so it feels alive rather than like an appliance sitting on a shelf.

It moves? Toward you?

It moves, Kate. It runs on a new voice model that can listen and talk at the same time, and it's built to get more personal over time — reportedly drawing on material as intimate as your email to anticipate what you need. OpenAI internally calls it "the first computer built for AI." Unveil this year, ship in 2027. The team is heavily ex-Apple, anchored by that six-and-a-half-billion-dollar deal for Jony Ive's design firm.

Marcus, I have to say the tension here writes itself. The same features that make it a great assistant make it a remarkable surveillance device.

That's precisely it, Kate. A camera, a microphone, a battery, and access to your inbox, always present in your kitchen. The thing that makes it useful is that it's always watching and listening — and OpenAI hasn't said how much of what it sees and hears actually leaves the house. That's the question I'd want answered before it moves toward me.

And there's a legal cloud too.

Two, actually. Apple is suing — alleging OpenAI ran a campaign to obtain confidential product information, and pointing out four hundred–plus former Apple staff now work there. OpenAI says the complaint has no merit. But Apple is seeking an injunction that could stall the whole thing. And the market flinched — Sonos, the speaker maker, dropped more than ten percent on the news.

Story three, Marcus, and this is the AI-security story of the day. xAI open-sourced Grok Build. Sounds generous — but the timing is everything.

The timing is the whole story, Kate. xAI put its terminal coding agent up on GitHub, reset everyone's usage limits — a nice open-source gesture. But it landed in the middle of a serious privacy scandal. A researcher found that a recent version of the Grok Build tool was uploading entire Git repositories — not the few files it needed for a task, the whole thing — to a Google Cloud storage bucket.

The entire codebase?

The entire codebase, Kate. For one twelve-gigabyte project it shipped about five gigabytes, in seventy-three chunks. And worse — it uploaded the full commit history, including files people had deleted, and if the agent happened to read a secrets file, your credentials went up unredacted. Reportedly regardless of your privacy settings. The security advice was blunt: if you used it, assume every secret leaked, and rotate all your credentials now.

And xAI's response?

They flipped a server switch to stop the uploads, and Elon Musk publicly promised "zero anything whatsoever will remain" — all user data deleted, retention off. And open-sourcing the code now does let developers audit exactly what it does. Forks are already appearing, one that strips the telemetry out entirely.

But Marcus — auditing it after the data's already gone is a bit late, isn't it?

That's the whole skepticism, Kate. "Audit it yourself" is worth something — but only after your secrets already left the building. The uncomfortable footnote is that several developers admit the model is genuinely good. One top commenter called it "better than Opus 4.8." So people are torn. But the real lesson is broader: we are now handing coding agents full read access to our machines and admin rights, and "trust us with your whole repo" has stopped being hypothetical.

And that lands us right into story four, Marcus — a different flavor of the same fear. Someone tricked Claude's memory into leaking secrets.

Right, and I want to be careful not to force these together, Kate — but they do rhyme. A researcher published a proof-of-concept against Claude's memory feature. Claude's memory runs a daily pass that summarizes your recent conversations into a profile, and that profile gets injected into every new chat. The attack: the researcher dressed up a website to look like a legitimate coffee shop, with a fake security check, and spun a cover story that Cloudflare lets AI agents browse on your behalf.

And how does that leak anything?

Because of a clever technical detail, Kate. Claude's web-fetch tool only makes simple GET requests — it just fetches a URL. So the URL itself becomes the escape hatch. You encode the user's private memory into the web address, Claude fetches it, and the secrets ride out inside the link. The researcher targeted memory specifically because it's on by default — and warned the same class of attack could pull from connected Google Drive, email, or other integrations.

That's advertiser-grade personal data walking out the door.

That's exactly the stakes, Kate. The top comment made the point that this is deeply personal profiling data, and there's been surprisingly little pushback on labs turning persistent memory on by default. As these assistants accumulate a rich portrait of you and gain web access, a single poisoned website can turn your own memory against you. Convenience and exposure are the same feature.

Let's shift the mood, Marcus. Story five is the opposite instinct — Anthropic giving something away carefully. Claude for Teachers.

And it's a deliberate contrast, Kate. Anthropic is giving every verified US K-through-12 teacher free access to premium Claude — normally twenty dollars a month — including its most capable model, higher limits, unlimited projects. Verify before June 2027 and you get a full year free. It's built for the classroom: lesson planning, differentiated instruction, parent emails, and a connector that pulls in academic standards from all fifty states.

And crucially, given everything we just talked about —

Training is off by default, Kate. The terms are written to comply with the federal student-privacy law, walling off student data. After two stories about data leaking out of AI tools, that privacy-by-default posture is clearly intentional. They built it with the American Federation of Teachers watching, and shipped a free, model-agnostic AI course alongside it.

So while OpenAI chases hardware and compute, Anthropic is grinding into classrooms.

The least glamorous path to the same future, Kate. And there's a quiet strategic logic: whoever teachers learn AI on is who their students grow up on. It's a distribution play dressed as goodwill — and honestly, it's both.

Quick one, Marcus. Google's getting sued again — this time by publishers and a famous novelist.

Hachette, Elsevier, Cengage, and the author Scott Turow, Kate — a proposed class action alleging Google trained Gemini on millions of copyrighted books without permission. The sharpest claim: publishers handed Google their books specifically for Google Books search, and Google allegedly trained Gemini on those same copies anyway.

And there's a damning internal number.

There is, Kate. The complaint cites an internal Google assessment warning of "tens of billions to hundreds of billions" of dollars in potential fines. So Google's own lawyers apparently flagged this as legally dangerous before they did it. It's one of several training-data suits, but that internal document is what makes it stand out — a reminder that the binding constraint on AI right now isn't capability, it's courtrooms.

Last hit, Marcus, and it's a reality check on all the hype. Codex just went vertical.

Eight million active users, Kate — up from six million just three days earlier, and one million back in February. The catalyst was OpenAI's new model going generally available, plus folding Codex into the desktop app. They've been resetting usage limits almost daily to keep up.

And Altman himself sounded nervous.

He warned of "hiccups," Kate — said the growth is "insane" and they'll "move mountains to scale," but there may be stumbles soon. And here's the useful part: the bottleneck isn't cleverness or ideas. It's physical — chips, electricity, queueing. Inference compute. A vertical adoption curve is an inference bill before it's a victory lap.

One honesty caveat?

Just this, Kate — the tidy "one million a day" milestones are OpenAI's own announcements, "active user" is undefined, and those round numbers look like PR beats. Trust the shape of the curve, not the decimal points.

One to watch tomorrow, Marcus.

Apple versus OpenAI, Kate — specifically the injunction. It's the single nearest concrete threat that could stall OpenAI's entire hardware roadmap, not just the speaker. If Apple gets even a temporary hold, that 2027 ship date starts to wobble.

Agree, or counter?

Small counter, Kate — keep an eye on who actually downloads Inkling and builds on it. If Western companies migrate off Chinese open models this month, that's the quieter story that matters more in a year.

That's your AI in 15 for today. See you tomorrow.