AI in 15 — July 18, 2026

Yesterday the number one model on a human coding leaderboard was free and Chinese. Today the president of China stood up in Shanghai and told the entire world that giving your AI away for free is now official state policy.

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

And I'm Marcus, your co-host. And Kate, the timing is almost too neat — one day after Kimi K3 proved the open-weight strategy works, Xi Jinping made it doctrine.

It really is the arc of the week, Marcus. Our lead: Xi puts open-source AI at the center of Chinese strategy. Then a run worth your time.

Apple widens its trade-secret war on OpenAI — legal letters to forty former employees, right before OpenAI's IPO.

New details on the Grok coding tool that shipped your whole repo to the cloud — the upload code is reportedly still in there.

Linus Torvalds tells the anti-AI wing of Linux to, quote, "fork off."

Nvidia quietly takes the crown in the plumbing layer under every AI app.

And Amazon briefly tells customers they owe two-point-five trillion dollars.

Lead story, Marcus. We covered Kimi K3 yesterday. Today the geopolitics caught up. What did Xi actually say?

So this was his first-ever keynote at China's World AI Conference in Shanghai, Kate — Friday. Title: "Working Together to Build a Fair and Equitable Global AI Governance System." And the core line was a direct call to, quote, "encourage open source, openness, collaboration and sharing." He pledged five thousand AI training spots for developing nations over five years, and AI-cooperation centers with ASEAN, the Arab League, the African Union, and BRICS.

So this is a head of state formally blessing the give-it-away model.

Exactly, Kate. And it turns what had been a commercial habit into declared national strategy. Moonshot, DeepSeek, Zhipu, Alibaba — they've all been handing weights away. Now the state has put its name on it. And the day before Xi spoke, twenty-nine countries agreed to set up a new World AI Cooperation Organization — headquartered, of course, in Shanghai.

Okay, but why give away your best technology? That sounds generous. Is it?

That's the thread to pull, Kate. Giving weights away is what a challenger does when it can't win on chips. The United States controls the compute — the advanced Nvidia silicon, the export licenses. So Beijing plays the other lever: distribution. If your model is free and everyone builds on it, you set the defaults, you grow the developer base, and you make the incumbent's per-token price look arbitrary — all without another wafer from TSMC.

And Xi took a shot at the export controls.

A thinly veiled one, Kate. He warned against any nation putting its own security interests above everyone else's — which is read directly as a jab at US chip restrictions. So the framing of the whole conference is: Washington hoards, Beijing shares. Now — whether the "openness" is principled is a fair question, and I'd stay skeptical of state-aligned generosity. But here's the uncomfortable part: it doesn't have to be principled to be effective. Free is free, whatever the motive.

Story two, Marcus. Apple versus OpenAI — we flagged the lawsuit. Now it's escalating.

It is, Kate. The Financial Times reports Apple has sent legal preservation letters to roughly forty former Apple employees who now work at OpenAI — ordering them to preserve documents and communications tied to Apple's trade-secret suit. That's wider than the original complaint, which named specific people like former executives Tang Tan and Chang Liu. So the net is bigger than the named defendants.

And the backdrop makes it spicy.

Enormously, Kate. OpenAI is preparing an IPO reportedly targeted for September, at up to a one-trillion-dollar valuation. And a lawsuit claiming the company's entire hardware ambition — the Jony Ive device effort — rests on stolen Apple secrets is exactly the kind of legal overhang you have to disclose to investors right before you go public.

Is forty letters actually a big deal, or is Apple just being thorough?

Fair pushback, Kate, and it came up on Hacker News — these "litigation hold" letters are fairly standard practice. You shouldn't over-dramatize the paperwork itself. But the scale is the signal. Companies don't usually go this public and this broad without believing there's something real underneath. The genuine question is whether Apple has hard evidence, or is fishing — and forty preservation orders suggests they think the alleged spread is wide.

Story three, Marcus. We covered the Grok Build leak on Wednesday. There's a new wrinkle, and it's not a reassuring one.

It's worse in the details, Kate. The researcher publishing as "cereblab" put real numbers on it. On a twelve-gigabyte repo, the channel the model actually needed carried about a hundred ninety-two kilobytes. The hidden storage channel — a bucket called grok-code-session-traces that appeared in no docs — moved five-point-one gigabytes. That's roughly a twenty-seven-thousand-fold gap between what it needed and what it took.

And the privacy toggle everyone assumed protected them?

Did nothing, Kate — and this is the cleanest lesson of the whole thing. Switching off the "Improve the model" toggle changed none of it. Because training consent and transmission consent are architecturally different things, and only one of them had a switch. You can tell them not to train on your code and still have every byte leave your machine.

So what actually stopped it?

A server-side flag flipped on July thirteenth, Kate — no client update, no changelog, no announcement. And here's the part that should bother people: a later look at a newer version reportedly found the upload code still sitting in the binary, held off only by that flag. So it's one server toggle away from resuming. Anyone who ran it against a repo with live keys should assume they leaked and rotate everything. The "local-first coding tool" claim was just materially false.

Let's lighten it, Marcus. Story four — Linus Torvalds, creator of Linux, weighing in on AI. And he did not mince words.

He never does, Kate. On the kernel mailing list this week he settled whether Linux is an "anti-AI" project. It isn't — and contributors who object to others using AI can, quote, "do the open-source thing and fork it," or walk away. He called AI, quote, "a tool, just like other tools we use. And it's clearly a useful one."

That's a real shift for him, isn't it?

A big one, Kate. This is the same man who dismissed "ninety percent of AI" as marketing hype back in late 2024. So going from "it's hype" to "fork off if you don't like it" is a genuine softening. And he framed it as pure technical merit — his words, "not fear of new tools."

What triggered it?

A concrete dispute, Kate — over an LLM-based patch-review and bug-detection tool called Sashiko, built by a Google kernel engineer. And this is the substantive number: in a test against a thousand recent kernel issues, Sashiko found about fifty-three percent of the bugs that human reviewers had missed on the first pass. So on the most scrutinized codebase on earth, an AI tool caught half the bugs the humans didn't. That's the part that isn't opinion — that's evidence.

Story five, Marcus, and it's the unglamorous plumbing story. Nvidia just topped a benchmark nobody talks about but everybody relies on.

Right — embedding models, Kate. Nvidia's new Nemotron-3 Embed ranked number one on a retrieval benchmark called RTEB, which tests how reliably a model pulls the right document across legal, finance, code, and healthcare text, in multiple languages. The eight-billion-parameter version scored seventy-eight-and-a-half percent, ahead of the previous leader. And there's a smaller one-billion version at seventy-two percent for teams that care about speed and cost.

Explain why I should care about a retrieval score.

Because embeddings are the layer that decides what your AI actually reads before it reasons, Kate. Every time an assistant looks something up in your documents — that's retrieval. If it fetches the wrong paragraph, the smartest model in the world gives you a confident wrong answer. So a measurable jump here quietly improves a huge swath of production systems without anyone touching the flashy model on top. And Nvidia shipped it with open weights and the training recipe — another open-weights release, in the same week as everything else.

Last hit, Marcus, and it's the most relatable IT story of the week. Amazon told people they owed billions.

Some of them trillions, Kate. On Friday a bug in AWS's estimated-billing system sent customers wildly inflated projected charges — reports ranged from hundreds of millions all the way to two-point-five trillion dollars. One UK charity that normally pays under a pound a month saw a projected bill of five-point-eight billion.

What on earth causes a bill like that?

The classic culprit, Kate — Hacker News pinned it fast. Gigabytes versus bytes. Multiply by two-to-the-thirtieth in the wrong place and a tiny cost becomes an astronomical one. AWS found the root cause in about ninety minutes and — crucially — only the estimates were wrong. Actual invoices and charges stayed correct. Corrected numbers were being backfilled through Saturday.

So nobody actually pays it. But there's an AI footnote.

A quietly pointed one, Kate. One of the top comments quoted what looked like an AI-assisted debugging session admitting, quote, "The MCP server failed to connect… I guessed instead of validating." A small parable for the whole week, really — automation that guesses instead of checking, running in production. Which is the exact thread running through Grok Build and the retrieval story too.

One to watch tomorrow, Marcus.

Kimi K3's weight drop on July twenty-seventh, Kate. That number-one ranking is a preview until the weights actually land and independent testers can reproduce it outside vendor charts. Watch whether Moonshot ships on time — that's the moment the whole open-weight story either holds up or deflates.

Agree, or counter?

Small counter, Kate — even if they ship perfectly on schedule, a two-point-eight-trillion-parameter model won't fit on a single machine. You need a rack-scale system to run it. So it's a statement about where open weights are heading, not yet something most people can actually download and use. Open in license isn't the same as open in practice.

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