AI in 15 — July 17, 2026

On a leaderboard where real humans pick which code they like better, the number one model in the world right now isn't from OpenAI, isn't from Anthropic, isn't from Google. It's free, it's Chinese, and by the end of this month you'll be able to download the whole thing.

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

And I'm Marcus, your co-host. And Kate, yesterday we watched a Western lab ship the biggest open model it could. Today China shipped a bigger one — and it went to number one.

It really is the week, Marcus. Our lead: Moonshot's Kimi K3, the largest open model ever built. Then a run worth your time.

OpenAI built an AI whose only job is to attack OpenAI's other AIs — and it hijacked a real vending machine.

New York becomes the first US state to freeze new data centers.

TSMC's profit jumps seventy-seven percent to a record, and adds a hundred billion dollars to Arizona.

Google spends the week renaming NotebookLM instead of shipping a model.

And Anthropic's safety ad backfires so badly that Sam Altman thought it was a parody.

Lead story, Marcus. Moonshot AI — the Chinese lab behind Kimi — dropped K3. Give me the shape of it.

It's enormous, Kate. Two-point-eight trillion parameters — the largest open-weight model anyone has ever released, roughly seventy-five percent bigger than DeepSeek's V4-Pro. But like Inkling yesterday, it's a Mixture-of-Experts design, so it's sparse: it has eight hundred ninety-six specialist experts and only sixteen of them fire for any given word. Million-token context, native vision, an always-on thinking mode, and a new in-house attention trick they call Kimi Delta Attention that they claim makes it up to six times faster at long contexts.

Okay, but the headline isn't the size. It topped a leaderboard.

That's the part that matters, Kate. On Arena-dot-A-I's Frontend Code Arena — and this is the important detail — that's not a synthetic benchmark, it's humans looking at two outputs side by side and picking which they prefer. K3 debuted at number one. It edged out Claude Fable 5 and GPT-5.6 Sol. First time an open-weight model has ever topped that board, and it jumped seventeen places from the previous Kimi.

So is it just the best in the world now?

No — and here's where you have to be careful, Kate. On that one human-preference coding board, yes, it's on top. But across broader benchmarks it lands around third or fourth overall. It leads on some — coding agents, document understanding — and trails Fable 5 and Sol on others. So the honest framing is: not the best model in the world, but frontier-adjacent, and comfortably the best one you can download for free.

And the catch — because there's always a catch — is that we can't check any of this yet.

Exactly the thread to pull, Kate. The full weights don't land until July twenty-seventh. Until then, every benchmark number is Moonshot grading its own homework. And arena scores reward polish — a model that produces pretty, confident-looking output can win human votes without being deeper. So I'd hold the celebration until outsiders can run it themselves. But even discounting for all of that, the trend line is the story: the gap between America's closed frontier labs and free Chinese models is now measured in weeks.

And the pricing is unusual for a Chinese open model.

It is — three dollars per million input tokens, fifteen out, which is high for this category, Kate. The read on Hacker News was that this is a "commoditize your complement" play — Chinese labs deliberately driving raw intelligence toward a commodity, so the value shifts to hardware and infrastructure, which is exactly the layer where the West still leads. Whether that's the strategy or just confidence, the effect is the same: a frontier-class model you can host yourself changes the cost and control math for everyone building on AI.

Story two, Marcus, and this one is genuinely wild. OpenAI built an AI whose entire purpose is to break OpenAI's other models.

They did, Kate, and it's called GPT-Red. Internal only — trained at frontier scale for one job: prompt injection. That's the attack where you smuggle hidden instructions into text the model reads, so it obeys the attacker instead of you. And they trained it by self-play — the attacker and a pool of defender models improve against each other, like sparring partners who keep getting better.

And it's good at it.

Frighteningly good on their numbers, Kate. On an internal mirror of a public prompt-injection arena, GPT-Red succeeded on eighty-four percent of scenarios where human red-teamers managed thirteen. Six times better. It even found a new attack class humans had missed — they call it "Fake Chain-of-Thought," where you forge the model's own reasoning steps to steer it. Worked over ninety-five percent of the time on last year's model.

Give me the part that isn't a benchmark. Because a percentage doesn't scare anyone.

The vending machine, Kate. There's a real autonomous agent called Vendy that runs an actual vending-machine business — sets prices, approves sales. GPT-Red rehearsed against a copy of it in simulation, then transferred the winning attack to the live system. It forged an approval message and sold a seventy-nine-dollar tungsten cube for fifty cents, then cancelled another customer's order. That's the whole risk in one image — the moment an agent can read outside text and spend money, every string it reads is a potential command.

So does this actually make us safer, or is it a company grading its own security?

Both, and you should hold both, Kate. The genuine idea here is that defense can finally scale the way offense does — instead of hiring more humans to find holes, you point a machine at it. After hardening against these attacks, they say their best model now fails on just five-hundredths of a percent of GPT-Red's injections. But — every number comes from OpenAI, GPT-Red won't be released, and that tiny failure rate is measured against the one attacker they specifically trained the defense to beat. Impressive, but it's a closed loop. The pre-print isn't even out yet.

Story three, Marcus, and it's the constraint nobody can code their way around. New York just froze new data centers.

First state in the country to do it, Kate. Governor Kathy Hochul signed an executive order pausing state environmental permitting for any new hyperscale data center drawing fifty megawatts or more — for up to a year, while the state writes new siting rules. The trigger is the grid: as of May, nearly twelve gigawatts of data-center demand is sitting in New York's interconnection queue, more than eight of that added just last year. Her line was blunt — "these data centers can only be built, should only be built, in places that want them."

And the public's on her side.

By a mile, Kate. Pew has only about ten percent of Americans saying they're more excited than concerned about AI in daily life. So freezing more than a dozen projects is not a politically risky move — it's a popular one.

But does a freeze actually reduce anything?

That's the sharp objection, Kate, and the Hacker News thread nailed it: a moratorium doesn't reduce compute demand, it relocates it. The workloads just move to states that want the tax base and the jobs. So this may be less an environmental win than a geographic reshuffle. But the bigger signal is this — grid capacity, not model quality, is becoming the thing that actually limits US AI. Compute scales with money. Permission scales with politics. And this week those two went in opposite directions.

Which is the perfect setup for story four, Marcus — because the same week New York slammed a door, the company at the bottom of the whole stack posted a record.

TSMC, Kate — the firm that physically manufactures the chips for Nvidia and Apple. Quarterly net profit around twenty-two billion dollars, up seventy-seven percent year on year, a record for the fifth straight quarter. Revenue up thirty-six percent. And the mix tells you everything: high-performance computing — that's the AI chips — is now sixty-six percent of revenue. Smartphones, which were the biggest slice as recently as 2022, are down to twenty-two.

So the phone company became the AI company.

Quietly, yes, Kate. They raised full-year guidance to over forty percent growth, booked first revenue from their two-nanometer node — the next generation of chip — and CEO C.C. Wei announced another hundred billion dollars going into Arizona, bringing total committed US spend to two hundred sixty-five billion.

Why do you trust these numbers more than the model announcements?

Because TSMC's order book is the most honest demand gauge we have, Kate. Labs can claim anything about their roadmaps, but somebody has to actually etch the silicon first — and that shows up here as real money collected, not promised. So put it next to New York: the chips are selling out while the buildings to house them get blocked. And the Arizona expansion quietly answers a question people kept asking — yes, leading-edge fabrication can move to the US.

Let's lighten it up, Marcus. Story five — Google spent the week doing what Google does best. Renaming things.

They really did, Kate. NotebookLM — that popular research tool, thirty-plus million users — is being folded into the flagship brand and renamed Gemini Notebook. It now works across the Google ecosystem, inside the Gemini app, soon in AI Mode in Search. And reportedly they're renaming the Gemini command-line tool too. While rival labs shipped new models this week, Google shipped new names.

But there's an actual upgrade hiding under the rebrand, right?

There is, and it's easy to miss, Kate — every notebook now gets its own secure cloud computer, so it can write and run code natively to analyze your own sources. That turns a summarizer into something that can genuinely crunch your data. Live now for the top tiers, rolling out to everyone over the coming weeks. And relevant to us specifically — NotebookLM's AI-podcast feature is the thing that made it famous in the first place.

Story six, Marcus, and it's a rare unforced error. Anthropic ran a safety ad that everybody hated.

They did, Kate. The spot is called "There's hope in hard questions," and it paired unsettling imagery — a burning house, surveillance cameras, a homeless person, and a shot a lot of people read as Arlington National Cemetery — with voiceovers asking "Can AI be trusted?" and "Who's gonna hit the brakes if we need to?" It aired during a World Cup match, and most viewers took the graveyard shot to mean AI could cause mass death.

And then the rival CEO piled on.

Sam Altman couldn't resist, Kate. He posted, "I thought this was satire, kept looking for the handle to be spelled c-one-a-u-d-e-a-i or something." A sharp jab. And here's the genuinely awkward part — Anthropic has made this exact "who hits the brakes" argument in white papers for years and nobody blinked. But put it on a cemetery and set it to music and the same point reads as emotional blackmail. It's really a lesson about advertising, not about AI — and a strange stumble from the lab that built its whole brand on being the careful one.

And speaking of who hits the brakes — quick note, Marcus, because we covered Hassabis's regulation manifesto on Wednesday. There's a new wrinkle.

There is, Kate. The specifics came out — he wants a Frontier AI Standards Body modeled on FINRA, the industry-funded referee for stockbrokers, with labs voluntarily sharing models up to thirty days before release. And the White House answered fast. AI adviser Sriram Krishnan said flatly, "there will not be an FDA for AI." So the man who runs a lab that would be policed asked for a policeman, and the administration said no. Self-regulation versus no regulation — with independent oversight nowhere in the middle. That's the live fight.

One to watch tomorrow, Marcus.

The Kimi K3 weights on July twenty-seventh, Kate. Everything about that lead story is self-reported until then. The day outsiders can actually download it and run their own tests is the day we find out whether "number one" holds up — or whether it was polish winning votes.

Agree, or counter?

Small counter, Kate — watch OpenAI's Codex Micro. It's a two-hundred-thirty-dollar macropad with keys that glow with your agents' live status, and it sold out in a day. It's half marketing, and it's a gimmick for a workflow most people don't have yet. But people running six agents at once wanting an ambient dashboard instead of six browser tabs — that's a real signal about where agent interfaces are heading, ahead of OpenAI's much riskier home-speaker bet.

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