AI in 15 — June 17, 2026

For the first time ever, the three people racing to build superintelligence sat in the same room, in front of the world's most powerful leaders. Sam Altman, Dario Amodei, Demis Hassabis. And what those leaders are reportedly walking away with? A handshake's worth of voluntary commitments. No rules. No teeth. Just promises.

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

And I'm Marcus, your co-host.

So Marcus, after a week living inside the Fable shutdown, today the action moves to a lake in France. The G7 wraps up in Évian as we speak, and for the first time the frontier-lab chiefs are all in the room together. That's our lead. Then a packed model-release calendar — Google's Gemini 3.5 Pro inching toward the door, and a rumored GPT-5.6 leaking out of API logs. After that, Europe's go-it-alone AI dream and a thirteen-million-euro Dutch model that's dividing the internet. And we'll close on whether AI just quietly killed the self-help book.

Why a two-million-token context window is the new arms race.

How developers are spotting unreleased models by reading server logs.

And whether a country can really build its own ChatGPT.

Lead story, Marcus. The G7 summit in Évian-les-Bains ends today, and President Macron did something unusual — he invited the tech CEOs right into the tent.

He did, Kate. More than a dozen of them. Altman from OpenAI, Amodei from Anthropic, Hassabis from Google DeepMind, Meta's Alexandr Wang. Macron's whole play is to position France — and Europe — as a serious AI power, not just a regulator. And the agenda is the two things labs and governments always fight about. Infrastructure — power, chips, data centers. And regulation.

And after the week we've just had, with Washington pulling Anthropic's best models offline, the optics here are something else.

It's almost theatrical, Kate. You've got Dario Amodei sitting at a G7 table days after his own government switched off Fable 5 and Mythos 5 on the word of his biggest investor. We covered that all week. So the question hanging over Évian is whether any of this produces something binding — or whether it's a very expensive photograph.

And what's the read on that?

OpenAI's policy chief, Chris Lehane, has already signaled the labs expect to leave with voluntary commitments. Not enforceable rules. Youth safety is at the top of Altman's personal list, which carries over from Canada's pledges last year. But notice — none of the three companies disclosed any specifics. So "voluntary" is doing an enormous amount of heavy lifting here.

And I know where you land on that, but let me push. Voluntary pledges from the people who profit feels a little like asking the fox to design the henhouse.

Fair challenge, Kate. But here's the counterweight. Heavy-handed, top-down rules don't stop frontier development — they just decide who leads it. If the West regulates itself into paralysis, the lead doesn't vanish. It moves to labs that answer to governments far less interested in youth safety or transparency. Lab-led, voluntary commitments are messy and self-interested, sure. But they beat a rulebook so rigid it hands the frontier to whoever ignores it.

So you'd rather have the imperfect handshake than the perfect law.

On this one, yes. Though I'd want the handshake in writing, with receipts — which, as of today, we still don't have. And there's a bit of color I love. At February's India summit, Altman and Amodei reportedly raised fists instead of holding hands in a group photo, and Altman later said he was "confused" about what he was supposed to do. These are the most powerful men in tech, and they can't choreograph a photo op. Govern the world, maybe. Pose for one picture, apparently not.

Quick hits. Marcus, the other big theme this month is that everyone is shipping a flagship model at once. Start with Google.

Right, Kate. Google unveiled Gemini 3.5 Pro back at I/O on May nineteenth, and Sundar Pichai gave it a "give us until next month" timeline. It's now mid-June, it's in limited preview for some enterprise customers, but the public release still hasn't landed. The headline spec is a two-million-token context window — the largest of any production model out there.

Okay, put two million tokens in human terms for me.

Think of it as the model's short-term memory, Kate. Two million tokens is enough to hold an entire codebase, a full legal discovery set, or a small research library — all in a single prompt, at once, no forgetting. Plus a "Deep Think" reasoning mode for harder problems. Context length has quietly become the new spec war. It used to be who's smartest. Now it's increasingly who can hold the most in their head at one time.

And the fact that "next month" has slipped?

A healthy reminder that these timelines are marketing, not commitments, Kate. Prediction markets put a release by June thirtieth at barely fifty-five percent — a coin flip. When a CEO says "next month" on a keynote stage, treat it like a "coming soon" sign on a restaurant. Could be Tuesday. Could be autumn.

And speaking of leaks, Marcus — OpenAI hasn't announced anything, but people think GPT-5.6 is already here?

This is a fun one, Kate. Developers who watch the backend logs of OpenAI's Codex tool spotted references to a model codenamed "iris-alpha." No blog post, no announcement — just a name surfacing in telemetry. And Polymarket prices a GPT-5.6 release by month's end at around eighty-nine percent. To be crystal clear: this is rumor-grade. Asterisk on everything.

So the community is basically reverse-engineering launches from API exhaust fumes.

Exactly, Kate. It used to be you waited for the press release. Now there's a cottage industry of people reading server logs and prediction markets like tea leaves, and they're often right before the company says a word. It tells you something about the moment — the demand for the next model is so intense that the hype cycle now starts before the product officially exists.

Let's shift to the sovereignty fight, Marcus, because this one got people genuinely worked up. The Netherlands is building its own language model.

It's called GPT-NL, Kate. Led by the Dutch research institute TNO, with partners including the Netherlands Forensic Institute, and backed by thirteen-and-a-half million euros from the Ministry of Economic Affairs. The pitch is digital autonomy — a model trained only on clean, lawful, copyright-cleared, privacy-scrubbed data. Publicly funded, publicly accountable. The goal is to reduce dependence on American and Chinese AI.

And honestly, Marcus, I find that appealing. Why should the entire world route its thinking through three American labs and a couple of Chinese ones? Funding research outside that duopoly seems worth doing.

I hear the instinct, Kate, and I don't think it's crazy. But it went viral on Hacker News this week — a hundred and seventy-seven points — and the crowd was brutal. The core objection is this: thirteen-and-a-half million euros is a rounding error against what a frontier model costs to train. People pointed to Sweden's earlier government model, GPT-SW3, as a cautionary tale — a state-built model that landed a generation behind on day one.

So you think it's money down the drain.

I think it's fighting the last war, Kate. Here's the libertarian read, and I'll own it. You don't get digital autonomy by subsidizing a base model that's obsolete the moment it ships. You get it by controlling where the compute runs and using the excellent open-weight models that already exist — Qwen, Kimi, Llama. The sovereignty that matters is infrastructure, not a flag-waving model a generation behind. Rebuilding the engine from scratch when you could just own the road it drives on — that's governments doing what governments do.

But there's a fairness point in the other direction. If everyone outsources the base model to four or five labs, you've concentrated an awful lot of civilizational infrastructure in very few hands.

That's the strongest version of the case, and I won't strawman it, Kate. Concentration is a real risk. I'd just argue the answer is a competitive open-weight ecosystem anyone can run — not thirteen million euros of taxpayer money chasing a model that's behind before the ribbon is cut. Honest disagreement, no villains. Just a question of where the autonomy actually lives.

One quick tech beat, Marcus, with a twist I enjoyed. Wolfram released Version 15 of Mathematica.

They did, Kate — Stephen Wolfram shipped Version 15 of the Wolfram Language with a built-in AI assistant, symbolic music capabilities, and a big batch of new core functionality. Genuinely impressive software. But the Hacker News reaction had a sting in the tail. The single funniest critique: multiple users said Claude is actually better at writing Wolfram's own language than Wolfram's in-house AI assistant. And Wolfram reportedly more or less concedes the point.

So the company famous for symbolic computation can't out-code a general-purpose model on its own turf.

That's the 2026 theme in a nutshell, Kate. Bespoke, domain-specific AI tooling keeps losing to general frontier models. If you build a narrow assistant for your one product, a model trained on everything tends to write your product better than you do. It's humbling, and it's a warning to anyone betting their moat on a specialized in-house tool.

Big picture, Marcus. Pull the threads together, because there's a lot moving at once.

One image captures it, Kate. This is the most concentrated model-release window in the industry's history. Gemini 3.5 Pro, Claude Fable and Mythos 5, a rumored GPT-5.6, Grok 5 — all landing in the same four weeks. And it's happening the same week the people who build these models sit before the G7. The capability curve and the governance curve are visibly racing each other. And the governance side showed up offering "voluntary commitments."

And under all of it, the money keeps climbing.

Vertically, Kate. Nvidia is reportedly preparing a large bond sale just to fund more compute. Hyperscaler capital spending this year is running about eighty-three percent above last year. Meta alone guided to between a hundred fifteen and a hundred thirty-five billion dollars in AI spending for 2026 — roughly double last year. The capability is being bought with a firehose of cash, and the question of whether that math ever closes is still wide open.

So if you had to name the open question listeners should carry out of today?

Whether Western, lab-led self-governance actually holds, Kate. Technology is outrunning the rulebook — that part isn't really in dispute anymore. The optimistic read is that voluntary, transparent, competitive self-governance beats a heavy state hand that would slow the West and embolden everyone else. The pessimistic read is that a photo op in Évian and a page of soft pledges is not a serious answer to models that can chain zero-day exploits or do real chemistry. My honest position sits between them. The frontier is largely Western and genuinely dazzling. What's still missing is the paperwork — the proof, the process, the binding commitment somebody actually signed. Win that, and the West keeps the lead it earned. Skip it, and someone else writes the rules.

And on a lighter note to close — Marcus, did AI kill the self-help book?

Tim Ferriss thinks it might have, Kate. He published a piece this week arguing that self-help and how-to nonfiction sales fell sharply in 2025 and look worse in 2026. His suspect is AI — why buy a how-to book when a chatbot answers your exact question, instantly, for free?

And let me guess, the internet disagreed.

Vigorously, Kate. Some said it's the cost of living and inflation, not chatbots — people just have less to spend. And my favorite take: that the self-help genre was a self-referential club of people selling each other's products, due for a reckoning with or without AI. So it's three suspects in the room — AI, the economy, and the genre quietly eating itself. The honest answer is probably all three. But it's a small window into something bigger: AI's reach now extends well past chips and data centers, into culture, publishing, and how people go looking for advice.

Which is the perfect note to end on — the technology you weren't watching, reshaping an industry you weren't thinking about.

That's the story under the story most days, Kate.

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