AI in 15 — July 11, 2026

Roughly four hundred former Apple employees now work at OpenAI. Apple says that's not a hiring streak — it's a heist. And today, one of the most secretive companies on Earth walked into federal court to prove it.

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

And I'm Marcus, your co-host. After a week of model launches, the biggest AI story today isn't a benchmark — it's a lawsuit. And it's a big one.

It really is, Marcus. Our lead: Apple sues OpenAI, alleging a coordinated trade-secret theft aimed straight at its hardware. Then a run worth your time.

An AI reportedly cracks a fifty-year-old math conjecture in under an hour.

OpenAI's number two steps down over illness, right as an IPO looms.

Humanoid robots perform their first live surgery — on pigs.

Europe waves through message-scanning that a majority of its own lawmakers voted against.

And Meta yanks an AI image feature after just days of backlash.

Lead story, Marcus. Apple filed suit yesterday against OpenAI. What's the actual allegation?

It's about as pointed as a corporate lawsuit gets, Kate. Apple filed in the Northern District of California accusing OpenAI of a systematic, leadership-directed effort to steal its hardware trade secrets — product designs, manufacturing processes, supply-chain strategy. The complaint names two former Apple people now at OpenAI. Tang Tan, who spent twenty-four years at Apple and was VP of product design for iPhone and Apple Watch — he's now OpenAI's Chief Hardware Officer. And Chang Liu, a senior electrical engineer who was there eight years.

And the four hundred employees figure — that's the framing?

Right, Apple says roughly four hundred former staff now work at OpenAI, and it frames that as a "pattern of theft" rather than normal poaching. But the vivid part is the specifics, Kate. Apple claims Tan used Apple's confidential internal project code names while recruiting, and asked job candidates to physically bring Apple hardware components to their interviews.

Wait — bring the actual hardware in? To an interview?

That's the allegation, Kate. And it goes further. Apple says Tan coached departing employees on how to evade Apple's security procedures — including telling recruits not to tell Apple they'd taken OpenAI jobs, so they could stay embedded and keep pulling information as long as possible. Liu, separately, allegedly never returned his Apple laptop and downloaded confidential technical documents. OpenAI's response was one dry sentence: "We have no interest in other companies' trade secrets."

So why does Apple care this much? OpenAI makes chatbots.

Because OpenAI is building hardware now, Kate. It bought Jony Ive's design firm for six and a half billion dollars last year to build a consumer AI device, and Ive's company is referenced in the filing — though Ive himself isn't a defendant. So this lands squarely on OpenAI's ambition to put a physical product in your pocket. And the comparison flying around Hacker News is Waymo versus Uber — the trade-secret suit that basically killed Uber's self-driving program.

That's a heavy comparison. Does it hold?

It's worth tempering, Kate — these are allegations, and OpenAI hasn't answered them in court yet. But two things make this more dangerous than a typical talent spat. One, Apple has effectively infinite legal resources and a reputation for grinding cases out for years. Two, code names and physical components are unusually concrete for a trade-secret claim — if Apple can actually document them, they're hard to wave away. And there's a quieter angle enterprises flagged: if you're a company handing OpenAI your data and IP, a lawsuit alleging it built a machine for extracting other people's secrets is not a great look.

So on top of everything else this week, OpenAI is now fighting Apple.

And that's the frame, Kate. Pair this with the New York Times sanctions fight we covered yesterday — the allegation OpenAI hid evidence and deleted outputs — and you've got a genuinely rough legal week. Both stories feed the same question: can this company be trusted with what it's given? That's a different kind of threat than a rival's cheaper model.

Story two, Marcus, and this is the one that might actually matter in fifty years. An AI reportedly proved a fifty-year-old math conjecture.

If it holds up, yes, Kate. OpenAI says GPT-5.6 Sol Ultra — running sixty-four subagents in just under an hour — produced a proof of the Cycle Double Cover Conjecture. That's a graph-theory problem posed independently by George Szekeres in 1973 and Paul Seymour in 1979, and it's resisted mathematicians for about half a century. In plain terms, it says for any bridgeless graph you can find a set of cycles that covers every edge exactly twice. OpenAI published both the prompt it used and the full proof.

Okay, but I have to stop you. "AI solves famous math problem" — how many times have we heard that and it fell apart?

Which is exactly the right instinct, Kate, and I want to be clear: this is a claim, not a settled fact. The mathematicians on Hacker News noticed the proof is strikingly short — and a short proof of a fifty-year problem cuts two ways. Either it's a clever trick every expert somehow missed, or there's a subtle flaw hiding in the brevity. Validating it will take the community weeks, not hours. Nobody should call this confirmed today.

But if it is real — how is this different from an AI winning at chess or writing code?

Because it's original mathematics, Kate — new knowledge with genuine intellectual prestige attached, not a game with known rules. And it fits a clean pattern of what these systems can now automate: tasks where correctness is checkable, the answer can be written as text, and there's prior work to build on. That's a real category. One more honest detail — commenters noticed how much of the prompt is spent basically begging the model to actually solve the problem instead of producing "vague optimism." So even at the frontier, these things still need heavy scaffolding to stay honest on the hard stuff.

So — impressive if true, and we'll know in a few weeks.

That's the whole story in one line, Kate. Report it with the asterisk firmly attached.

Story three, Marcus. A leadership shake-up at OpenAI. Fidji Simo is stepping down.

She is, Kate — and it's for a serious personal reason. Simo joined in May 2025 as CEO of Applications, reporting directly to Sam Altman — effectively the company's number two on the product and business side. She's stepping back to a part-time advisory role after a severe relapse of POTS, a neuroimmune condition. Her duties get split three ways: among Greg Brockman, CFO Sarah Friar, and Chief Strategy Officer Jason Kwon.

And the timing is the newsworthy part.

It is, Kate. She'd been seen as a candidate for expanded responsibility, and this comes as OpenAI is reportedly prepping a possible IPO. Losing your number two on the applications side — the part of the company that turns models into revenue — right before you go public is not nothing. It's a human story first, but it also thins the bench at a delicate moment. Splitting one person's job across three executives is rarely as clean in practice as it looks on the org chart.

Story four, Marcus, and this one's wild. Humanoid robots did their first live surgery.

On live pigs, Kate, and the detail that matters is the price tag. In a UC San Diego trial, surgeon-controlled Unitree G1 humanoids — these are around thirteen to twenty thousand dollars each, about seventy pounds — removed gallbladders from live pigs through standard laparoscopic ports. One operation used a humanoid with a human assistant; another used two humanoids working together. Published in Nature.

And why does the cost matter so much?

Because the incumbent, Intuitive's da Vinci surgical system, often runs half a million to several million dollars, Kate. These are teleoperated — a human surgeon is still driving — but a cheap, general-purpose humanoid hints at remote surgical care reaching smaller hospitals that could never afford a da Vinci. It's a striking crossover of robotics, teleoperation, and AI. And worth noting, geopolitically: it's a Chinese-made robot being used in a US lab to push a US research result. The supply chains here are tangled in ways that'll matter later.

Still early, though.

Very early, Kate — pigs, not people, and a surgeon in full control. But cheap hardware doing precision work is exactly how a capability goes from a lab curiosity to something in a rural clinic a decade out. Worth watching the cost curve, not just the demo.

Story five, Marcus, and this one bothers me. Europe passed message-scanning that most of its own lawmakers opposed.

The procedure here is genuinely strange, Kate. On July ninth, the European Parliament had a motion to reject so-called Chat Control — suspicionless scanning of private messages. Three hundred and fourteen MEPs voted to kill the scanning. Two hundred and seventy-six voted to keep it. So more members opposed it than supported it.

So it should have died. What happened?

It needed an absolute majority to reject — three hundred and sixty-one votes — and it fell short of that bar, Kate. So even though a clear plurality voted against, the scanning survives on a technicality. It extends warrantless scanning across Instagram, Discord, Snapchat, Gmail, and iCloud until 2028. Former MEP Patrick Breyer called it "a farce that damages democracy."

And the AI angle — because this is our beat.

This is fundamentally an AI deployment, Kate. Continent-scale message scanning means a wall of classifiers making automated guesses about the private chats of roughly half a billion people. And every classifier at that scale generates false positives — innocent messages flagged, at volume. So you're building the largest automated surveillance system in the democratic world and hoping the models are accurate enough not to ruin innocent lives. The permanent version of these rules gets negotiated in September, so this isn't over.

That's the part I'd keep an eye on.

Agreed, Kate. The technicality got the headlines. The September negotiation is where it actually gets decided.

Last hit, Marcus, and it's a fast one. Meta pulled an AI image feature just days after launching it.

Right, this ties to the Muse Image rollout we covered earlier in the week, Kate — the one that could pull public Instagram photos into AI generations when an account got tagged. After days of user anger over consent and data use, Meta withdrew the feature. One critic summed up the mood: "AI companies see people's images and data as raw material to be exploited."

So the backlash actually worked.

This time, yes, Kate — and that's the signal worth taking. Consumer patience for opt-out-by-default AI image tooling is thin, and it snapped in days, not months. It pairs with broader unease about Meta putting facial recognition into its Ray-Ban glasses. Meta's whole edge is that it already knows who you are and what you look like — which is exactly why people flinch when it turns that into a feature. Distribution is a moat right up until it becomes a liability.

One to watch tomorrow, Marcus.

The Apple versus OpenAI docket, Kate. Whether OpenAI files a substantive response or a motion to dismiss tells us if this is a quick settlement or a multi-year fight hanging over its hardware push.

Agree, or counter?

Small counter, Kate. The lawsuit's the drama, but the story that actually changes what we think AI can do is that math proof — if it survives peer scrutiny. Watch the mathematicians, not just the lawyers.

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