AI in 15 — July 19, 2026

A machine just closed a math problem that had sat open for thirty years. It did it in about ninety minutes. And the catch — the part almost nobody in the excited headlines mentioned — is that a human professor had been feeding that same problem to the machine's predecessors for a full year first.

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

And I'm Marcus, your co-host. And Kate, after a week of open-model fireworks, today we get a quieter but arguably deeper question — is AI starting to actually do original research?

That's our lead, Marcus — GPT-5.6 Sol and a thirty-year-old statistics conjecture. Then a run worth your time.

Google delays its next flagship model, and Alphabet sheds two hundred billion dollars in a single day.

An inference startup you may not have heard of just hit a seventeen-and-a-half-billion-dollar valuation.

DeepMind points its most powerful AI at the next pandemic.

And George Lucas compares AI skeptics to people defending the horse and buggy.

Lead story, Marcus. Walk me through this math result — because on the surface it sounds like the sci-fi moment everyone's been waiting for.

It does, and the underlying facts are genuinely impressive, Kate. Edgar Dobriban — he's a professor at Wharton — reports that OpenAI's GPT-5.6 Sol Pro resolved a roughly thirty-year-old open question about something called the Benjamini-Hochberg procedure. Now, that sounds obscure, but it's one of the workhorse tools of modern science. Any time researchers run hundreds of tests at once — think gene studies, drug trials — they need a way to control how many false positives sneak through. Benjamini-Hochberg is that method. Everyone assumed it stayed reliable even when the data points are correlated. Nobody had ever proven it.

And Sol Pro proved it.

In about ninety minutes, Kate. And for contrast — the previous model, GPT-5.5, had been thrown at the same problem and failed after roughly twenty hours running across multiple agents. So on its face, that's a real generational jump.

Okay. So why do I hear caution in your voice?

Because the framing matters enormously, Kate. Three things to hold onto. First — Dobriban is clear that the model combined existing techniques rather than inventing genuinely new mathematics. It was clever assembly, not invention from nothing. Second — the practical gap it found was tiny. We're talking a false-discovery rate of about zero-point-one-oh-four against a target of zero-point-one. Mathematically interesting, practically almost nothing.

And the third?

The third is the one that reframes the whole headline, Kate. A commenter on Hacker News pointed out that Dobriban had spent roughly a year feeding this exact problem to GPT-5.4 and 5.5 first — priming it, refining the setup, building up context. So that headline "ninety minutes" is really the last mile of a year of human-guided work. It's a relay race where a human ran the first twenty-six miles and the model sprinted the last hundred yards.

So is this a breakthrough or not?

It's a real marker — a Berkeley statistician called it a sign of advancing capability whose consequences reach far beyond math, and I think that's fair. But it's a marker of AI as a research accelerator, a tireless collaborator, not an autonomous mathematician. The honest version of the story is a human and a machine did this together. And that's still a big deal — it's just a different big deal than the headline sells.

Story two, Marcus, and this is the one that moved real money. Google was supposed to ship its next flagship. Instead, it slipped — and the market did not take it kindly.

It did not, Kate. Alphabet shares fell about four-point-four percent on Thursday — that erased roughly two hundred billion dollars in market value in a single day. The trigger was a Bloomberg report that Gemini 3.5 Pro, Google's next flagship model, is months behind schedule. Sundar Pichai had signaled it back at their developer conference in May, and it was expected in June. It's now in partner testing with no firm date.

What went wrong?

The reporting points to a late-June update to the training data that produced disappointing results — and notably, the weakness was in coding, Kate. Which is exactly the wrong place to be weak right now. The report describes frustrated engineers and managers genuinely worried Google is slipping.

And the timing is brutal, isn't it? Because coding is where everyone else is surging.

That's the whole sting, Kate. In just the last two weeks, OpenAI shipped GPT-5.6 Sol with big claims about agentic coding. Kimi K3 — the Chinese open model we've been tracking all week — debuted at number one on a human coding leaderboard. So Google stumbles on coding in the exact fortnight its rivals plant flags on coding. The contrast writes itself.

But play devil's advocate for me. Is a two-hundred-billion-dollar drop actually justified by a delay?

Great question, and I'd push back on the market here, Kate. This is a delay, not a failure. Gemini remains genuinely competitive on broad intelligence — it's a top-tier model family today. A company choosing not to ship something that missed its internal bar is arguably discipline, not weakness. A two-hundred-billion-dollar single-day swing on a slipped date tells you as much about how jittery this market has become as it does about Google's actual position. When the whole sector is priced for perfection, a missed month reads as a crisis.

So nerves as much as substance.

Exactly, Kate. Watch the eventual model, not the stock chart. If 3.5 Pro lands strong in a few months, this Thursday becomes a footnote.

Story three, Marcus, and this is the business-side mirror of everything we've covered this week. A company called Fireworks just raised at a seventeen-and-a-half-billion-dollar valuation. Most listeners have never heard of them. Why should they care?

Because Fireworks sits at the exact pressure point this whole week has been about, Kate. They closed roughly one-and-a-half billion dollars in new funding at that seventeen-and-a-half-billion valuation — Nvidia is one of the backers. They were founded by ex-Meta engineers back in 2022, and they run what's called inference infrastructure. That's the plumbing — the machinery that actually runs an AI model and serves its answers once the model's already been built.

So not building the models. Running them.

Running them, cheaply and reliably, at massive scale, Kate. And the numbers show why investors are excited — they say they crossed a billion dollars in annualized revenue, up about five-fold year on year, and they're now serving more than forty trillion tokens a day. A year ago that figure was fifteen trillion. And crucially, more and more of what they run is open-weight models.

And that's the connection to Kimi K3.

That's the whole thesis, Kate. Here's the logic. As frontier intelligence gets commoditized — as free open models like K3 close the gap on the expensive closed ones — the raw model stops being where the money is. The value slides to whoever can run those models fastest and cheapest. The picks and shovels. Fireworks is a pure bet that a wave of enterprise demand is migrating onto cheaper open models, and somebody has to serve them at industrial scale.

And Nvidia backing them is a little on-the-nose.

Beautifully so, Kate. Nvidia sells the shovels to the shovel company. Whether enterprises use closed models or open ones, whether it's OpenAI or Kimi, the tokens still run on Nvidia silicon. Backing Fireworks means Nvidia wins the inference boom from a second angle. It's a tidy little illustration of who collects a toll no matter which model wins.

Story four, Marcus. Let's go somewhere completely different — because this one isn't about chatbots or coding. DeepMind is aiming its AI at pandemics.

Right, and this is a genuinely serious effort, Kate. DeepMind, together with its sister company Isomorphic Labs, laid out what they're calling a bioresilience push. The idea is a three-layer defense against biological threats — using frontier AI for pathogen surveillance, so spotting a dangerous outbreak earlier, and then for dramatically faster design of countermeasures. Think vaccines and treatments developed in a fraction of the usual time.

And they're not doing it alone.

No — it's backed by more than fifteen partnerships, Kate, and the names are heavyweight. Lawrence Livermore National Laboratory, the UK's AI Security Institute, CEPI — the coalition behind pandemic vaccine work — and the Francis Crick Institute. So this is the AI-for-drug-discovery thread applied specifically to biosecurity, with real institutional muscle behind it.

But Marcus, there's an obvious tension here, right? The same AI that designs a vaccine faster could design a threat faster.

And to their credit, that dual-use problem is baked right into the framing, Kate. The stated goal is two-sided — prevent the models themselves from being misused to design something dangerous, and simultaneously arm the defenders so that if something does emerge, the response is faster than ever. That's the honest way to think about this technology. The same capability that helps a doctor helps a bad actor, and you can't have one without confronting the other. What I like here is they're not pretending the risk away — they're building the defensive side deliberately rather than hoping nobody weaponizes the offensive side.

So less hype, more infrastructure.

Exactly, Kate. This is AI doing something genuinely consequential and unglamorous. No leaderboard, no viral demo — just some of the most powerful models on earth pointed at the thing that could actually harm the most people. That's a story worth more attention than it'll get.

Let's land on something lighter, Marcus. George Lucas — yes, that George Lucas — weighed in on AI. And he was not gentle with the skeptics.

He was not, Kate. Lucas compared people rejecting AI to, quote, insisting the horse and buggy is where it's at. Which is a pointed line from a filmmaker whose entire career was built on inventing new technology to tell stories — the man basically rebuilt Hollywood's special-effects industry from scratch.

And it echoes another shift we mentioned yesterday.

It does, Kate — Linus Torvalds, the creator of Linux, telling the anti-AI wing of his project to essentially fork off, calling AI a useful tool. So within the same week you've got a legendary filmmaker and a legendary programmer, two people with zero incentive to shill, both landing in roughly the same place — this is a tool, use it or get left behind.

Do you buy it, or is it easy for the giants to say?

A bit of both, Kate. I'd note both men are established enough that disruption doesn't threaten them personally — it's easier to cheer the new tool when your legacy is secure. But there's a real signal in the pattern. The loudest voices calling AI pure hype a year ago are quietly softening. And when the skeptics start conceding the tool is useful, that tells you something the benchmarks don't.

One to watch tomorrow, Marcus.

It's still Kimi K3's weight drop on July twenty-seventh, Kate — but with a sharper edge now. We've spent all week saying the number-one ranking is a preview until you can download it. Here's what I'll actually be watching: not just whether Moonshot ships on time, but whether inference shops like Fireworks — the company we just discussed — stand it up fast and cheap. Because that's the real test.

Agree, or counter?

One refinement, Kate. The weights are only half the story. A two-point-eight-trillion-parameter model won't fit on a single machine — you need a rack-scale system to run it. So even a perfect launch leaves open the question that actually matters for most teams: not can you download it, but can you afford to run it. Open in license isn't the same as open in your budget. That gap is where the next month gets decided.

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