AI in 15 — October 12, 2026

Five thousand. That's how many people on Earth, by Andrej Karpathy's count, have seen what frontier AI can really do. And he says even they are starting to fall behind it.

Welcome to AI in 15 for Monday, October 12, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Fresh week, Kate, and some very big claims to pick apart.

Then let's get into it. Karpathy says the gap in understanding what AI can do has become a funnel.

Terence Tao asks what mathematics is for once problem-solving stops being the hard part.

Sakana AI releases a model that coaches and grades itself.

Plus Nicolas Cage takes on Amazon, AI-built clones of Word and Photoshop, Google gives Gemini agents their own work accounts, and the rising price of art made by humans.

Marcus, Karpathy posted this on Saturday. He went back to something he wrote in April and said the gap is "widening." What's the argument?

He says it's no longer two groups talking past each other. It's "a sharp funnel," and he gives rough numbers, which he calls napkin math himself. About six billion people, roughly three quarters of humanity, have "barely come in contact with LLMs at all." Then one to two billion are casual users of free chatbots. Then about twenty million people, mostly in code and math, have watched an agent finish something in an afternoon that used to take weeks or months.

And at the narrow end?

About five thousand people with access to frontier systems inside the labs. For that group, he says, "human review and comprehension are starting to fall behind."

So if you're listening to this show, you're probably in the one-to-two-billion group. Maybe the twenty million.

Most likely. And his point is that what you see in a free chatbot is a long way behind what the labs see internally. He gives three reasons. The first is ambition: "The impact scales with ambition, problem size, and horizon." Ask for a limerick and you get a limerick. Hand it a month-long project and you see something very different.

And the second?

Jaggedness. Capability peaks in work that's "digital, verifiable and economically valuable." That comes from how these models are trained. Reinforcement learning works best when you can automatically check the answer, so code and math race ahead while messier work lags. The third is access. Free tier, paid tier and internal-only models are basically different products that share a name.

How did people take it?

It was the top AI story on Hacker News, more than five hundred comments, and plenty of them pushed back. Some doubted that even the twenty million agree agents can finish big projects from a prompt. They pointed to maintenance, token bills, and the fact that someone still has to understand the codebase. Others said the funnel mostly shows how narrow the strong slice really is.

Where do you come down?

The shape seems right to me. The numbers are guesses, and the five thousand can't be checked. But look at last week. OpenAI's internal model produced hundreds of math results faster than mathematicians can read them. Anthropic's agents did things on the open web that nobody noticed for ten weeks. Both of those are what "comprehension falling behind" looks like in practice. My caveat is that Karpathy himself says the strength is uneven. Being in the five thousand doesn't mean the model is good at everything. It means you've seen where it peaks.

So the free chatbot is the trailer, not the movie.

And some of the movie is still out of focus.

Quick hits. We've covered OpenAI's mass release of math proofs all week. Marcus, what's new is the reaction from mathematicians themselves, starting with Terence Tao.

Tao gave a talk at Caltech called "Math 2.0," and the slides were number two on Hacker News this weekend. Quick recap: OpenAI put three hundred and seventy-two results on GitHub, about three hours of compute each on average, with Lean proofs for many of them. Tao's question is what progress in mathematics means if solving problems is no longer the bottleneck.

And his answer?

The line everyone is quoting is "open problems are lighthouses, not destinations." Famous problems matter mostly because of the theory people build while chasing them. If a machine just lands on the island, you get the result but not the map.

I like that. Who else is weighing in?

This part is new for us. Twenty-five Fields Medal winners, the Nobel of mathematics, signed an open letter called "A Severe Misalignment of AI in Mathematics." Their argument is that solving problems was always a stand-in for building understanding. Timothy Gowers warned that the literature could balloon while no human community actually understands it.

That's a different complaint from last week's "nobody asked for this."

Much more serious. And Tao's slides take it out of math. He asks whether you would take an AI-designed cancer cure if no human understood how it worked. Drug-discovery researchers on Hacker News had a good answer: medicine already uses plenty of drugs whose mechanism isn't fully understood. Aspirin was used for decades before anyone understood how it worked.

And the people at the bottom of the ladder? I've read that some mathematicians are rewriting grant applications they just submitted.

That's the human cost. If your three-year project was one of those problems, you need a new plan. Here's what I'd ask OpenAI: you published average compute, but not the prompts and not the cost per problem. If you want the field to adjust, give it the information it needs to check you.

Next, Sakana AI in Tokyo, working with co-authors from UC Berkeley, released something called MASS, Multi-Agent Self-Supervision. Marcus, self-improving AI?

A careful version of it. One model plays three roles: it does the work, it redesigns the workflow, and it judges the results. In the inner loop, it runs a team of agents and compares the output with the current best workflow, which they call the champion, then tweaks the workflow. The weights don't change in that step.

And the outer loop?

It runs the champion workflow eighteen times per task, keeps up to sixteen of the best runs, and retrains itself on those transcripts using LoRA. That's a lightweight fine-tuning method, a bit like adding a thin layer of new habits rather than rebuilding the whole brain. The base model is Alibaba's open-weight Qwen3.6, twenty-seven billion parameters. GPT-5.5 and Claude Opus 4.8 were used only as outside judges.

Did it work?

Partly, and they're honest about where it didn't. Against the original model, the win rate went from about fifty-four percent after round one to seventy percent after round two. On research-style benchmarks it got up to one point six times more score per output token. On coding, Terminal-Bench and SWE-bench, nothing. Slightly worse, actually.

So what's the weak link?

The optimizer role, the part that redesigns the workflow. Swapping in GPT-5.5 as the optimizer helped far more than upgrading the judge. In other words, it's better at grading its own homework than at figuring out how to study. Hacker News said the released code doesn't match the paper, and that "recursive self-improvement" oversells what's really better teamwork among agents. My summary: two rounds of gains on synthetic tasks, open code, clear limits. Interesting, but not takeoff.

Over to Hollywood. At New York Comic Con on Saturday, Nicolas Cage said he refused to sign an AI waiver for Amazon's Spider-Noir.

And he didn't hold back. Quote: "Amazon is a very AI-friendly company, to the tune of fifty billion dollars invested in OpenAI. I am not an AI-friendly actor, so it's like, connect those dots!" He added, "I'm probably not going to be working with that studio again, but that's okay," and announced he's joining HBO's True Detective.

Very Nicolas Cage. What was actually in the waiver?

That's the gap in the story. Amazon hasn't commented, and nobody has said what the waiver covered. It could be digital de-aging, voice replicas, or training rights on his performance, and those are very different things. What's clear is that AI terms are now a negotiating point for individual stars, not just for unions. If a star of his size can walk away and land a job at HBO, studios will have to compete on those terms.

So "AI-free" becomes a perk you can bargain for.

Which leads nicely into our art story later.

Now something going viral on GitHub: AI-built clones of famous software. Marcus, someone rebuilt Microsoft Word?

A Rust "reimplementation" of Word from an account called storytold got a lot of attention on Hacker News, both praise and plenty of bug reports. Then there's "Doory," an open-source Canva clone apparently written entirely by Claude and pushed in a single commit. Phoronix covered AI-built Rust clones of Photoshop, Premiere and Lightroom. We mentioned Theo Browne's TypeScript compiler port last week. This is the same idea pointed at consumer apps.

Are these actually usable?

Mixed. A demo that opens a document is not thirty years of edge cases in Word. But the trend is real. The cost of a decent first version of a big application has dropped a lot.

And the legal question?

That's the interesting part. "Clean-room" traditionally means the engineers who write the copy never saw the original code. It's how Compaq legally cloned the IBM PC BIOS in the eighties. But if the engineer is a model trained on a big chunk of public code, can you call it clean? Nobody has tested that in court yet. When someone clones a product with deep pockets and a legal team, we'll find out.

Enterprise news. Google is giving Gemini agents their own work accounts.

Each agent gets its own email address, calendar, Drive storage and company directory entry. Coworkers can delegate to it like a colleague: send it a file, add it to a meeting, assign it a task. It's in private preview, and there's no extra charge for Gemini Enterprise customers.

So my next coworker might literally be a bot with a calendar.

And here's why I think it's actually smart. Remember Nadella's point yesterday about treating agents as insiders, and the bank statement posted to the wrong Slack channel. If an agent has its own identity, you can give it its own permissions, audit its actions and switch it off without touching anyone's personal account. That's much safer than agents borrowing a human's login. The risk is the flip side: an agent in the company directory can also receive phishing emails.

Good point. Even the bots need security training.

Quick money round.

Butterfly Effect, the parent of the agent startup Manus, raised more than five hundred million dollars, led by Boyu and IDG, with Tencent participating. It's their first round since Meta's acquisition fell through. Jeff Holden, who used to run product at Uber, brought Atomic Machines out of stealth with two hundred and fifty million dollars. It builds tiny machines from code, starting with cooling and power parts for AI data centers.

So someone's building the shovels for the gold rush.

Tiny shovels. And on open models, a reminder: Mistral's trillion-parameter Large 4 is supposed to release its weights on October twenty-seventh. The only independent score so far, from Artificial Analysis, puts the preview at thirty-eight. China's GLM-5.3 scores forty-five. So the Western comeback in open models still has ground to make up.

Let's finish with art. The BBC reports a "human premium." Buyers are paying more for art that's verifiably made by people.

And "AI-free" badges are spreading. It's a familiar economic pattern. When machines make something cheap and plentiful, the handmade version becomes a luxury. Think hand-thrown pottery after factories, or vinyl after streaming. The hard part is verification. A badge only means something if somebody checks it, and that market barely exists yet.

So the next startup boom is certifying humans.

Probably using AI to do the certifying.

One to watch: Apple's event tomorrow. It's billed as a smart-home event, but the real question is how much of it is about AI and Siri.

Agreed. If Siri doesn't get a lot smarter, Apple's smart home is just a nicer light switch.

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