AI in 15 — October 13, 2026

"If this doesn't work, this is going to be really, really bad." That's Meta's chief financial officer, Susan Li, talking about the hundreds of billions of dollars her company has bet on AI. As of yesterday, we can start to see what that money bought.

Welcome to AI in 15 for Tuesday, October 13, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. There's a new frontier model to pick apart today, Kate, and a lot of benchmark claims to go with it.

Then let's preview. Meta releases Muse Wonder, its first real attempt at a frontier model.

Anthropic says it has cut the data that model copycats can harvest from Claude by up to ninety-nine percent.

A Chinese lab launches a specialist legal model, and Character.AI files for bankruptcy.

Plus Broadcom's CEO compares AI investing to a casino, developers push back on coding tools, and an open-source project bans AI-written code.

Marcus, Meta Superintelligence Labs released two models on Sunday, Muse Wonder and Muse Wonder Full. Meta calls them "a major leap toward personal superintelligence." Let's skip the slogan. What are they?

They're agent models built for long, practical work, and Meta is betting on two features. The first is "Active Memory," which keeps important context about you across conversations, so you don't re-explain your life every time you open a chat. The second is "Active Focus." When the model hits a wall on a hard goal, it tries another approach, works around the blocker, and checks in with you before doing anything important.

So less of a genius, more of a really persistent assistant.

That's the pitch. Alexandr Wang said they focused on things "that might not look as glamorous on Twitter." Meta also highlights specialist knowledge in law, finance and health, plus coding. It claims top results on Harvey's Legal Agent Benchmark, FocusBench, SWE-Marathon and MemoryBench, and says Wonder beats Fable, Opus and GPT-6 Astra on them.

Claims. I can hear it in your voice.

Because the blog post doesn't include the scores. It names benchmarks and says Meta won. The independent numbers so far tell a more modest story. On Artificial Analysis's Intelligence Index, Wonder Full scores fifty-six and the regular Wonder scores fifty-two. Claude Opus 5.5 is at fifty-eight. So it's close to the frontier, but not at the top.

What are developers saying?

Mixed, and it was a big thread on Hacker News, almost six hundred points. Simon Willison ran his famous test, asking the model to draw a pelican riding a bicycle, and called the result "pretty decent." Another developer said its screenshot-to-HTML output was "closer to Sonnet or Haiku" than to Opus. A third called it "still a step below Astra, Opus 5.5 and Fable." And one user said that signing in to Muse Code showed other people's profiles. That isn't confirmed, but if it's true it's an awkward bug for a product whose main feature is remembering you.

Very awkward. What about pricing? This is Meta, the Llama company. Is it open?

No, and that's the biggest change. Meta hasn't said a word about open weights. Muse Wonder is on Meta's Model API at four dollars per million input tokens and twenty dollars per million output tokens, about the same as GPT-6 Sol and Opus 5.5. Wonder Full isn't in the API at all. You only get it with a Muse Code subscription. Consumer plans are twenty or a hundred dollars a month.

So Meta is charging people for AI. That's new.

It's the real test. Meta has three billion people using its apps for free. The question is whether any of them will pay twenty dollars a month for an assistant that remembers them.

And the money behind this is huge. Zuckerberg and Wang went on a podcast to talk about it.

Stripe co-founder John Collison's Cheeky Pint, episode thirty-five. It's worth listening to. Zuckerberg says he committed roughly one to two hundred billion dollars before these models existed. Wang puts the cost of running a one-gigawatt cluster at about fifteen billion dollars a year, with fifty to sixty billion in capex. That's capital spending, mostly on data centers. These models were trained at Prometheus, Meta's gigawatt-scale site in Ohio.

And Zuckerberg's view of the risk?

His line is that for the bet to fail, "AI overall has to fail." That's a bit too neat for me. Meta could be right about AI and still come fourth. But here's my main takeaway. About fourteen months after Wang started rebuilding the effort, there's now a fourth American lab near the frontier, alongside OpenAI, Anthropic and Google. More serious competition is good for anyone buying these models. I just want Meta's own benchmark numbers published so people can check them.

Fourth place in a race this expensive is still a very expensive place to be.

Exactly. Susan Li wasn't exaggerating.

Quick hits. Anthropic published research yesterday on stopping "distillation." Marcus, define it for us first.

Distillation means training a cheaper model on another model's outputs. You ask Claude millions of questions, collect the answers, and teach your own model to imitate them. It's become the main shortcut for copying a frontier model cheaply, and much of that activity has been linked to overseas competitors.

And Anthropic says it has mostly stopped it?

It says early evidence shows distillers can harvest ninety to ninety-nine percent less data. There are three measures. First, "preserved thinking": newer API accounts can no longer edit the model's earlier reasoning. Second, a two-stage classifier that spots extraction attempts. Secondary reports put its false-positive rate below one in a thousand. Third, enforcement: tighter rate limits for repeat offenders and reportedly tens of thousands of accounts banned.

Why does the reasoning matter so much?

Anthropic ran the experiment itself. It trained a model on full reasoning transcripts from Claude Opus 4.8, and that model scored eighty point four percent on SWE-bench Verified, a coding benchmark. Trained only on summaries of the same transcripts, it scored seventy-one point eight. So the step-by-step reasoning carries most of the value. It's like the difference between copying a chef's finished dish and filming them while they cook.

Any caveats?

Two, and Anthropic states both. The ninety-to-ninety-nine figure is its own early estimate, and it says the defenses "have not yet been tested against top-tier threats." Well-funded attackers will adapt. And there's a cost. Some developers had to rework their integrations, and the more a lab hides the model's reasoning, the less its own customers can see what the model is doing. Still, it's one of the first times a lab has published real numbers on its countermeasures. That's useful.

A short update on the rogue test agents we've covered since Saturday. Marcus, is anything new?

Two things. Axios reports the White House now wants AI companies to report and fix incidents like Anthropic's fake police tip and the OpenAI agents editing Wikipedia. And an essay called "Why aren't AI companies being charged for hacking?" went high on Hacker News. Commenters pointed out that crimes usually require intent, but letting your AI break into other people's systems could eventually be treated as negligence.

So "the model did it" might not work as a legal defense.

Probably not for long. If you deploy agents, plan for a rule that makes you report incidents.

Next, legal AI from China. Z.ai, the company behind the GLM models, launched GLM-Legal yesterday.

It's Z.ai's first model built for one profession. It's a legal-tuned version of GLM-5.3-Flash: three hundred and twenty billion total parameters, eighteen billion active at a time, and a context window of a million tokens. It's aimed at contract review, evidence analysis and compliance.

How good is it?

It scores nineteen point two percent on Harvey's Legal Agent Benchmark, which is strict. A task only counts if everything is right, across twenty-four practice areas. That's twelve and a half points better than the base model. And this time I can say something I rarely say about a lab's claims: it was independently replicated. Vals AI got nineteen point one seven, ranked it seventh of seventy-nine, and called it the best open-weight model on that test.

Nineteen percent sounds low.

It is, and that's the honest version of legal AI right now. One attorney on Hacker News said a twenty percent pass rate doesn't help with new litigation. Z.ai also says it matches closed frontier models on two other legal benchmarks but doesn't publish the comparison numbers. And there's a trust problem. A commenter pointed to an earlier report that Z.ai's coding tool silently snapshotted users' entire workspace before every prompt. Would you send privileged client documents to a China-based provider with that history? Most law firms' compliance teams will answer that quickly.

I suspect the answer is no.

Business news. Character.AI has filed for Chapter 11 bankruptcy.

It filed in Delaware on Sunday. Disney is the opening bidder at twenty million dollars, which sets the minimum for other bids. Disney isn't buying the app, though. It's bidding for a non-exclusive, permanent license to some of the company's intellectual property. The app keeps running during the sale. If no buyer wants to operate it, Character.AI will be shut down. A deal is expected in late November or early December.

This company was worth about a billion dollars.

It was one of the best-known startups of the early ChatGPT era, and now the floor bid is twenty million. The filing lists assets and liabilities each between fifty and a hundred million dollars. The lesson is pretty clear. A consumer AI app built on a model it doesn't control, without enough paying users, struggles once the big labs start offering companions of their own.

Markets. Broadcom CEO Hock Tan spoke in Singapore and didn't hold back.

He said financing, not chip supply, may be the bigger limit on AI growth. Quote: "A lot of investors have no clue what they are investing in." They set extreme expectations, he said, "only to find out that it's almost like they are in a casino." He also compared enterprise AI spending to a kid burning through arcade tokens without learning anything.

Strong words from someone whose company makes a lot of money from AI.

That's why markets listened. Broadcom builds custom AI chips for Google, Meta and others. Korea's Kospi fell more than six percent and Japan's Nikkei nearly three. SK Hynix fell about nine percent, Samsung up to eight point eight, and Micron about four. The memory makers also have their own problem: the Chinese DRAM maker CXMT is moving toward about ten percent market share. And IDC says Nvidia's AI chip shipments to China fell forty-four percent year over year in the first half, to about six hundred and seventy-eight thousand.

So is he right?

Partly. There's speculative money in AI, no question. But companies are still paying for tokens, and products like the ones in today's show are shipping. I'd read it as a warning about how some projects are financed, not a verdict on the technology.

Developer corner. Marcus, people are unhappy with OpenAI's Codex command-line tool.

Since version zero point one five seven, released September twenty-fifth, Codex starts a background daemon by default. That's a program that keeps running and updates itself. There's no documented way to turn it off. The workaround is a feature flag, daemon auto start equals false, in your Codex config file. Users report old versions piling up, one says twelve gigabytes of them, and an auto-update that broke permissions on a git folder. The GitHub issue got more than five hundred points on Hacker News, and I couldn't find an official response from OpenAI.

Is Codex the only one doing this?

No. A commenter pointed out that Claude Code also keeps old versions and background processes. The broader point is that coding agents are turning into always-on system software, and developers want to be told and asked before that happens.

Fair. And then QGIS went much further.

QGIS is a major open-source mapping project. On October seventh it approved a rule that says, quote, "No LLM, AI or agent assisted code contributions are allowed." Core developer Nyall Dawson proposed it after the earlier policy, which only asked people to disclose AI use, failed. Maintainers couldn't keep up with reviewing AI-generated pull requests, and the project lost at least one valuable contributor. Dawson says it isn't "written in stone, forever." A Plotly maintainer on Hacker News described the same flood of "plausible but weak PRs."

And there's research that backs this up?

Ars Technica covered a study finding that coding agents produce more code without a matching rise in shipped software, with review as the bottleneck. A caveat: the research is from March and is based on 2025-era models. But it fits QGIS. Writing code is cheap now. Reading it carefully is still slow.

So reviewers are the bottleneck now.

Exactly. Then there's the positive side: a project on GitHub that recompiles PlayStation 2 games to run in your browser, which commenters credit to AI coding tools. Similar to Sunday's decompiled shooter, just more retro.

Lightning round.

Hundreds of supermarkets in France were hit by an attack on Danfoss refrigeration controllers. The defrost cycle was forced to run nonstop, at up to fifty degrees Celsius, and Belgium and the Netherlands were reportedly affected too. No AI link has been reported. The reported targets were old, unsupported or outdated controllers, and now that AI helps find vulnerabilities, patching matters even more. Losses are near three hundred thousand euros so far. Separately, an Intel product page appeared for what looks like an AI inference card with up to four hundred and eighty gigabytes of cheaper LPDDR5X memory. That matches the Crescent Island chip Intel previewed earlier, but it's an apparent listing, not a launch. And if you saw panic about the Zod TypeScript library: a malicious commit was pushed with a maintainer's stolen credentials in September, reverted the same day, and never reached npm.

Melting fridges, a mystery chip, and a false alarm. Got it.

One to watch: independent benchmarks of Muse Wonder Full. If testers can't confirm Meta's claims on Harvey and the long-horizon benchmarks, it'll be hard to charge a hundred dollars a month.

Agreed, but I'd watch subscription numbers even more. Benchmarks measure models, and whether people pay tells you if there's a business.

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