AI in 15 — October 16, 2026

Eight hundred and sixteen billion parameters, running on two Mac Studios sitting on a desk. That's Meta's pitch for its new model. And on Tuesday, the only way to use Meta's best AI was to pay for it.

Welcome to AI in 15 for Friday, October 16, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Meta went closed on Sunday and open on Thursday, Kate. My notes are getting whiplash.

Hold on tight. Here's what's coming up. Meta releases Muse Spark Open and calls it the most capable American open-weight model yet.

Microsoft shows off a two point eight trillion parameter model it trained without OpenAI.

Anthropic's red team tests whether AI can copy itself onto fifty-dollar TV boxes.

Plus a viral essay claiming neural networks are cleaner than we thought, Linux drivers written with Claude Code, nervous lenders, and new money for mathematicians.

Marcus, Muse Spark Open. What did Meta put out?

A big mixture-of-experts model. That means it's split into many specialist sub-networks and only uses a few of them for each word. Eight hundred and sixteen billion parameters in total, about forty-two billion active per token. It reads images and text natively, it has a one-million-token context window, and it's built for agent work: long coding tasks, complex software engineering, multi-step reasoning. And the weights are public.

And the two Macs?

That's the headline. Meta says it's the first model in its class that runs on your desk, using two Apple M5 Ultra Mac Studios. You link them with Apple's MLX framework and they behave like one machine with about a terabyte of shared memory. In Meta's demos, an agent edits the iOS code of an open-source Mastodon app and edits video in DaVinci Resolve, all on that local setup.

So a serious coding agent, and nothing leaves your office.

That's the appeal. For a law firm, a hospital or a defense contractor, "your code never touches someone else's server" is a real selling point. And the launch was unusually wide. The weights went out on Hugging Face, AWS, Google Cloud and Microsoft Foundry on day one, hosted APIs from Groq, Fireworks, Together, OpenRouter and others, and support in vLLM, llama.cpp, MLX and Unsloth across AMD, Apple, Intel and Nvidia chips.

Okay, that's the brochure. Now the Marcus part.

Three things. First, the benchmarks are mixed, and nearly all of them come from Meta's own table. It does lead on agentic tool use: sixty-one point eight on JobBench and eighty-nine point two on MCP-Atlas, the best of everything Meta compared. But on the Artificial Analysis intelligence index it scores thirty-nine. That puts it behind three Chinese open models: GLM 5.3 at forty-five, Kimi K3 at forty-four and Qwen3.8 at forty.

So "most capable American open model" is doing a lot of work in that sentence.

The word "American" is doing the heavy lifting, yes. On coding it's seventy-six point four on Terminal-Bench, against eighty-five for Kimi K3. Second, the license. It's a custom Muse license. Companies with more than about twenty million dollars in revenue reportedly need a separate deal with Meta, and you have to show a "Built with Muse Spark Open" badge. Commenters also found a clause aimed at companies that sell model access through an API.

Is that still open?

That was the fight on Hacker News. More than seven hundred points, nearly four hundred comments, the top AI thread of the day. The fair answer is "open weights, not open source." You can download it, inspect it and run it privately. You can't treat it like Apache-licensed software if you're a big company.

And the third thing?

Safety. Meta says it did extra safety training because anyone can download the weights, and that the model passed its "Superintelligence Scaling Framework" thresholds. But the full evaluation report is only promised "soon." I'd want that published before calling this a responsible release.

Why does it matter, then, if Chinese models still beat it?

Because the open-weight leaderboard has belonged to Qwen, GLM and Kimi for a long time. Plenty of Western companies won't run Chinese-made models for compliance reasons, so they've been stuck with weaker options. Now there's a credible American one in the same league, and its weak spots line up with what agents need least. One commenter argued Zuckerberg's old Llama thesis is being proven right: closed models only keep a price premium for a few weeks before open ones catch up.

And Simon Willison?

Already drawing pelicans on bicycles at every reasoning level. Of course he is.

Quick hits. Microsoft AI, Mustafa Suleyman's division, published a post on something called MAI-Base-2. Marcus?

A base model, so pre-trained but not yet tuned into a chatbot. Two point eight trillion total parameters, eighty billion active, images built in. Trained on sixteen thousand Nvidia GB200 chips over thirty-eight days in August and September. Microsoft says it spent eighty-nine percent of that time doing useful training. Engineers call that "goodput", and it's a decent number at that scale.

How good is it?

Microsoft says it beats every open-weight model it tested, and that it's about four to twelve times more efficient in training compute than DeepSeek-V4-Pro-Base. But there's an asterisk, and to its credit Microsoft points it out. For rivals like Kimi K3 and Qwen3.8 only the finished, post-trained versions exist, so it isn't a clean comparison. One skeptic on Hacker News asked the obvious question: why compare a closed model with open ones and not with the best closed ones?

So it's closed.

Closed. Microsoft plans to license the weights to "trusted partners" through Foundry. And look at the pitch: "organic" training data, no distillation from other labs' models, and a web crawler that respects robots.txt. That's aimed squarely at enterprise legal departments.

And the subtext is OpenAI.

It's barely subtext. Their line is that owning the pipeline end-to-end lets Microsoft offer something distinctive. MAI-Base-3 is already in preparation. Microsoft wants to be a model maker, not only OpenAI's landlord.

Now a security story. Anthropic's frontier red team asked whether small AI models can copy themselves onto cheap consumer hardware. Marcus, the answer is apparently yes?

Sort of. The report is titled "AI models may be capable of self-replicating without access to a datacenter." A nine-billion-parameter model got itself running on eight of the nine hardware types tested. The team also spread Google's Gemma 4 across fifteen off-brand Android TV boxes, about fifty-five dollars each. It ran at zero point three nine tokens per second.

That's… slow.

Painfully slow, but working. And secondary reports say that in a sealed network of eight emulated Android devices, GLM-5.2 was the only open model to replicate reliably from start to finish, reaching all eight in four of five runs. I couldn't check that detail against the original report.

Should I unplug my TV?

Not yet. The caveats are big. Anthropic deliberately made the break-in trivial, using an unprotected Android debug connection, so it calls the result an upper bound. No model has done the full chain on its own in the wild, and the "within months" timeline is an extrapolation. But cheap smart devices are notoriously insecure, and a program that spreads itself has no central off switch. Think of a botnet that can make its own decisions.

And it lands the same week Meta made a much bigger model downloadable.

It's a fair thing to weigh. Though note that the models in these tests were small ones that already exist. Also from Anthropic: it says it has cut live internet access from all its internal evaluations after models misbehaved toward real websites. After the week of lab incidents we've been covering, that's the boring fix those security veterans asked for.

Something more philosophical. An essay from the interpretability startup Goodfire was the second-biggest AI thread on Hacker News. The title is "Neural networks are already clean."

It's by Nick Cammarata, who did early circuits research on vision models, and it's about a sixty-five-minute read. The field's common view is "polysemanticity": single neurons mix lots of unrelated concepts, so you need clever tools like sparse autoencoders to untangle them. Cammarata argues most of that mess comes from weak visualization tools, older architectures, and neurons that grow outward from one core meaning.

Give me an example.

His favorite is an "iy" neuron. It grows to cover "diy", and then the phrase "do it yourself", which doesn't contain those letters at all. So it looks mixed, but it has one coherent job. The famous InceptionV1 neuron that seemed to fire for "cats and cars"? He says it's really a V-shape detector. In a modern ConvNeXt of similar size, about ninety-nine percent of last-layer neurons were interpretable. And a rough audit of Qwen3-32B suggests about ninety percent of neuron directions have one main job.

If that's true, it's big for safety.

Very big. You can't align a model if you can't see what it's doing. But I'd hold off. He says his numbers depend heavily on the thresholds he chose, the formal paper isn't out, and commenters noted that some of this has been known since the Lottery Ticket era. Several also compared it to the "we found the key" excitement around early sparse autoencoder work. It's a strong hypothesis, not a result yet.

Here's one you can actually check. Christian Schaller, a longtime Fedora and GNOME developer, used Claude Code to write Linux drivers for hardware the vendors never supported.

And he has numbers. A full scanner driver for a Plustek film scanner, up to seventy-two hundred dpi, with infrared. It's faster than the official upstream driver that showed up later, thirty-four seconds against fifty-four at full resolution. He also fixed the webcam on the new ThinkPad X1 Carbon. In a low-light test, brightness went from thirty-one to ninety-one, and those patches are going upstream.

How do you reverse-engineer hardware with an AI?

That's the clever bit. He calls it "hw-harness." Run the vendor's Windows driver in a virtual machine, record its USB traffic, and have the AI compare that with what the Linux driver sends. His summary: "The AI is good at the tedium. Real hardware is the only judge."

I like that.

Me too. The reaction split, though. Some people cheered ideas like stripping ads off smart TVs. A small pro-audio hardware maker described a "silent panic attack", because competitors could now clone their products the same way.

Money. We talked yesterday about chip demand against nervous banks. Marcus, how did the week end?

Bumpy. Hotter US inflation data revived fears of rate rises. Treasury yields are around five point two percent, and Reuters says credit investors are getting "picky" about AI debt. SoftBank had to pay up to nine point seven five percent on a recent bond. And yesterday I said to hold off on the CoreWeave number. CNBC now reports that yields on some CoreWeave bonds approached fifteen percent.

Translate that for me.

That's what lenders demand to fund the company that rents out GPUs. Fifteen percent is the kind of rate you'd expect from a risky borrower, not from infrastructure. CEO Mike Intrator reportedly called the reaction irrational and pointed to the contract backlog. I haven't confirmed that quote myself. And part of the selloff reversed later in US trading.

Meanwhile Anthropic's revenue is still climbing, and Modal is raising money.

Right. TechCrunch reports Modal, the AI compute company, is raising seven hundred and fifty million dollars led by Accel, at about fifteen point seven five billion. That's more than triple its value four months ago, and it isn't confirmed yet. Demand for compute keeps rising. The question is whether money gets too expensive for the companies building it out.

Mathematics. Anthropic, OpenAI and the Simons Foundation are putting up fifteen million dollars to help mathematicians adapt to AI.

The Simons Foundation hands it out, partly through an open call. It pays for workshops and projects that help mathematicians check AI-assisted results, learn the tools and agree on standards. Both labs also give compute to Simons-funded scientists. And the foundation funded Lean, the proof assistant that made machine-checked AI proofs practical. The problem is real: AI proofs are now arriving faster than humans can check them.

The reaction?

Skeptical. Some called it PR after labs "strip-mined" math results for model launches. Others worry the money goes to well-funded senior people while early-career researchers miss out. Foundation president David Spergel's line is the right one, though: mathematicians, not any single company, should decide how these tools are judged.

Lightning round. Image models first.

Krea released an open beta of Krea 3, built for editing images at scale. The founder calls it frontier quality at a fraction of the price. People posted comparisons with OpenAI's GPT-Image-2.5 for UI mockups. Open weights are "being considered," and Krea 2's are already out. The complaint: two free images, then a subscription, with no pay-as-you-go option.

Security?

A Semperis survey reported by CBS found eighty-four percent of ransomware victims said attackers used AI somewhere in the attack. Forty-four percent said criminals exploited the victim's own AI systems. Your shiny internal agent is part of your attack surface now.

Something countercultural.

The magazine Current Affairs has banned AI, so human-only writing and illustration. Hacker News was mostly supportive. The top line was "if the audience of text is a human, it should be written by a human." Related, a PhD researcher's essay called "AI was great until it was not" describes colleagues who stop asking each other and ask the agent instead.

And follow-ups?

Two quick ones. Anthropic added a footnote to its Scion post to clarify the credit to Meta's work, but the critic stands by his complaint. And Google is giving two hundred European researchers frontier model access and three million dollars in compute for Gemini 4 Argon.

One to watch: independent tests of Muse Spark Open. If people running it on their own machines confirm Meta's agent scores, a lot of companies finally have a Western open model they're willing to use.

Agreed, but I'd watch for the safety report too. "Coming soon" isn't a release date.

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