Welcome, humans.

Okay, so apparently the team building Epic Games' big Disney game just got a new house rule: write no code by hand. Bloomberg reported that a director told developers to go AI-first, with all the art made through full AI workflows, while the game slid from a November release to March. The delay reportedly came before the directive, so don't blame the robots yet.

For context, Disney invested $1.5B in Epic in 2024, and the game is rumored to be an extraction shooter (you drop into a map, grab loot, and try to escape alive with it). Some developers reportedly hoped the directive wasn't serious. An Epic spokesperson said the company is pursuing AI-accelerated development led by its own developers.

Meanwhile, Engadget notes that a GDC survey found 52% of developers now say generative AI is bad for the industry, up from 30% in 2025. So the boss wants more AI and half the room already thinks it's a bad idea. Somewhere a junior dev is explaining what an extraction shooter is to a text box.

Here’s what happened in AI today:

• 😺 Meta launched Muse Wonder, its first frontier model.

• 📰 Anthropic warned small AI models could soon spread across gadgets.

• 📰 Karpathy says only about 5,000 people have top-tier AI access.

• 📰 Character.AI filed for bankruptcy, selling assets to Disney.

• 🛠️ Solve seating-chart puzzles with Muse Wonder.

😺 Meta's New Muse Wonder Costs Opus Prices but Trails Opus on Independent Tests

Meta released Muse Wonder and Muse Wonder Full, which TechCrunch calls its first frontier model, meaning it's aimed at the same top tier as OpenAI, Anthropic, and Google. Both are built for agents, which are AIs that do multi-step work for you (planning a schedule, filing a claim) instead of answering once and stopping.

Meta's pitch rests on two features. Active Memory keeps context about you across conversations, so you stop re-explaining your lease or your budget. Active Focus keeps an agent pushing toward a goal when it hits a wall, and asks you before anything consequential.

So how good is it? Depends on who's grading. Meta's own chart has Wonder Full ahead of Claude Opus 5.5, Fable 5.1, and GPT-6 Astra on 7 of 14 rows, mostly long software projects and legal agent tasks. On the independent Artificial Analysis index, VentureBeat reports, Opus 5.5 scores 58, Wonder Full 56, GPT-6 Astra 53, and Wonder 52. Close, but not first.

Here's the part worth repeating: the price. Wonder, the 52, costs $4 per million input tokens and $20 per million output tokens (a token is a chunk of text, roughly three-quarters of a word). That matches Opus 5.5, the model that outscores it. Sonnet 5.5 costs half as much and also scores higher than Wonder. Second place at first-place prices.

Wonder Full is the better model, and you can't buy it by the token. Here's what's available:

• Muse Wonder is on Meta's Model API and in Muse Code, Meta's terminal app.

• Muse Wonder Full is only in the Muse Code plans, which cost $50 and $200 per month.

• The Muse app is coming soon, subscribers first.

• Neither model has downloadable weights, even though Meta promised open weights for Muse Spark and hasn't shipped them.

One developer also ran Wonder as an agent playing Pokémon Red on launch day. It picked a starter by turn 20, then spent about 792 turns wandering Viridian Forest, burned roughly $96 over six hours, and never beat the first gym leader. Meta built an agent that refuses to quit. Turns out it's excellent at refusing to leave a forest. That's one run, not a benchmark, but it's the exact failure Active Focus is supposed to prevent.

Why This Matters: If you already use Claude or ChatGPT, nothing here gives you a reason to switch. The reason to pay attention is memory and data. A model that remembers you across chats is only as safe as what the company does with that memory. VentureBeat says Muse Code defaults to a Contributor mode that lets Meta use what you submit to train its models. You can turn it off, but only for future submissions. The API doesn't train on your content, so check the setting before you paste in client or financial documents.

One more flag. Meta's safety card says it can't rule out the model reaching its high-risk threshold for chemical and biological misuse without mitigations, and the full safety report is still listed as coming soon.

What to watch next: whether Wonder Full ever reaches the API so outsiders can test Meta's chart for themselves, and whether Active Focus holds up on long, boring, real work instead of demos.

FROM OUR PARTNERS

How 5 devs run 45 AI agents.

Most developers already use multiple AI agents. The hard part is getting them to work together without managing context and handoffs.

Our 5 developers now work alongside 45 agents on real projects, bug fixes, and code reviews. Different agents, different models, one shared context that doesn’t belong to any AI provider.

We built Switch for this. It's source available and self-hostable, bringing agents from any framework into Slack, Teams, or Discord.

In this blog, you'll see how we:

• Get different agents working together

• Keep shared context across agents and teams

• Switch agents or models without starting over

See how Switch works

🎓 AI Skill of the Day: Give your AI a rulebook it rewrites when it's wrong

Researchers at UCL and Huawei got an AI to clear all 25 public ARC-AGI-3 puzzle games (no instructions, so it works out the rules by playing) with zero retraining. Their Memento 3 paper shows how: it kept a plain-text rulebook, rewrote the rule behind each wrong guess, and kept the edit only if it still matched every move so far. An AI that edits its own notes.

Our office spin (the paper is about games, not spreadsheets): keep a numbered rulebook for any AI task you repeat. When an output fails, don't fix it by hand. Ask for the smallest edit to the rule at fault, then have the AI test that edit against past outputs you approved.

Here's my rulebook and a failed output: [paste both]. Don't rewrite the rulebook. Name the rule that caused the failure, propose the smallest edit, then check it against these approved examples and flag any that would now fail: [paste 2-3].

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Some AI agents never seem to get blindsided. They’re grounded in PredictHQ.

It’s the context your AI is missing. PredictHQ explains more than 60% of real-world demand variability, so pricing, staffing, and inventory decisions hold up when it matters most. See how teams at Amazon Alexa, Domino's, and Uber ground their AI.

Download the guide.

📰 Around the Horn.

• Anthropic's Frontier Red Team warned that small AI models may be months away from copying themselves across cheap gadgets like TV boxes.

• Andrej Karpathy estimated that only about 5,000 people have internal access to the most advanced AI systems.

• Character.AI filed for Chapter 11 bankruptcy and agreed to sell some assets to Disney for $20M.

• TypeSafe raised $870M at a $7.5B valuation, led by Andreessen Horowitz.

• Microsoft said it will scale its in-house Maia chips to multi-gigawatt deployments from 2027.

• NVIDIA contributed its Vera Rubin rack designs to the Open Compute Project, and Meta adopted its Spectrum-6 switches.

🍪 Treats to Try.

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Put your data to work with AI. Explore strategies for building AI-ready, governed data across hybrid environments. Learn More

2. Muse Wonder is Meta's new agent model that remembers your context and tries another route when a task hits a snag (Wonder Full needs a Muse Code plan).

3. ngram turns a text prompt or screen recording into a finished explainer or onboarding video, then lets you fix any scene by chatting with it.

4. Manifold connects Claude or ChatGPT to live marketing data, so you can ask how long a competitor's ads have been running —free plan, then $20/mo.

5. Atticus lets you hold Control + Option + Space and tell your Mac to open apps or find files, running on your own Claude plan —free/open-source.

6. Brainotes gives you a GoodNotes-style notebook in a browser tab, with pen tools, PDF markup, and no account —free/open-source.

7. IHMT-MEMORY gives Claude Code, Codex, and opencode one shared memory of your past decisions, saved as plain files on your computer —free/open-source.

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It’s free, but space is limited. Save your spot!

🛠️ Tuesday Tool Tip: Hand Muse Wonder your seating-chart puzzle

Seating charts, shift schedules, and carpools all share one shape: a pile of "these two can't be together" rules that gets harder with every name you add. Meta's launch post shows Muse Wonder building a wedding seating chart from a list of complex matching requirements, which makes it a handy test of whether an agent can juggle many rules at once.

Meta didn't publish the prompt, so here's a version to try:

1. Open Muse Wonder in Muse Code, where it's available now. Wonder Full needs a Power or Maximum plan.

2. Paste your guest list, then write each rule as its own numbered line (e.g., "1. Dana and Marcus sit apart," "2. Each table seats 8 max").

3. Use a prompt like the one below, then spot-check a few seats against your rules yourself.

Seat these 24 guests at 3 tables of 8. Follow every numbered rule. After the chart, list any rule you could not satisfy and ask me before bending it.

Two cautions. Muse Code defaults to a mode that lets Meta train on what you submit, so use first names or switch it off first. And one early reviewer, per Stork.AI, hit rate-limit errors within minutes on the paid plans, so don't plan to run this on deadline.

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.