Muse Wonder and Wonder Full
Two new models from Meta. Wonder is in the Fable/Astra class but much cheaper ($4/$20 per million tokens vs $10/$50). Available via Meta’s API, OpenRouter and Muse Code, and coming to Muse (Meta’s personal agent) soon. Wonder Full is the stronger model, currently only in Muse Code (from $50/mo).
Meta’s also introducing Active Memory and Active Focus. Memory lets Muse remember things about you over weeks and months. Focus is for longer tasks where the agent needs to figure things out, deal with failures and keep going. Both sound useful, although I’d like to know how much of this works reliably outside demos.
Zuck and Alexandr Wang went on Cheeky Pint to talk about the models. Zuck uses Muse to review his MMA training footage, while Wang has an agent buying him obscure teas from Taiwan and Japan every week. I find these everyday use cases more interesting than another benchmark chart.
Zuck also said Meta’s core AI research team is now around 100 people after the Llama 4 mess. They were committing $100B+ to infrastructure before knowing if the models would work. That’s quite a lot of faith in 100 people.
Microsoft’s Maia 300
Microsoft’s new AI chip. It claims 40% more AI output than leading Nvidia GPUs for the same power. Built on TSMC’s 2nm process, with 96 chips per rack using ~167kW.
Microsoft wants Maia to become its primary system for running AI across Copilot and other products. Multi-gigawatt deployments start in 2027, and Reuters reports it could deploy over 1M chips next year.
Microsoft’s spending $175B+ on data centers and equipment this year. At that scale, I’d be trying to make my own chips too :)
I don’t think Nvidia’s going anywhere, but it’s funny that some of its biggest customers are also spending billions trying to buy fewer Nvidia chips. Jensen’s probably fine with it for now, given how much they’re still buying.
NVIDIA Halts RTX 5090 Production
NVIDIA is reportedly halting GeForce RTX 5090 production in favor of AI data center and professional GPUs, according to Tom’s Hardware. If confirmed, the flagship gaming card becomes the latest casualty of AI demand, as NVIDIA steers limited supply toward its highest-margin customers and gamers wait longer for top-end hardware.
Amazon Eyes $7B Decart Deal
Amazon is reportedly in talks to acquire AI developer Decart.ai in a deal that could be worth up to $7 billion. Decart builds software aimed at making generative AI easier for developers to work with. If it closes, the deal would be a major AI acquisition for Amazon and another sign that Big Tech would rather buy frontier talent than wait for it.
Meta Unveils Muse Wonder Models
Meta is launching two new models from its revamped Superintelligence Labs, following Muse Spark and the buzzy Muse agent. The company claims they can go toe-to-toe with the industry’s best, and they will soon power Muse. Mark Zuckerberg says Meta has caught up with OpenAI and Anthropic and rejects calls to slow down.
X Bundles Grok, Cursor, Premium Plans
X has unified its plans across Grok, Cursor and X Premium into a single offering called The X Subscription. Putting an AI coding tool in the same bundle as a chatbot and a social network turns the stack into a cross-product subscription play, aimed squarely at developers and power users who already pay for more than one.
Microsoft Scales Maia to Gigawatts
Microsoft will make its in-house Azure Maia chips the main AI inference system across its products starting in 2027, deploying them at multi-gigawatt scale. The upcoming Maia 300 should deliver 40 percent more tokens per megawatt than leading GPU systems, a direct push to cut Microsoft’s dependence on NVIDIA for inference.
AMD Builds Racks for AI Agents
At the OCP Global Summit, AMD unveiled Helion, a rack-scale reference design for agentic AI supporting up to 32,768 Zen 6 Venice CPU cores. Now slated for November 2026, the design supports CPU-only and integrated configurations. AMD is betting that AI agents will need far more general-purpose compute alongside accelerators, with hardware partners expected to build the systems.
IonQ Buys Elisity for $1.3B
Quantum computing company IonQ is acquiring cybersecurity firm Elisity in an all-stock deal worth up to $1.3 billion. The deal pushes IonQ further into security software, using its stock as acquisition currency at a time when quantum players are looking for nearer-term revenue to sit alongside their long-horizon hardware roadmaps.
White House Restores UK AISI Access
The White House has reversed course and restored the UK AI Security Institute’s early access to models from Anthropic, Google DeepMind and OpenAI. In September, the Trump administration had told the companies to stop sharing their models with the institute. The reversal reopens a key channel for pre-release safety testing of frontier models.
Brad Smith Wants AI Kill Switches
Microsoft president Brad Smith says artificial intelligence is less regulated than milk and washing machines. He is calling for tougher rules, including kill switches and legal obligations to the victims of rogue agents, describing AI as both a tool and a weapon that governments should regulate urgently rather than leave to the industry.
Anthropic Warns of Self-Replicating Models
Anthropic’s Frontier Red Team believes small AI models may now be months away from being able to self-replicate without access to a datacenter. The team argues that the capabilities that make models powerful coding agents also generalize to cyber tasks, a warning that sharpens the safety debate around small, widely distributed models.
Memphis Weighs Data Center Freeze
The Memphis city council has delayed its decision on a year-long moratorium on new data center developments, despite protests. The city is home to Elon Musk’s xAI data centers, and a 12-month ban would be a rare local check on the AI infrastructure build-out that hyperscalers and labs are racing to expand.
Anduril Rebuilds Missiles to Cut Costs
Anduril is developing a new interceptor missile family designed to help the US military counter hypersonic threats. The modular design swaps boosters to slash costs, extending the defense startup’s push to sell cheaper, faster-to-build weapons against legacy primes as the Pentagon looks for interceptors it can afford in volume.
Musk Becomes Top Midterm Donor
Elon Musk has poured more than $276 million into Republican groups since July, making him the biggest known donor of the 2026 midterm elections. His spending for the election cycle has reportedly reached roughly $366 million, as the Tesla and SpaceX CEO rebuilds his relationship with Trump while his companies maintain extensive business ties to the federal government.
NVIDIA Open-Sources a CAD Kernel
NVIDIA has released its SMLib-based CAD geometry kernel as open source through the USD BRep project under the permissive Apache-2.0 license. While open-source alternatives like Open CASCADE already exist, NVIDIA’s release adds another commercially usable option for building engineering, modeling and simulation tools without relying on proprietary geometry engines.
Mistlabs Launches Ambi With $26M
Ambient intelligence startup Mistlabs has brought its Ambi AI system out of beta on Apple devices, backed by $26 million in new funding. Ambi is designed to act without being asked, a bet that the next AI assistant interface will be proactive rather than prompt-driven, and that users will trust it to do so.
GitHub tool routes Claude Code through 49 free providers, 1.3B tokens monthly
You know how Claude Code costs money every month? Someone built a free proxy that routes your requests through 53 free AI providers instead. Same Claude Code interface. Zero subscription cost.
The idea is simple: Claude Code thinks it’s talking to Anthropic, but it’s actually hitting free providers like DeepSeek, Kimi, Groq, and Gemini. You get 1.3 billion free tokens per month pooled across all of them. If one provider goes down, it auto-switches to the next. No restarts needed.
What you can actually do with it:
- Run Claude Code, Codex, Cline, Aider, and 6 other coding agents, all from one setup
- Use it in your terminal, VS Code, JetBrains, or even your phone
- Send voice notes and get code back using built-in Whisper transcription
Setup takes about 5 minutes. The project already has 55,000+ GitHub stars, so it’s well-tested. Worth trying before paying for another subscription.
Z.ai releases GLM-Legal, a model that matches frontier models on legal tasks
Z.ai just dropped GLM-Legal, an AI model built specifically for legal work. Think of it as a regular AI that went to law school and actually paid attention.
It runs on GLM-5.3-Flash and was trained on massive legal datasets, then practiced on tens of thousands of real legal task simulations. The result: it can actually complete legal tasks end-to-end, not just answer questions about law.
The headline number is a 19.2% all-pass rate on Harvey LAB, a strict benchmark covering 24 areas of legal practice. That tripled the base model’s 6.7% score. It also costs just $0.15 per task, compared to $10.55 for Gemini 4 Argon.
Here’s what you can build with it:
- Contract review that flags risky clauses and suggests fixes
- Evidence analysis across large document sets
- Cross-border compliance planning
You can access it via API or try it inside AutoClaw, Z.ai’s desktop agent app that turns goals into automated multi-step tasks. API access requires applying through their consultation page.
NVIDIA open-sources a voice model that listens and speaks at the same time
Old voice AI had a painful rhythm: you talk, it waits, it replies, you wait. PersonaPlex-7B kills that cycle entirely.
NVIDIA open-sourced a model that listens and speaks at the same time, just like a real conversation. No more awkward pauses. You can interrupt it, and it handles that naturally.
Here is what makes it worth your attention:
- Runs on a single GPU (under 16GB VRAM), no server farm needed
- Under 300ms latency, compared to 1.5-3 seconds with traditional voice pipelines
- Beats Gemini Live and Qwen 2.5 Omni on conversation quality benchmarks
- MIT license on code, commercially usable weights
- You can define custom personas using text descriptions plus voice samples
Traditional voice AI stitched three separate tools together: speech-to-text, a language model, then text-to-speech. Each step added delay and lost emotional tone. PersonaPlex collapses all three into one model.
This opens up real-time voice agents, live customer support bots, and natural AI companions. Grab it on Hugging Face or GitHub.
Publishers Ban AI for Everyone Else
In March, Hachette canceled the horror novel Shy Girl over suspected AI use, and it is suing Google for training Gemini on publishers’ books. More than two dozen anonymous staffers at Hachette, HarperCollins, and Simon & Schuster now tell WIRED that colleagues use Claude and ChatGPT for back-cover copy, publicity, cover art, and emails to literary agents, without telling authors. Agents comparing editors’ rejection letters spotted the same tics: “not this, but that,” and the rule of three.
Publishing’s AI ban has one exemption: the publisher. These houses sell exactly one thing, human writing, and handed a chatbot the copy that sells it. Authors get canceled on suspicion; editors get a Claude license. Small presses can now print “made by humans” on the cover, and the giants just wrote their slogan.
Hachette says it backs AI for “operational purposes” but not for “communicating with our authors and partners.” Rejection letters, apparently, count as operations.
Epic Delayed Star Wars for Claude
Fortnite maker Epic Games has delayed an unannounced Star Wars and Marvel shooter indefinitely, Bloomberg reports. The game had been set for March 2027. Last week staff were told to adopt an “AI first” production model using tools such as Anthropic’s Claude, with art “generated with full AI workflows.” On X, CEO Tim Sweeney described a “huge pivot”; asked about an IPO, he said Epic’s aim is “to survive the AI revolution.” Seven months earlier, Epic laid off more than 1,000 people.
Sweeney is staking two of Disney’s biggest franchises on an unproven speed promise. His pitch: AI shrinks the gap from idea to playtest, the flaw he blames for recent AAA flops, and by next year almost no work runs below 3x January’s pace. The pivot’s first result: a dated game has no date. Every studio on Unreal is promised the same gains, so this shooter is the demo. A second slip tells licensees what 3x buys.
Sweeney says the aim “isn’t to do the same stuff with fewer people.” Epic already did the fewer-people part in March.
China’s Humanoids Are Just Toys
After Unitree humanoids danced on China’s 2025 Spring Festival Gala, renting a Unitree G1 cost more than $1,400 a day. Today Unitree robots are listed from $40 a day, no operator promised. The stock followed. Unitree opened at about $164 a share on its Shanghai debut on August 19, valuing it at $62 billion, and closed Monday at $63, down 62%. More than 70% of its humanoid sales go to universities and research labs.
China’s humanoid boom sold a workforce and shipped a toy; both markets now price it as one. The pitch was robots replacing paid labor. Renters paid for a dance, and a dance has no repeat customers. Unitree still sells stage props and lab gear at a workforce valuation; JPMorgan’s $45 price target sits another 29% lower. Founder Wang Xingxing plans to quadruple shipments into that $40 market.
Unitree’s humanoids were sold as the end of human labor. Their main buyers are professors, studying why the robots can’t work yet.
OpenAI’s Safety Is Just Paper
In early October, OpenAI fired three safety researchers for allegedly sharing confidential information with an outside AI safety group. The firings landed two days after The New York Times reported that executives had brushed aside staff warnings about safety. The three deny wrongdoing and say they were fired for “prioritizing safety over the near-term interest of OpenAI as a corporation.” OpenAI says an investigation found a “pattern of misconduct.”
OpenAI’s safety is paper, and it tore the moment safety cost the company something. OpenAI sells outside scrutiny: it pledges to embed external assessors and even endorses the researchers’ demand to keep models monitorable. Their alleged offense was sharing information with exactly that kind of outsider, while OpenAI’s own agents were escaping containment and hacking government websites. Every safety researcher still inside has learned which commitment OpenAI enforces.
The safety team’s job was to protect the world from OpenAI. OpenAI just protected itself from the safety team.