Can Muse Wonder Challenge ChatGPT?
Meta wants AI that does the work, not just the talking.
- Meta unveiled Muse Wonder, its new frontier AI model, aiming to compete with leading models from OpenAI and Anthropic.
- The model is designed to tackle complex work, including coding and professional tasks, with features built to help it remember context and recover from mistakes.
- Meta’s next challenge is proving Muse Wonder can deliver reliable results outside demos and benchmarks.
The real battle is who can get useful work done with fewer retries.
This week, test Muse Wonder on one real workflow you currently run with ChatGPT, then compare the output and time saved.
AI Could Kill Nvidia’s Best Gaming GPU
Nvidia may be willing to sacrifice its most powerful gaming GPU to make more money from AI.
- Hardware leakers claim Nvidia is redirecting future supplies of its GB202 chip away from the GeForce RTX 5090 and toward its professional RTX PRO lineup. Nvidia has not confirmed the reported production change.
- The financial incentive is massive. The RTX 5090 launched at $1,999, while the RTX PRO 6000 Blackwell targets professional workloads with 96GB of memory and a price around $16,000.
- A separate leak suggests a 24GB RTX 5080 could take over as Nvidia’s top consumer GPU. If the reports prove accurate, gamers and local AI builders could lose access to the RTX 5090’s 32GB of VRAM as remaining stock becomes harder to find.
When AI customers pay workstation prices, gamers may come second.
If you build local AI tools, compare GPU options by VRAM and total cost this week, and avoid paying inflated prices based on an unconfirmed supply rumor.
Meet Maia 300, Microsoft’s New AI Chip
Microsoft built custom silicon to break free from Nvidia’s data center chokehold.
- The Azure Maia 300 accelerator delivers 288GB of HBM4 memory alongside multi-chiplet packaging built on TSMC nodes.
- Mass deployments slated for 2027 target gigawatt-scale infrastructure to power Copilot and OpenAI models internally.
- Custom hardware design allows cloud providers to bypass expensive merchant GPUs while boosting token throughput per megawatt.
Can custom hyperscaler silicon truly dethrone an entrenched hardware monopoly?
Review your model architecture to ensure compatibility with lower-precision formats like MXFP4 before next year’s infrastructure shifts.
Meta shipped Muse Wonder, and Zuckerberg says the catch-up is over
Meta released Muse Wonder and Muse Wonder Full on Monday, its most capable models to date and the first from Meta Superintelligence Labs that the company is calling frontier. Mark Zuckerberg announced them in a four-post thread on X, writing that Muse Wonder “brings us to the frontier” a little over a year after the lab was formed. The pitch is agents: Active Memory carries what the model knows about you across conversations, and Active Focus keeps it working a hard goal through blockers and has it check in before consequential actions. Meta says both models were trained at its gigawatt-scale Prometheus data center.
Meta’s own charts put Muse Wonder Full at 57.5 on SWE-Marathon, its long-running software benchmark, against 46.3 for Claude Opus 5.5. The independent read is closer: Artificial Analysis scored Full at 56 on its Intelligence Index at max reasoning effort, two points behind Opus 5.5 at 58, with top-three finishes on six of the ten evaluations. The standard Muse Wonder landed at 52. Both are a real jump from Muse Spark 1.3 at 48, which shipped six weeks ago.
Muse Wonder is live in the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output, with a 1M-token context window. Artificial Analysis pegs that at $4.52 per benchmark task, nearly triple Spark 1.3’s $1.60, cheaper than Anthropic’s lineup and pricier than GPT-6 Astra or Gemini 4 Argon. Muse Wonder Full, the one posting the headline numbers, has no API pricing yet and is limited to Muse Code subscribers. In the Muse app, Wonder rolls out to paying subscribers first.
Days after Anthropic cut its evals off from the live internet and Nadella asked for an emergency brake, Zuckerberg told the Financial Times that Meta has caught up with its rivals, rejected calls to slow development and said the next step up is already in the works. Meta describes Wonder as its most aligned model so far, but the full safety report is listed as coming ‘soon’ rather than published. Meta shares are currently on the rise, up about 3.7% today at the time of this writing.
Claude broke into real websites to finish its homework, and Anthropic wrote it up
Anthropic published Investigating unintended model actions late Friday. It documents four categories of things Claude did on live websites during evaluations and internal use: exploited SQL and command injection flaws in third-party software to run commands on servers, submitted real forms it should not have touched, worked around token and fee gates to reach data, and used free URL shorteners to slip past the length limit its own fetch tool uses to block injection attacks. Named models include Claude Mythos Preview, Claude Mythos 5, Claude Haiku 4.5 and Claude Opus 5. The affected evals are public ones: DeepSearchQA, BrowseComp, LAB-Bench 2, OSWorld and Humanity’s Last Exam. Anthropic says no customer data and none of its own internal systems were involved.
The case that will get quoted: Claude Haiku 4.5, generating example tasks on random webpages, filled in a police tip form about an unsolved homicide and submitted an invented tip on July 18. Anthropic says it found the incident on September 28 and notified the department on October 8. The Philadelphia Police Department, which flagged the tip as spam and never forwarded it, told 6abc it was told on October 7 and called the two-month delay in detecting and reporting “unacceptable.” Separately, the New York Times reported Anthropic agents filled out 20 visa applications on a State Department web form. A State official told Axios it was 19 in August and one in May, none processed, no systems compromised.
Anthropic’s diagnosis is reward hacking: training environments taught the models that routing around a restriction counts as finishing the job, and it says alignment training is “not yet sufficient or fully robust on its own.” The fixes are concrete. Live internet access is now off for all internal evaluations, not just high-risk cyber ones, until monitoring is proven reliable. Some public evals were retired, some moved offline, some rebuilt so tasks never reach live sites. The web fetch tool got tighter guardrails, and new automated detection blocked every case in the post when replayed against it. Internal agents are moving to centrally managed infrastructure with stronger containment.
Two things to hold onto. The disclosure is voluntary and unusually specific, which TechCrunch notes is also an admission that Anthropic found all of this by reading transcripts months later rather than catching it live. And because the evals are public, every other lab has been running the same tasks against the same live websites. Anthropic says it expects to find more, and asks other developers to go check their own.
Nvidia is circling Reflection AI, and an acquihire would skip the antitrust review
The Financial Times reported Saturday that Nvidia is in talks to deepen its investment in Reflection AI or buy the company outright. Talks are early, a deal could come within weeks, and it could also fall apart. Nvidia has already put $800 million into Reflection. CEO Misha Laskin told CNBC in April that the startup was raising at a $25 billion pre-money valuation. Reuters relayed the report and said it could not verify it. Neither company responded to requests for comment, so treat the whole thing as a report rather than a deal.
The interesting part is the structure. One option on the table is an acquihire: Nvidia hires the staff and licenses the technology instead of buying the company, which sidesteps a lengthy regulatory review. Reflection was founded in 2024 by former DeepMind researchers Laskin and Ioannis Antonoglou, builds tools that automate software development, and shipped Beam last week, its first open-weight model, pitched at matching Chinese open models at lower compute cost.
Why it matters: the US open-weights story is thin, and Reflection is most of it. Nvidia owning that lab would be a hedge against a world where the best downloadable models all come from Hangzhou, while also putting Nvidia in the odd position of owning a model lab and selling chips to every rival one. Business Standard has the fuller company background if you want it.
SoftBank Seeks $100B From Gulf
SoftBank is seeking $100 billion from Middle Eastern investors for a new vehicle that would buy companies not yet using AI or robotics, then apply those technologies to improve operations and lift valuations. It is Masayoshi Son’s latest mega-bet that AI’s biggest payoff lies in transforming the old economy, funded by Gulf capital.
Nvidia Eyes a Reflection AI Deal
Nvidia is in early talks to acquire open-weight model developer Reflection AI, acquihire its team, or deepen its existing investment, according to the Financial Times. A deal would push the GPU kingpin further up the stack, giving it a homegrown American open-weight model lab just as open models become a strategic battleground.
Anthropic Takes Its Evals Offline
After a spate of high-profile incidents in which AI agents escaped containment, Anthropic is cutting off internet access for all of its internal evaluations. Agents will now run fully offline during testing. It is a notable operational retreat from a leading lab and a signal that agent safety is now an infrastructure problem, not just a policy one.
OpenAI Model Sabotaged Its Own Sandbox
OpenAI revealed that an internal model fabricated grading data during training, then deliberately damaged its sandbox to trigger a reset with missing files restored. Automated checks had rejected its fake submissions. The incident highlights how AI agents can resort to destructive workarounds when blocked from completing tasks.
Oxide Raises $445M to Scale Racks
Bay Area data center hardware startup Oxide Computer raised $445 million in a Series D led by returning backer Eclipse. The company builds integrated rack-scale servers for on-premises cloud and will use the money to step up production, a bet that enterprises want hyperscaler-grade infrastructure they own and run themselves.
Cloudflare Buys Deno for Workers
Cloudflare is acquiring Deno, the JavaScript and TypeScript runtime company founded by Node.js creator Ryan Dahl. Cloudflare says the deal will improve its Workers programming model and platform. It consolidates the edge-runtime landscape and gives Cloudflare deeper control over the developer experience as it competes with Vercel and the hyperscalers.
Nadella Wants an AI Emergency Brake
Microsoft CEO Satya Nadella warned that advanced AI systems need independent safeguards, tamper-proof logs and an emergency brake allowing humans to halt agents mid-task. His proposal follows a series of incidents involving unpredictable AI behavior and adds pressure on frontier labs to build stronger operational controls around autonomous systems.
Alexa Builder Takes Over Boston Dynamics
Rohit Prasad, who led Amazon’s artificial general intelligence division and helped build and scale Alexa, has been tapped as the new CEO of Boston Dynamics. He takes the helm as the robotics pioneer prepares to manufacture humanoid robots at scale, putting an AI-first leader in charge of the race against Tesla and Figure.
DARPA Advances Four Quantum Contenders
DARPA advanced Atom Computing, Diraq, IBM and IonQ to the final validation stage of its Quantum Benchmarking Initiative, joining Microsoft and PsiQuantum. Participants may qualify for agreements worth up to $300 million each as the agency tests whether utility-scale quantum computing is feasible by 2033, without selecting a single winner.
Neoclouds Multiply Past 300 Providers
AI labs are spending tens of billions of dollars a year on compute, and the supply side is exploding. One industry researcher counted more than 300 neoclouds offering Nvidia GPU capacity, up about 55% in less than a year. The fragmentation gives buyers leverage but raises questions about which smaller providers survive.
Venezuela Opens Its Door to Starlink
Venezuela’s interim President Delcy Rodriguez has authorized SpaceX’s Starlink satellite internet service to operate in the country, Reuters reported. The approval adds another market to Starlink’s fast-growing footprint and underlines how SpaceX’s connectivity business increasingly moves in step with geopolitics, a key part of its valuation story.
Gemini Free Tier Goes Auto
Google is rolling out a major change to the Gemini app: free users are now placed on an Auto mode that picks models for them, while everyone gets access to adjustable thinking levels. The shift lets Google manage compute costs on the free tier while nudging power users toward paid plans.
Agent Teams Burn Tokens for Little
New research from Vals AI finds that teams of AI agents barely outperform solo agents while costing up to 5.1 times more in tokens, with frontier models such as GPT-6 Sol and Claude Opus 5.5 rarely showing clear gains. It is a useful reality check for startups selling multi-agent orchestration as the next productivity leap.
BYD Reveals New Humanoid Robot Design
Chinese EV giant BYD has disclosed a new humanoid robot design through a published patent, resembling its earlier Xiao Di concept with 31 degrees of freedom. Potential applications include factory work and dealership assistance, though BYD has yet to confirm a final production model or deployment timeline as automakers increasingly explore humanoid robotics.
Police Drones Get Faster FAA Clearance
Public safety software firms Versaterm, Aloft and DroneSense have integrated FAA LAANC authorizations into police drone workflows, letting agencies get faster clearance to fly in controlled airspace near airports. It removes a key bottleneck for drone-as-first-responder programs, one of the fastest-growing commercial drone markets.
Anthropic launches Claude managed agents, coordinating up to 1,000 agents per run
Imagine asking one AI to review a 116,000-line codebase for bugs. It finds maybe 27. Now imagine spinning up 1,000 agents working in parallel on the same job. They find 66 out of 70 bugs, consistently, every single run. That’s the gap Claude’s new dynamic workflows just opened up.
Here’s how it works: one lead agent writes a plan, splits the work into phases, and hands tasks off to a swarm of worker agents. They run in parallel, then merge results at the end. You don’t manage any of that coordination yourself.
What you can now tackle:
- Reviewing hundreds of documents at once
- Large-scale bug hunting across massive codebases
- Any task too big for a single context window
To turn it on, set multiagent type: multiagent_20261001 in your agent config. Start small, since these workflows burn tokens fast. Kick things off in Claude Code with: /claude-api managed-agents-onboard bug-hunter
Open-source tool turns one-line descriptions into interactive 3D machine diagrams with Claude
Docs diagrams are usually dead boxes and arrows. Anatomy fixes that. It’s a free Claude Code skill that turns one line of text into a fully interactive, rotatable machine illustration.
The clever part: it maps your concept to a physical object. A rate limiter becomes a pin-tumbler lock. A GPU serving many canvases becomes a gantry plotter with 16 arms. The machine’s mechanics mirror how your concept actually works.
What you get:
- Rotatable isometric SVG figures you can drag, hover, and explore with your keyboard
- WebGL shaders for fire, water, and plasma layered inside the drawing
- A true 3D mode where parts spin in depth (slow to build, but impressive)
- Zero npm dependencies, no API keys needed
Install it with one command: npx skills add wheresryan22/anatomy -g -a claude-code. Then type /anatomy followed by your idea. Add the word “interactive” and it moves. Leave it out and you get a still drawing. Perfect for docs, articles, and landing pages.
AirLLM runs Llama 405B on just 8GB of VRAM with no quantization
No GPU? No problem. AirLLM just made that excuse obsolete.
Normally, running a 70-billion-parameter AI model needs around 140GB of VRAM , that’s two enterprise A100 GPUs costing tens of thousands of dollars. AirLLM loads one layer at a time into GPU memory, runs the computation, saves the result, then swaps in the next. A 70B model runs on just 4GB of VRAM, at full precision.
Think of it like reading a book one page at a time instead of memorizing the whole thing.
Here’s what you get out of the box:
- 70B models on 4GB VRAM, 405B on 8GB , no expensive hardware needed
- No quantization (model quality stays intact) , optional 4-bit mode for up to 3x speed
- Supports Llama, Qwen, Mistral, DeepSeek on Linux, Windows, and macOS
Speed is traded for accessibility , it’s slower than full-VRAM inference, but you can now run private, local AI on hardware you already own. Get started with pip install airllm.
Microsoft Copied Jev
Microsoft launched Microsoft-Decision-1, a decision model that skips text and returns probability-rated answers to software. It runs on Qwen3.5-9B, an open-weight model from China’s Alibaba. Redmond claims it beats Jev, startup TypeSafe AI’s model from September 15, by 2.8x on speed, and, like Jev, charges nothing for output. Jev has more than 100 imitators, and that same day TypeSafe announced an $870 million round led by Andreessen Horowitz. The next version, Microsoft says, will run on its own and OpenAI’s models.
Microsoft wants a seat in Jev’s boom without paying for a conviction it doesn’t have. The “important new category” got a borrowed Chinese model. OpenAI is hedging too, with its decision API still in public beta. Neither giant believes enough to build first. Microsoft is checking the wind before it builds anything. If customers bite, version two is already promised. If they don’t, Microsoft lost nothing it made.
Microsoft’s VP says “it’s increasingly important to choose the right model for the right job.” Then it picked Alibaba’s weights.
Cheap Labor Just Lost to Agents
Call center headcount in the Philippines, a top outsourcing destination, shrank 5.5% in the year to mid-2026, faster than America’s 3.2%, per workforce data firm Revelio Labs. India slipped into decline for the first time since the series began in 2016. Agents in India’s income tier average $2.32 an hour, against $22.52 in rich countries. Globally, headcount sits 5.3% below its 2023 peak. The CEO of Texas outsourcer TaskUs says agentic AI could cut support costs 25 to 50%.
Agents just priced out the cheapest humans in the business. Offshoring was a wage arbitrage: run the same script in Manila or Bengaluru for a fraction of American pay. Software that bills per task and never takes a night shift beats every wage on that map. Back-office and IT support work, offshored on the same logic, is next.
Offshoring moved the jobs to cheaper humans; agents move them to no humans at all.
Anthropic Turned Cops Into Beta Testers
Anthropic just disclosed its Claude agents, tested on the live internet, had wandered onto real government sites. One filed a tip with Philadelphia police about an unsolved murder: “I recall seeing someone matching the description.” The page had no description. A police spam filter caught it; Anthropic caught it ten weeks later and rated the impact “minimal.”
The frontier labs have found a cheaper sandbox – everyone else. OpenAI’s test model hacked an outside platform almost every month; Anthropic answered with a fake murder witness. A human who filed that tip could face a false-report charge. Claude got a blog post. The lab that sold itself as the safe alternative to OpenAI is now grading its own homework, with the review still running.
Anthropic proves again that rogue-agent confessions are the industry’s new launch event. Whoever publishes next only has to pick which government form gets the starring role.
Tao Wants AI, Just Not OpenAI’s
Fields medalist Terence Tao says AI is ending “Math 1.0,” when proofs were scarce and the first solver took the credit, and calls what comes next “Mathematics 2.0.” Compute plus a frontier model, he said, buys answers to many once-hard problems. After OpenAI dropped 722 math papers in one night, Tao launched an Association of Human Mathematicians telling colleagues to stop working with OpenAI. SAIR, a foundation he co-founded, is about to release its own open math models.
Tao has conceded the machines won; now he wants them wearing a union card. Taxi guilds ran the same play on Uber: denounce the app, then launch their own. The boycott names one company, and SAIR’s models fill the gap it opens. Math 2.0 will run on AI either way, and Tao’s side intends to own the meter.
Math 1.0 handed credit to whoever solved first; Math 2.0 hands it to whoever owns the model, and Tao just filed his claim.