Meta brings its most capable open model to Apple’s Macs
Meta released the weights for Muse Spark Open, one of the strongest American-made open models currently available. The model is optimized to run on Apple silicon, but there’s one catch: you’ll need at least two top-spec M5 Ultra Mac Studios. The release signals a thawing relationship between the two rivals, with Meta’s Alexandr Wang telling Axios, “We’re excited for mommy and daddy to be playing nice again.”
Goodfire researcher says neural networks may be cleaner than previously thought
In an informal research essay, Nick Cammarata — an interpretability researcher who previously worked at OpenAI — argues the field has been wrong about a core assumption: that models are nearly impossible to interpret directly. He says this is a misconception that stemmed from flawed architectures and weak inspection tools. If he’s right, it could help researchers better understand what models are doing before releasing them to the public.
Amazon’s Bee unveils two AI wearables that take notes for you
Bee Frontier, worn on your wrist, and Bee Eclipse, which attaches to your phone, are lightweight devices that summarize conversations to help you remember what matters most. On privacy, Amazon says a glowing light indicates when the device is listening and recordings are never stored long-term. They’re Bee’s first new devices since Amazon acquired it last year. Watch how they work.
How to build a personal decision simulator with Claude
- Open Claude > Projects > New Project. Name it after the decision you’re working on.
- In the Project, click the + in the knowledge section and upload any useful files, notes, budgets, comparison tables, or research.
- Start a new chat inside the Project.
- Sample Prompt: “Build an interactive decision simulator for [decision] comparing [Option A/B/C] across [criteria]. Use adjustable weights, 1–10 scores, and automatic weighted totals for each option. Add editable Best/Base/Worst Case scenarios and show how each changes the rankings. Highlight the leading option, key trade-offs, and the variables driving the result most.”
- Open the Artifact Claude creates beside the conversation and check that the weights, scores, scenarios, and totals are interactive.
- Adjust the sliders or input fields yourself to test different priorities, costs, risks, or assumptions and watch the ranking update.
- Use the chat beside the Artifact to request any structural changes, such as adding a criterion, changing the scoring scale, or adding another scenario.
- Reopen the finished simulator later from the Artifacts tab and reuse it with new decisions or updated assumptions.
Meta’s Open AI Needs 1TB of Mac Memory
Apparently, corporate grudges have a RAM limit.
Meta released Muse Spark Open on Thursday, trailing right behind this week’s proprietary Muse Wonder launch. But unlike its corporate siblings, Spark Open lets developers download the raw weights for free—provided you don’t mind dropping a small mortgage on the hardware to run it.
Packing 816 billion parameters and a massive 1-million-token context window, the multimodal beast natively digests text, code, audio, video, and images. Meta claims its internal trials included fully on-device operations like editing DaVinci Resolve video timelines and pushing code fixes directly to an iOS app—all without pinging a remote cloud server.
Your trusty MacBook Air can sit this one out. Running the model locally requires roughly 1TB of unified RAM. That’s either four 256GB Mac Studios linked together or two 512GB M5 Ultra behemoths when that larger configuration expected later this month finally ships. The Wall Street Journal estimates compatible Mac setups will cross the $40,000 threshold. So while the software download is technically free, it makes for a shockingly pricey desk ornament.
After half a decade of passive-aggressive privacy feuds, Apple and Meta actually worked together to optimize the model for Apple’s MLX framework. Meta AI chief Alexandr Wang credited former Apple executive Daniel Gross, now at Meta, with brokering the unlikely truce. His assessment to Axios: “We’re excited for mommy and daddy to be playing nice again.”
Don’t shed a tear for Jensen Huang, either. Meta trained the model on Nvidia’s GB300 NVL72 systems, and Team Green is happily supporting GPU deployments for anyone not buying Mac hardware by the pallet.
On-device execution keeps sensitive developer prompts off Meta’s servers and could dramatically slice long-term cloud bills for heavy users. Apple gets to sell eye-wateringly expensive hardware, Meta gets broad open-weight adoption, and your IT department gets to explain to finance why its coding assistant needs four separate computers to run.
Apple’s China AI Rollout Leaves New Siri Behind
Apple Intelligence is clearing customs. Its newest Siri is still waiting.
Apple is reportedly aiming for a December rollout of Apple Intelligence in mainland China via iOS 27.3, but the fancy new Siri AI is getting left at the border. After years of regulatory purgatory, the initial release will offer a remarkably trimmed-down toolset.
There’s a smaller appetizer coming sooner. Apple Executive Chairman Tim Cook told The Paper on Thursday that a few core features have finally scored approval for a late-October beta. That test includes Live Translation, Writing Tools, and Clean Up in Photos—though Cook coyly dodged confirming the December timeline.
Bloomberg’s reported December lineup adds notification summaries and original Genmoji to the mix, while strictly banishing Image Playground, upgraded Genmoji, and the revamped Siri. Throw in app-specific restrictions on text generation, and you’re getting Apple’s greatest hits from 2024 instead of the new album.
The red tape doesn’t end at launch, either. Apple’s partnership with Alibaba reportedly includes a screening mechanism that lets regulators thoroughly babysit future AI model updates. Meanwhile, local rivals like Huawei, Xiaomi, and Oppo are happily flaunting their own native AI capabilities while Apple plays catch-up.
For IT teams, mainland China just became a completely separate deployment puzzle. Apple’s documentation confirms that existing compatible hardware will support the initial capabilities, saving you an immediate upgrade headache. Just triple-check those local data-handling terms before promising your employees a futuristic assistant experience.
Designed in California. Availability assembled separately.
Microsoft’s Maia 300 Takes Aim at AI Costs
Microsoft unveiled the Maia 300 earlier this week, slating production deployments for 2027 with a grand mission to make its homegrown silicon the primary engine for AI inference across its product suite. That’s the grunt work of actually running trained models to answer questions instead of just staring blankly at a prompt.
This expansion comes with an appetite that would make a utility company blush. Microsoft is plotting multi-gigawatt deployments of the Maia family starting next year, while Reuters reports a stash of potentially more than 1 million Maia 300 chips in 2027. Looks like Nvidia’s favorite customer is finally getting serious about cutting the cord.
Microsoft expects the Maia 300 to pump out 40% more tokens per megawatt than “leading GPU-based systems”—conveniently refusing to name whose GPUs those might be. It also boasts up to seven times the rack-scale token throughput of its own Maia 200. Sure, comparing apples to your own older, smaller apples is fun, but neither metric actually establishes a benchmark victory over Nvidia’s Rubin platform.
The hardware crams 96 accelerators into a 167-kilowatt rack. Its chiplet design relies on TSMC’s N2P process for the actual brainpower, while memory and networking are engineered to reduce the energy wasted shuffling data around. Because even AI needs to cut down on its miserable daily commute.
For Azure partners and customers, the ultimate payoff remains a giant question mark. Maia is strictly an inside job serving Microsoft’s own infrastructure right now, with no separate rental offering on the menu. As Channel Insider wisely points out, Microsoft’s newfound hardware efficiency doesn’t automatically trickle down into customer savings. You might want to wait for actual service performance and contract pricing before promising your boss a smaller cloud bill.
Right now, Microsoft’s only customer for cheaper AI is… Microsoft.
Meta put an 816B open-weight model on two Mac desktops
Meta Superintelligence Labs released Muse Spark Open on Thursday and called it the most capable US open-weight model to date. It is natively multimodal, 816B total parameters with 42B active, and carries a 1M-token context window, aimed at coding and long-horizon agentic work. Weights are up on Hugging Face, Microsoft Foundry, Google Cloud Model Garden and Amazon SageMaker, with hosted APIs from Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI.
The headline is the hardware. Meta says it runs on two Apple M5 Ultra Mac Studios over MLX, each with up to 512GB of unified memory and 1.2TB/s of bandwidth. Read the fine print: the model card asks for at least 700GB of accelerator memory for the NVFP4 build and 2TB for BF16, and the M5 Ultra Mac Studio is not shipping until late October at an unannounced price. Meta published no measured token rate, and the local demos are company-run. Together AI lists it at $1.25 in and $4.25 out per million tokens.
On the scoreboard it is good, not dominant. Meta’s own table has it leading JobBench at 61.8 and MCP-Atlas at 89.2 while trailing GLM 5.3 and Kimi K3 on terminal coding, DeepSWE and GDPVal-AA. Its Artificial Analysis Intelligence Index score is 39 per Meta’s model card, against independently listed scores of 46 for MiMo-V2.6-Pro, 45 for GLM-5.3, 44 for Kimi K3 and 40 for Qwen3.8. The Muse Model License is not quite open either: model-as-a-service businesses over $20M in annual gross revenue have to notify Meta and sign a separate agreement, and products must say “Built with Muse Spark Open.”
Note what Meta did not release. Its closed flagship Muse Wonder scores 52 on the same index, and Bloomberg read the launch as Meta keeping its best tech closed. That is the honest frame. The useful part is that a US lab is competitive on open weights again after a year of ceding that ground to Chinese labs, and that the deployment story now includes a desk instead of a datacenter.
Microsoft finished a 2.8T base model and will not open it
Microsoft AI published its pre-training report for MAI-Base-2 on Thursday: a mixture-of-experts model with 2.8T total parameters and 80B active, MoE in every layer, trained on a 50T-token mixture whose unique-token corpus grew about 36% over MAI-Base-1. The vision-language portion is 4.5T unique tokens and 18% of the mix, and web data carries a March 2026 cutoff. The run took 16,000 GB200 GPUs over 38 days across August and September, at 21% average MFU and 89% goodput.
The efficiency claim is the point of the post. Microsoft reports MAI-Base-2 is 4.2x to 12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base and 6.3x to 12.7x more efficient than its own predecessor, and says it achieves lower bits-per-byte than every open-weight model it tested on held-out evaluations, naming DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. Two caveats that Microsoft states itself: the efficiency figures are estimates from scaling-law fits rather than direct measurements, and the eval scores live in the charts rather than the text. Separately, CEO Mustafa Suleyman told Business Insider it beat a comparable OpenAI base model on about a quarter of the compute, based on private tests nobody outside Microsoft can check.
The model is still in post-training. There is no release date, and the weights go to trusted partners through Foundry, not to Hugging Face. Pair that with Monday’s Maia 300 announcement, which claims 40% more tokens per watt than leading GPU systems with 288GB of memory and 18TB/s of bandwidth, entering production in 2027 as Microsoft’s primary inference system, and the strategy reads clearly: own the silicon, own the base model, keep both.
Nothing here is usable today. It matters because a pre-training report with real numbers from the company that resells OpenAI’s models is a statement about who Microsoft intends to depend on.
FTC probes OpenAI, Anthropic
The Federal Trade Commission has launched formal investigations into OpenAI and Anthropic over AI safety practices and consumer protection concerns. The probe marks the first major federal regulatory action targeting the two largest frontier AI labs simultaneously, signaling a shift from voluntary commitments to potential enforcement.
- The FTC is examining whether the companies’ public safety claims align with their actual practices, focusing on potential deceptive advertising under existing consumer protection laws.
- Investigators are also reviewing the voluntary safety commitments both companies made to the White House in 2023, assessing whether those pledges have been meaningfully implemented.
- Neither OpenAI nor Anthropic has commented publicly on the investigation, though both have previously stated they welcome regulatory oversight of AI development.
For enterprise buyers, this investigation raises immediate questions about vendor risk and compliance exposure. Companies deploying OpenAI or Anthropic tools should monitor the probe closely, as findings could trigger new disclosure requirements or usage restrictions. If the FTC determines safety claims were overstated, it could reshape how AI vendors market their products and force enterprises to reassess their AI procurement criteria.
Meta launches its “most capable” open-weight AI model, built to run locally on Macs
The 816 billion parameter model, called Muse Spark Open, comes with a 1 million token context window and native multimodal capabilities. Meta says it outperforms other popular open-weight models on standard benchmarks, with particular strength in complex coding tasks and agentic workflows.
The catch: running it locally requires at least 1 terabyte of RAM, effectively limiting on-device use to high-end Mac Studio machines equipped with M5 Ultra chips. Meta demonstrated the model paired with its Muse Code terminal, showing it could handle full build cycles including bug fixes and new features for a Mastodon client, as well as reviewing and cropping footage in DaVinci Resolve.
Open-weight models offer cost efficiency and transparency that proprietary systems don’t. If Muse Spark Open delivers on Meta’s performance claims, it could reshape how enterprises approach local AI deployment for coding and creative workflows.
Meta gives the West its best open-weight model
Meta just released Muse Spark Open, an open-weight, 816B-parameter multimodal model built for coding and agentic work, and worked with Apple so that it could run locally on Mac Studios — a first for a model in this size class, Meta says.
- Two M5 Ultra Mac Studios with 1TB combined memory can run Spark Open via Apple’s updated MLX toolkit, with Meta saying the model shipped iOS app fixes and edited video in DaVinci Resolve without the cloud.
- On Artificial Analysis’ Intelligence Index, it sits at 39, ahead of the West’s open-weight models but behind China’s GLM 5.3 (45), Kimi K3 (44), and Qwen3.8 (40).
- Spark Open either matches or tops open rivals on multimodal and knowledge tasks, but Chinese models still code better, with Kimi, GLM, and Qwen all ahead.
- It’s the third major Western open launch this month, after Reflection AI’s Beam and Mistral’s 1T-parameter Le Chonk, whose weights arrive Oct. 27.
Meta built its AI rep on open Llama models before going closed. Now it’s playing both sides with Wonder at the frontier and Spark Open giving devs serious AI they can run on their hardware. China still leads the open-model race, but Meta’s return puts an American heavyweight back in contention.
Turn a ChatGPT answer into a tool you can use
In this guide, you’ll try ChatGPT’s new Intelligent UI by creating a workshop cost calculator. You’ll change the headcount, check the figures, and ask it to add a budget buffer.
Step-by-step:
- Open ChatGPT in a web browser and start a Chat with GPT-6. Choose Medium where available and leave Pro reasoning off for this exercise
- Paste the guide’s prompt to create a calculator with editable costs, budget, and attendance. Check that the starting total is $700 for 20 people
- Change attendance to 30 using its controls. The total should rise to $950 and show that you are $50 over budget
- Try zero attendees and check that the fixed room fee stays. Test the budget boundary too, so a useful-looking calculator does not hide a wrong result
- Return to 20 people and ask for a 10% buffer. Check for a $770 final cost and $130 left, then change the headcount again to test the update
Pro tip: Tell ChatGPT which inputs should change and which rules stay fixed. After each revision, check one result you know and one new case before using the tool.
Report: Harvard maps AI’s gender divide
Harvard researchers just found consistent gender gaps in how Americans use AI, get support for it, understand it, and appear in coverage of it, drawing on a survey of 5,400+ adults, browsing data, and 78K news articles.
- 29% of women said they’d never used AI vs. 22% of men, though browsing data showed no real gap in trying it, only in how much time users spent on it.
- At work, 64% of employed women reported no AI training vs. 49% of men, with the study linking training and encouragement to higher use and confidence for both.
- Men answered more AI questions correctly while women chose “not sure” more often, though more men also claimed to recognize made-up AI figures.
- AI headlines named men 14.6x as often as women, a gap the researchers partly tie to male-led companies, though the gap also held among quoted experts.
This is one of the broadest looks yet at how AI’s gender gaps are taking shape. Access looks nearly equal today, but the real divide shows up in training, confidence, and who gets heard. As AI labs pitch superintelligence for everyone, these gaps show how far that promise still has to go.