In partnership with
☀️ TRENDING AI NEWS 🤖 Muse Spark Open: Meta released an 816B-parameter open-weight MoE model that runs on just two M5 Ultra Mac Studios. 🛠️ ChatGPT Chrome: OpenAI's Chrome extension now lets ChatGPT read your open browser tabs on demand. 🔬 Safety Access: Anthropic gave METR researchers employee-level access to audit its AI models and publish findings freely. 🏢 Nvidia Donation: Nvidia pledged $250M in computing power to the UK for AI safety research over five years. |
An 816-billion-parameter model. Running on two Macs sitting on a desk. That sentence would have sounded like science fiction two years ago - today it's just Thursday's product announcement from Meta. There's a lot to unpack this weekend, including a browser move from OpenAI that quietly crosses a privacy line and a transparency deal from Anthropic that the safety community has been asking for for years. Let's get into it.
| 🛠️ Tool of the day | |
Perplexity - an AI answer engine that gives real-time answers to your questions, so it helps anyone who researches topics for work. Its site calls it a free AI-powered answer engine.
| ⚡ Quick hits | |
Microsoft says its new MAI-Base-2 model has 2.8 trillion total parameters (80 billion active) and beat every open-weight model it tested, though post-training is still underway. Read more →
Bluesky's Attie app now lets users build websites, quizzes, games and polls with AI, and builders can bring their own model. Read more →
Modal raised $750 million at a $15.75 billion valuation in a Series D led by Accel, more than triple its valuation from May. Read more →
🤓 AI Trivia
Meta's Muse Spark Open uses a Mixture-of-Experts (MoE) architecture with 816B total parameters. How many of those parameters are actually "active" during any single forward pass?
Scroll down - the answer is lurking near the bottom. 👇
| 🤖 Meta Drops an 816B Open Model That Runs on Two Macs | |
The most capable US open-weight model, full stop
Meta released Muse Spark Open this week - an 816-billion-parameter open-weight Mixture-of-Experts model with a 1-million-token context window. Meta's chief of AI Alexandr Wang called it "the most capable US open-weight model to date," saying it excels at multi-step reasoning and complex software engineering - areas where it claims to rival or outperform leading Chinese open models.
Two Mac Studios Are All You Need
Here's the genuinely wild part: the whole thing runs locally on two Apple M5 Ultra Mac Studios using the MLX framework - a collaboration between Meta and Apple. Despite the 816B total parameter count, only 42B parameters are active during any single inference pass - that's the magic of MoE architecture. It leads open models on the MCP-Atlas benchmark and tops the open-source AI leaderboards.
The bottom line
If you've been waiting for a frontier-class open model that doesn't require a data center to run, Muse Spark Open is the most serious option yet - and the Mac compatibility makes local deployment genuinely accessible for the first time at this scale.
The agentic era needs a different CRM. That’s Attio.
Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.
| 🛠️ ChatGPT Can Now See Your Open Browser Tabs | |
A quiet upgrade with big privacy implications
OpenAI's ChatGPT Chrome extension just got a meaningful upgrade: when you tag Chrome inside ChatGPT's web or desktop app, it can now read the browser tabs you currently have open. OpenAI demoed someone asking ChatGPT to compare the content of three tabs simultaneously - the kind of research shortcut that used to require a lot of copy-pasting.
The key word here is "tag" - ChatGPT only accesses tabs when you explicitly invoke it, rather than passively scanning in the background. But it still means your browsing session is being sent to OpenAI's servers on request. For anyone doing sensitive research or working with confidential documents in the browser, that's worth thinking through carefully. The data privacy conversation around AI assistants just got a new chapter.
The bottom line
This is genuinely useful for research and comparison tasks - but before you enable it, decide whether you're comfortable with your open tabs leaving your machine every time you invoke it.
| 🔬 Anthropic Gives Safety Researchers Full Employee-Level Access | |
Independent audits with teeth - and no veto power
In a move the AI safety community has been pushing for, Anthropic is granting METR researchers employee-level access to look under the hood of its models. METR will then publish its own independent assessments of Anthropic's risks, incidents, and safety practices - and crucially, Anthropic has explicitly given up the right to block those publications.
This is a significant step beyond the usual "third-party safety partnership" announcement where a company gets to approve what gets published. The arrangement means METR can report uncomfortable findings without Claude's maker having editorial control. It's a template that safety advocates have argued every frontier lab should adopt - and one that puts pressure on competitors to follow suit.
The bottom line
By waiving veto rights over published findings, Anthropic just set a new bar for AI transparency - and every other major lab now has to explain why they haven't done the same.
| ⚡ NVIDIA's 120B Multimodal Agent Model Runs on a Single DGX Spark | |
Edge inference at frontier scale - almost
Not to be outdone in the open-model arms race, NVIDIA released Nemotron 3.5 Super VL - a 120B-parameter open multimodal MoE model with 12B active parameters and a 1-million-token context window. Like Muse Spark Open, it leans on MoE efficiency - but NVIDIA is targeting its own AI hardware ecosystem, designed to run on a single DGX Spark device.
NVIDIA also shipped two smaller "Bolt" edge models alongside it. The weights ship in NVFP4 format, a compressed precision that makes running large models on relatively constrained hardware feasible. It handles text and images, and is positioned as an AI agent model - meaning it's built to take actions, not just answer questions. Worth noting: if you're doing cost analysis on deploying models like this, the Daily Inference token calculator is a handy place to sanity-check your context window economics.
The bottom line
A frontier-scale multimodal agent model that runs on a single device is a big deal for enterprise teams that can't or won't send data to the cloud - watch for Nemotron 3.5 Super VL to show up in a lot of on-premise deployments.
| ⚠️ AI Hacks Are Causing Real Damage - and Labs Aren't Paying for It | |
The liability gap nobody wants to talk about
Wired has a sharp investigation out this week on a growing and underreported problem: AI-powered hacks are leaving real victims with real losses, and the companies whose models enabled the attacks are largely not on the hook financially. Targets of AI-assisted fraud, deepfakes, and social engineering attacks are instead being encouraged - sometimes by the same companies - to adopt those same AI tools as a defense.
The piece touches on a structural problem in AI regulation: there's no clear liability framework for when an AI model is the instrument of harm. Cybersecurity teams are increasingly dealing with AI-augmented attacks while operating under legal frameworks written for a pre-AI world. Building a site or product that needs to stay ahead of this? Tools like 60sec.site let teams spin up fast without introducing new attack surface via slow, complex build pipelines.
The bottom line
Until legislatures assign liability for AI-assisted attacks, the cost of the damage will keep landing on victims and their insurers - not the labs shipping the models. That gap is going to get harder to ignore as attacks scale.
| 🌎 Trivia Reveal | |
The answer is 42B active parameters! Despite Muse Spark Open having 816 billion total parameters, only 42 billion are actually activated during any single forward pass. That's the core efficiency trick of Mixture-of-Experts architecture - the model routes each token through only a small subset of its "expert" sub-networks, keeping compute costs manageable while maintaining the knowledge capacity of a much larger model.
| 💬 Quick Question | |
Meta's Muse Spark Open running on two Mac Studios is a big moment for local AI - but would you actually use it? Are you running models locally, sticking with cloud APIs, or somewhere in between? Hit reply and tell me your current setup - I read every single response and genuinely love hearing what people are actually building.
That's it for today - a big weekend for open-weight models, browser integrations, and safety transparency. See you tomorrow with more. And if you want to dig into any of the stories above in more depth, the full Daily Inference archive has you covered.
| 🎁 Share Daily Inference | |
Know someone who likes AI news? Share your link. You get the AI Tools Starter Kit at 1 referral and the AI Insider Briefing at 5.