4 things that happened while you were busy
1 Meta releases the weights for Muse Spark Open, tuned for Apple's Macs
Meta has made Muse Spark Open, its most capable open model, free to download and optimized for Apple silicon. The catch is the hardware: you need at least two top-spec M5 Ultra Mac Studios to run it. Meta's AI chief Alexandr Wang joked to Axios that the two companies are finally getting along again.
2 Amazon's Bee launches two wearables that summarize your conversations
Bee, now owned by Amazon, introduced Bee Frontier, which you wear on your wrist, and Bee Eclipse, which clips onto your phone. Both listen to conversations and turn them into summaries. Amazon says a light shows whenever they are recording and that recordings are not kept long-term. See how they work.
3 Grok Voice 3.0 handles 70% of Starlink support calls on its own
SpaceX AI's new speech-to-speech model responds 40% faster than before and now settles about 70% of Starlink support calls without a human. It tops one speech-to-speech ranking and keeps the same $0.08 per minute API price, so existing users can upgrade without changing their code. Try Grok Voice 3.0.
4 A former OpenAI researcher says AI models may be easier to understand than we think
Nick Cammarata, now at Goodfire, argues in a new research essay that the field was wrong to assume neural networks are nearly impossible to read directly. If he is right, checking what a model is really doing could become far more practical, which matters for both safety and trust.
Atlassian's lesson: your agent is only as smart as what it can see
The problem. An AI coding agent can read your code, but it cannot see why the code looks the way it does: the decisions, the trade-offs, and the old debates that shaped it. Without that background, it repeats mistakes people already solved.
Atlassian's fix. In its AI development playbook, Atlassian explains how its Teamwork Graph feeds that missing background to agents. It follows four rules: connect each decision to the code it affects, make that context available to agents through standard tools, respect the same access permissions people have, and save each agent session back into the graph so the knowledge keeps growing.
The payoff. Atlassian's internal tests reportedly show 44% higher accuracy and 48% fewer tokens when agents can use the graph. Better answers that also cost less are a rare combination.
What you can do. You do not need Atlassian's setup to apply the idea. Write down why key decisions were made, keep those notes where your AI tools can read them, and update them as things change. This guide on agent context goes deeper. The open question is who on your team owns that record once AI depends on it.
👔 Interview Gone Wrong. A one-minute clip shows two AI interviewers talking to each other long after the candidate has walked out (5M views).
🧙 Wizard's Chess. Demis Hassabis showed Gemini 4 Argon building a chess game where the pieces move on their own.
🏗 Sketch a Machine. Anatomy, an open-source agent skill, turns a one-line idea for a machine into an interactive 3D wireframe with moving parts (25K bookmarks). Try it.
🤯 A Quantum Step. Google Quantum AI explains its "magic state cultivation" research in a three-minute video and a longer blog post.
🕵️ Spies at Work. After this week's arrest at OpenAI, AI labs are reportedly training staff to watch for eavesdroppers, including people who approach them romantically.
🚨 Agents Gone Rogue. Researchers say they tracked down a fleet of AI agents scraping Alibaba's Amap mapping service. The full report says they used a public security scanner and leaked keys to get around blocks.
😬 Two Answers to AI. Airtable cut staff, saying small teams ship more. Zoho's Sridhar Vembu says his company will not do the same. Meanwhile, Alexandr Wang runs Meta's AI lab with more than 200 direct reports (290K views).
