October 15, 2026
TechCrunch reported that Meta, working with Apple, has launched Muse Spark Open, an open-weight model built to run locally on Apple silicon. That's a bigger deal than it sounds. Most frontier-grade models still live on someone else's servers, which means your data makes a round trip through a third party every time you use them. A model that runs natively on a Mac changes the privacy math and the latency math at once, and it signals that Meta is willing to compete on openness rather than just raw capability. For businesses, this matters less because of the specific model and more because of the precedent: open-weight, on-device AI is becoming a real alternative to API calls, not just a hobbyist curiosity. Teams building internal tools should watch this trend closely, because it changes what's possible to run securely in-house, which is exactly the kind of tradeoff we walk clients through in our own security posture discussions.
At TechCrunch Disrupt, Amazon devices chief Panos Panay laid out a vision where ambient AI anticipates what you need instead of waiting for you to open an app. It's a compelling pitch, and it's not a new one -- every major device maker wants to be the layer that makes the others invisible. But there's a real tension here that I don't think Panay fully addressed: an app you open gives you a moment to decide whether to trust it. An ambient system acting on your behalf, by definition, removes that checkpoint. For consumers, that's a convenience bet. For businesses, it's a governance problem -- who approves what an always-on agent does with company data before it does it? Until that question has a good answer, I'd treat ambient-AI hype with more skepticism than enthusiasm, even as I think the direction is probably inevitable. If you're automating workflows internally, the discipline that matters is building in approval steps and visibility, the same principle behind our own workflow automation approach -- automation should extend your judgment, not replace it silently.
The most underrated story of the week is Anthropic granting METR researchers employee-level access to look under the hood, with METR free to publish its own findings on Anthropic's risks, incidents, and practices without Anthropic holding veto power. Most companies, AI or otherwise, carefully stage what outside auditors see and retain final say over what gets published. Anthropic giving that up is a genuine, costly signal -- it only works as a trust-building move if the access is real and the findings get published even when they're unflattering. I think this should become the baseline expectation for any company claiming to build safety-critical systems, not a one-off headline. It also puts quiet pressure on competitors: if Anthropic can survive independent scrutiny, the absence of similar openness elsewhere starts to look like a choice. Business buyers evaluating AI vendors should start asking a version of this question directly -- not just "is your model good," but "who checks your homework, and do they get to publish what they find." That's the same spirit behind why we publish our own incident response practices rather than keeping them behind closed doors.
Three different stories, one shared question: as AI moves onto more devices, into more ambient layers, and under more scrutiny, who actually controls what it does and who it answers to? Meta is betting openness wins. Amazon is betting invisibility wins. Anthropic is betting that surviving outside scrutiny wins. None of these bets are wrong exactly, but they're not the same bet, and a business picking AI tools this year is implicitly picking which of those philosophies it wants running inside its company.
Which of these three approaches -- open and local, ambient and ask-nothing, or audited and transparent -- would you actually want making decisions with your company's data?
Sources