Stanford HAI has published a pointed argument that 'open-weight' models — the kind Meta releases publicly — are not the same thing as 'open source,' and that the distinction matters for science and society in ways the field has been quietly glossing over. It's a clean piece of intellectual honesty in a space where 'open' has become a marketing term as much as a technical one. The argument: releasing weights without training data, architecture details, or reproducible methodology gives you a model you can run, not one you can actually audit or build on in the deep sense. That gap is where a lot of AI governance assumptions are currently parked.

Meanwhile, Google's Gemini is getting a real-world agentic test in India — purchasing from Flipkart on users' behalf — which is a more interesting signal than a benchmark. Agentic shopping is where the frictions between AI assistants and platform ecosystems get concrete fast, as Meta's ongoing data dispute with Amazon has already demonstrated. When the agent is the interface, every platform has to decide whether to cooperate or compete with it.

Meta's Muse is getting a second round of scrutiny today — its biometric keychain form factor is drawing both design criticism (Wired thinks it looks like a children's toy despite being adults-only) and genuine commercial curiosity. The product is essentially a bet that ambient AI hardware can live on your keychain rather than your face, which is either the right answer to the smart glasses adoption problem or a very expensive detour. The previous reporting that Amazon has already cut off Muse's shopping data access means Meta is building this ecosystem with real platform headwinds from day one.

AI and microscopy have combined to decode the full neural architecture of a fruit fly — not a metaphor, an actual wiring diagram of a biological brain. This is the kind of result that sits awkwardly between 'interesting neuroscience' and 'directly useful AI research,' and the honest answer is probably both. Connectome mapping at this resolution generates training data for neuromorphic architectures and tests assumptions about how biological information processing differs from the transformer stack. Whether it changes anything near-term is unclear; that it's now possible is not nothing.