Amazon Wants to Be Your Doctor Now
Amazon launched a Health AI agent this week. Prime members get 24/7 access to health guidance through One Medical, and the thing can interpret lab results, manage prescriptions, book appointments, and handle over 30 common conditions with a real provider in the loop.
This matters because healthcare has been stuck in fax-machine mode for about forty years. If anybody can drag it into 2026, it's the company that figured out how to get toilet paper to your door in two hours.
My take? I'm cautiously into it. Getting a straight answer about a lab result without spending three days on hold is a real benefit for regular folks. But I'll tell you what, the minute this thing starts upselling me on Prime Pharmacy deliveries in the middle of a medical conversation, I'm out.
Grok 4.20 Tries to Stop Making Stuff Up
xAI dropped Grok 4.20 and it's all about factuality. The model scored highest among March frontier releases on benchmarks measuring accuracy on news and events from the prior 30 days.
Current-events accuracy has been the soft underbelly of every big model. You ask about something that happened last Tuesday and you get confident nonsense. Closing that gap is not glamorous work, but it's the work that actually makes these tools usable.
Honestly, if every lab stopped chasing bigger numbers on math benchmarks and just made their models stop hallucinating about last week, the whole field would be better off. Grok is pointed in the right direction here, and credit where it's due.
Google Figures Out How to Cram More in Less
Google showed off TurboQuant at ICLR 2026. It's an algorithm that slashes the memory overhead of the KV cache using vector rotation and some Johnson-Lindenstrauss compression math.
Translation for the rest of us: longer context windows, less memory burned, faster responses. That's the holy trinity of inference and Google just pushed it a little further down the road.
Look, this is the kind of boring infrastructure work that doesn't make headlines but quietly makes everything else possible. Every time some team figures out how to fit more model into less hardware, we all win. Keep it coming.
The Through Line
Notice what's not on this list? Nobody announced a new AGI. Nobody claimed their model achieved consciousness. Nobody shipped a demo video that'll never become a real product.
This was a week of grown-up AI news. Health tools that do real chores. Models that stop lying about the news. Infrastructure that makes everything cheaper. That's what progress actually looks like when the cameras aren't rolling.
More of this, please.