ma2tic

The brief · 2 April 2026

The AI brief, 2026-04-02

Today: in-house models everywhere (Gemma 4, Foundry, MAI on the image side last week), a debate that refuses to die over deregulation in the face of China, and a whale that might soon talk back to you. A five-minute scan, coffee included.

3 min read J / K to navigate

FDA shifts the bar for priority medical AI

Previously, an AI could earn breakthrough device status just by matching a doctor's performance, only faster. STAT News reports the FDA is now favoring AI that does something no doctor can do at all, like spotting multiple cancers on a single scan or predicting mortality risk from cancer or heart failure. That changes which tools show up in clinics first.

practitioners › STAT News analysis, April 2, 2026, on FDA breakthrough device designations, featuring examples of multi-cancer detection from single scans and cancer/heart-failure mortality prediction.

STAT News →

AI Scribes Save Doctors 16 Minutes of Paperwork a Day

A JAMA study across five US health systems finally puts numbers on what physicians had been saying for months. Less time on the chart means more time for the patient, or just a less exhausting day. The open question is whether the gains hold once the novelty wears off.

practitioners › 13.4 fewer minutes on the patient chart and 16 fewer minutes of documentation per day, measured across five health systems, study published online April 1, 2026 in JAMA.

JAMA →

Snowflake can now read your PDFs right where they live

Previously, analyzing a PDF or Word file stored in Snowflake meant exporting it to another tool first. AI_COMPLETE runs multimodal analysis in place, directly on files already sitting in storage. For data teams buried in contracts, invoices, or technical diagrams, that cuts out a whole round trip.

practitioners › AI_COMPLETE document intelligence is now generally available on Snowflake, reading PDFs, Word docs, charts, and diagrams with no export required.

docs.snowflake.com →

Google Drops Gemma 4, Its Beefiest Open Models Yet

Google has released a new family of open-weight models, ranging from a small model built for on-device use to a large one designed for reasoning and agentic tasks. You can download them, fine-tune them, and run them on your own hardware or with any cloud provider you like. The Apache 2.0 license keeps the door open for commercial use without murky clauses.

practitioners › Four sizes available (E2B, E4B, 26B MoE, 31B Dense), Apache 2.0 licensed, with the 31B ranking #3 among open models worldwide on Arena AI as of April 2, 2026.

blog.google →

Microsoft ships three in-house models on Foundry

Microsoft is cutting its reliance on OpenAI alone, pushing its own models for transcription, voice, and image generation. If you're building on Azure or Foundry, you now have a native alternative for these three building blocks without routing through a third-party provider.

practitioners › MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 (already announced March 19) are now live in Microsoft Foundry and MAI Playground.

Microsoft →