The brief · 17 April 2026

The AI brief, 2026-04-17

Today AI is learning to recognize any cell under a microscope, two new models are betting on efficiency, and regulators are taking back the reins on both sides of the Atlantic. Six stories to scan, from the lab to the central bank.

3 min read J / K to navigate

Fed leaves generative AI out of its updated model risk rulebook

This is a regulatory gap, not a green light. US banks will still need to document their credit and fraud models under the new SR 26-2, but anything involving ChatGPT or AI agents remains without written rules for now. This doesn't change much for you directly, unless you work at a bank and are waiting for an official framework to roll out an internal AI copilot.

practitioners › The Fed, OCC, and FDIC published SR 26-2 on April 17, 2026, replacing SR 11-7 after 15 years; generative AI and agents are explicitly excluded from scope as "new and rapidly evolving."

Federal Reserve →

xAI Rolls Out Grok 4.3 Beta With Native Video Input

You can now show Grok a video instead of describing it in text. It also adds document and slide generation right in the chat, pushing the tool closer to a real office assistant.

practitioners › Knowledge cutoff December 2025, limited to SuperGrok and Premium+ subscribers, improved tool-calling.

xAI →

Alibaba drops Qwen3.6-35B-A3B, a model that only activates 3B of its 35B parameters

This is a MoE model: it only lights up a small slice of its network per query, so inference costs stay low despite the total size. It codes autonomously on real-world tasks with a 73.4% score on SWE-Bench Verified, and it's open source. Worth testing if you want a coding agent without relying on a closed API.

practitioners › Weights available on Hugging Face, also accessible via Qwen Studio and API, multimodal.

qwen.ai →

Alabama Bans Health Insurers From Letting AI Make Coverage Decisions Alone

If an insurer uses AI to decide on a claim or prior authorization, a human clinician must now sign off before it's final. Alabama joins the growing list of states refusing to let coverage decisions rest solely on an algorithm.

practitioners › SB 63, signed into law April 17, 2026: mandatory human oversight, periodic accuracy audits, non-discrimination certification, no fully automated decisions allowed.

Alabama Senate Bill 63 →

Under the radar

Caltech Unveils an AI That Recognizes Any Cell Under the Microscope

Until now, every cell type needed its own segmentation model, hand-trained for one specific imaging setup. CellSAM changes that: a single model spots cells across cancer biopsies, immune cells, or other tissue types, with no retraining required each time. For a lab that can't afford to hire a data scientist, that's the difference between months of calibration and a tool that works right out of the box.

practitioners › Published in Nature Methods, free for researchers, trained on hand-annotated data.

Nature Methods →

Rights bodies warn EU: don't gut the AI Act in the name of simplification

The Digital Omnibus is pitched as a simplification of the AI Act and GDPR, but ENNHRI and Equinet worry it's being used as cover to water down protections. If you're deploying high-risk AI in Europe, registration deadlines and supervisory authority powers could stay as they are or get watered down, either way your compliance workload is on the line.

practitioners › Joint statement issued April 17, 2026 by ENNHRI (human rights) and Equinet (equality bodies), calling to preserve registration duties, high-risk AI deadlines, and GDPR safeguards.

ennhri.org →