🚀 RELEASES & LAUNCHES
Instinct gives each AI assistant its own email address
Instinct gave each assistant a dedicated email address for autonomous account and support tasks.- Instinct’s invite-only personal-assistant users can claim dedicated addresses at mail.instinct.com.
- The feature lets Instinct create and manage accounts, contact businesses, handle support requests, and process forwarded emails.
- Instinct can act autonomously on email threads and check back when it needs user input.
- The separate inbox reduces users’ need to share personal email addresses or intervene in agent-run tasks.
IBM releases 385M-parameter Granite PatchTST-FM-r2 forecasting model
IBM released Granite PatchTST-FM-r2, an open zero-shot time-series forecasting model.- It targets developers and enterprises building time-series forecasting and imputation systems.
- The model has roughly 385 million parameters, supports contexts up to 8,192 steps, and predicts 99 quantiles without fine-tuning.
- PatchTST-FM-r2 uses conformer blocks and 50%-overlapping patches with a 16-step patch length and 8-step stride.
- IBM reports GIFT-Eval geometric-mean scores of 0.467 CRPS and 0.6846 MASE, ranking second among replicable zero-shot models.
- Weights, code, architecture, and inference tools were released under Apache 2.0 and OpenMDW 1.0 licenses.
Apple redesigns Health app with AI-powered Health Age and readiness score
Apple announced an Apple Intelligence-powered redesign of its Health app.- For Apple Intelligence-compatible iPhone and iPad users, with Health Age requiring an Apple Watch.
- The redesigned app arrives later this year with Insights and Longevity tabs, Health Age, personalized recommendations, and a 0–10 readiness score.
- Users can add blood-test data through a U.S. Quest Diagnostics panel covering more than 50 biomarkers for $119.
- The app uses Apple Watch, iPhone, clinical records, and compatible third-party data to turn health metrics into personalized guidance while keeping assessments on-device where reported.
Apple adds Reference Image to verify iPhone 18 Pro photos
Apple introduced Reference Image to help verify whether iPhone 18 Pro photos were edited.- The opt-in feature targets photojournalists, photographers, and everyday viewers.
- The iPhone 18 Pro sensor captures signed data that Private Cloud Compute turns into an unalterable reference image.
- Photos displays the reference image beside the original as a digital negative for comparison.
- Apple is providing APIs for third-party apps and adding SynthID support for AI-generated or edited images.
- The feature gives builders a way to expose capture and editing provenance, rather than prove a scene’s context or truth.
Shipt launches Ask Shipt AI assistant for prompt- and photo-based carts
Shipt launched Ask Shipt, an AI assistant that turns prompts and photos into grocery carts.- Target-owned Shipt customers can use Ask Shipt in the app and on Shipt.com across more than 100 retailers.
- The assistant builds carts from text requests, meal photos, budgets and occasions, while surfacing store promotions.
- Customers can review, modify or remove recommendations before checkout, adding conversational cart-building to Shipt’s marketplace.
Mentria runs a 27B one-bit model in Chrome at up to 30 tok/s
Mentria ran a 27B one-bit model in Chrome at up to 30 tokens/s on a 6 GB RTX 3060.- The browser engine targets users seeking local, private LLM inference without an installation or server.
- Mentria repacked Prism ML’s Bonsai-27B model for WebGPU/WGSL inference in Chrome.
- The model uses about 1.14 bits per parameter and fits 27 billion parameters into 3.8 GB of GPU memory.
- The reported benchmark reached 25–30 tokens/s on an RTX 3060 Laptop GPU with 6 GB of VRAM.
- Browser-based one-bit inference enables larger local models on consumer GPUs while keeping data on-device.
🔬 RESEARCH
- Google DeepMind maps 9 billion human DNA variants with AlphaGenome Atlas
Google DeepMind released AlphaGenome Atlas, a searchable map of 9 billion human DNA variants.- Academic researchers, clinical researchers and biologists can access the Atlas through a no-code portal, API and Google Antigravity skill.
- The resource precomputes AlphaGenome predictions for every possible single-letter human DNA change in a roughly 1-petabyte dataset.
- Its AlphaGenome Variant Impact score combines AlphaGenome and AlphaMissense predictions to rank variants across coding and non-coding regions.
- The Atlas lets researchers look up and prioritize variants before laboratory testing; collaborators reported 22% more non-coding genetic associations in a 54,000-plus-person UK Biobank analysis.
Sources 📎 Coverage: arstechnica.com · 👁 via Ars Technica AI
🏛️ POLICY & REGULATION
- Microsoft and AFT agree to binding AI privacy rules for US schools
Microsoft and AFT agreed to binding AI safety and privacy rules for US schools.- The agreement is for US school districts using Microsoft AI tools, negotiated with the American Federation of Teachers and its New York City affiliate.
- It includes 10 contractually enforceable principles, including no model training on student or educator data, limited data collection, family disclosures, no AI companions, and human review for high-risk decisions.
- Districts can add the terms to new or existing Microsoft agreements beginning in November without renegotiating their full contracts.
- The standard gives schools enforceable controls over AI data use, oversight, and safety requirements.
Sources 📎 Coverage: theverge.com · 👁 via The Verge AI