Friday, July 24, 2026

ChatGPT Gets Your Health Context

ChatGPT Gets Your Health Context

Today’s Overview

Good morning, AI is getting more personal and more agentic at the same time. OpenAI is bringing connected health data into ChatGPT, Sierra is buying long-horizon agent tech, and voice assistants are moving from chatty demos into desktop work. There is also a packed research slate on education, deep research agents, robot tracking, and spatial reasoning. Let's dive in.

Top Stories

OpenAI launches Health in ChatGPT

OpenAI launched Health in ChatGPT for U.S. adults, letting users connect Apple Health and supported medical records for more contextual health conversations. The company said connected health data and related chats will not be used to train foundation models or target ads.

  • The rollout starts with logged-in U.S. users who are 18 and older and spans web and iOS across Free, Go, Plus, and Pro plans.
  • OpenAI says users can connect sources including Apple Health, U.S. medical records, One Medical, and Function Health to ground answers in their own information.
  • By default, ChatGPT asks before using connected health data, and users can disconnect accounts anytime with synced source data deleted from OpenAI systems within 30 days.

Sierra acquires Takeoff for long-horizon agents

Sierra acquired Takeoff, a long-horizon AI agent platform. The deal gives Sierra immediate access to technology for autonomous, multistep enterprise applications.

  • The acquisition was publicly announced on July 23, 2026 by Bret Taylor, who said the Takeoff team would join Sierra.
  • The article says no financial terms valuation, closing date, separate press release, or investors were disclosed.
  • Takeoff is described as a platform built in less than a year as Sierra moves to strengthen its agent stack.

OpenAI and Anthropic race on voice AI

OpenAI and Anthropic shipped voice upgrades at nearly the same time, underscoring voice as a key battleground in how people work with AI. OpenAI moved ChatGPT Voice into the desktop app and Codex, while Anthropic expanded Claude voice to stronger models and connected apps.

  • OpenAI's upgrade brings ChatGPT Voice to macOS and Windows desktop apps for paid plans.
  • The OpenAI flow is built around GPT-Live, which can listen, think, and reply at the same time instead of forcing a slower turn-by-turn cadence.
  • Anthropic's side of the race now lets Claude voice use Opus and Sonnet and connect to tools such as Gmail and Slack.

Research & Analysis

K12-KGraph targets curriculum cognition in educational LLMs

K12-KGraph is a curriculum-aligned knowledge graph extracted from official People's Education Press textbooks in mathematics, physics, chemistry, and biology across primary, middle, and high school. The authors use it to build K12-Bench, a 23,640-question multi-select benchmark, and K12-Train, a 7,335-sample supervised fine-tuning corpus for text and multimodal tasks.

  • The work frames the target capability as curriculum cognition rather than simple exam-question answering.
  • Its graph covers nine node types and fourteen relation types for curriculum structure and visual grounding.
  • The benchmark spans five task families: Ground, Prereq, Neighbor, Evidence, and Locate to test different forms of curriculum understanding.

AREX explores self-improving deep research agents

AREX is a family of recursively self-improving deep research agents that alternates between evidence gathering and constraint-wise self-auditing. The system is trained through agentic mid-training and long-horizon reinforcement learning, with models reported at 4B dense and 122B-A10B Mixture-of-Experts scales.

  • The paper builds on a discovery-verification asymmetry where finding answers is costly but checking candidate answers can be decomposed into smaller constraints.
  • AREX uses an autonomous context-update tool to compress long interaction histories into a compact improvement state without relying on an external model.
  • Training emphasizes decisive moments where evidence is acquired or bad research directions are corrected to reduce the weakness of sparse final rewards.

ReferTrack grounds robot tracking in visual references

ReferTrack is a referring-then-tracking system for embodied visual tracking, where a mobile agent follows a natural-language target using only onboard vision. It first selects the target from indexed bounding boxes, then predicts tracking waypoints conditioned on that visual decision.

  • The method is designed for a single forward-facing camera instead of relying on a multi-camera setup.
  • To preserve target motion over time, ReferTrack keeps a sliding-window queue of previously selected bounding boxes.
  • The authors say code, data, and checkpoints will be released on GitHub for others to try.

ProVisE tests spatial reasoning in pixels

ProVisE, short for Protocolized Visual Evaluation, is a framework for evaluating spatial cognition in image-generation models. The paper argues that some spatial tasks are better answered directly in pixel space than through coordinates, options, or text.

  • The paper's core critique is an answer-interface mismatch between image-generation models and text-heavy spatial benchmarks.
  • SpatialGen-Bench contains 470 samples across 14 spatial subtasks, four capability levels, and diverse answer forms.
  • ProVisE includes an Agentic builder that constructs and validates task-specific protocols for new benchmarks.

Trending AI Tools

  • Grok Build Workflows Runs complex coding tasks across hundreds of background agents with orchestration scripts and skeptic verification checks.

  • Claude Voice Mode Adds Opus and Sonnet support, connected tool access, more languages, and public beta availability across mobile, desktop, and web.

  • ChatGPT Voice on desktop Lets paid users direct ChatGPT and Codex agents by voice in the macOS and Windows desktop apps.

Quick Hits

  • Microsoft MAI models bring MAI-Image-2.5-Pro for image generation and editing plus MAI-Voice-2-Flash for faster, lower-cost voice applications in public preview.

  • Kimi K3 copying claim centers on Michael Kratsios accusing Moonshot AI of using distillation from Anthropic's Fable 5 as Moonshot prepares to publish K3 weights on July 27.

  • Jacob Tsimerman joins OpenAI to work on AI safety, with the move coming the same day he was awarded the 2026 Fields Medal.

  • FLUX 3 is Black Forest Labs' new model positioned for image, video, audio, and robotics use cases.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.