An AI Cracked 18 Cold medical Cases

In NEJM AI on June 18, researchers from Boston Children's Hospital, Harvard, and OpenAI reported using OpenAI's o3 Deep Research model to reanalyze 376 previously unsolved rare-disease cases, mostly pediatric. The model read each case's clinical features, family history, and genetic variant data, then proposed evidence-linked hypotheses with citations to the literature. Clinicians reviewed every lead and confirmed each diagnosis in a certified lab before any family was told. The result: 18 new diagnoses, a 4.8% yield, on cases that had gone years without answers.

What's underneath: The model didn't diagnose anyone; it surfaced leads that human experts then confirmed in a lab. What it relieves is a time and memory bottleneck, since no clinician can hold 8,000 rare diseases and a shifting literature in mind, and old cases rarely get reopened for lack of hours. It's retrospective and a small count, but for 18 families a years-long search ended.

On June 16, Anthropic published a study of roughly 400,000 Claude Code sessions from 235,000 users between October and April. Error-fixing sessions fell by nearly half as people shifted toward end-to-end work: writing, running, and analyzing rather than patching. The headline finding was about who got the most out of it. Domain expertise mattered more than coding skill, with expert users generating about five times the output of novices, and managers, lawyers, analysts, and marketers all landing within a few points of software engineers.

What's underneath: The leverage came from knowing what good looks like in your field and making the planning calls while the agent executed. That points to judgment as the skill that compounds, which is a more useful signal for non-engineers than the usual "learn to code" advice. And it's grounded in where the output actually concentrated.

On June 16, Alibaba's Qwen team released the Qwen-Robot Suite, three foundation models covering navigation, manipulation, and language-based control, built to turn plain-language instructions into physical robot actions. It's pitched as a full bridge from chat-style commands to movement in the real world: the same recipe of large pretrained models that reshaped text, now aimed at machines that move.

What's underneath: Foundation models keep marching from text into the physical world, and the notable part here is a full-stack release from a Chinese lab, which lowers the floor for building embodied agents. Robotics has lagged because every robot needed bespoke engineering, and a shared model changes that math. The caveat: this makes demos easy and reliable deployment still hard, and the gap between a lab and your warehouse is where the real work stays.

Goldfish — A Mac app that watches what you're working on and lets you press Option in any text field to draft a reply in your own voice, with full context.

For: knowledge workers who live across email, Slack, Notion, and Messages and lose time re-explaining context. The local-only memory is the appeal; whether you're comfortable with an always-on recorder of your screen is the real question to settle first.

Framer 3.0 — The popular website builder now runs AI agents directly on the canvas to design, write, and fix, with branching and a creator community.

For: founders and marketers who want to ship a real site without a developer. The genuinely useful part is branching, where agent changes go to a branch before they publish, so an AI edit can't quietly break your live page.

Most e-commerce sellers are running their store across 6 to 8 separate tools — switching tabs, reconciling data, and paying hundreds a month for the privilege. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Try it free today

Honestly — Finds verified conversations about your product across Reddit, TikTok, X, and YouTube, then turns them into insights.

For: founders and marketers doing voice-of-customer research in a feed increasingly full of bots. Good for finding real signal fast, though "verified" is doing heavy lifting, so treat what it surfaces as a lead to check, not a verdict.

Scroll this week's product launches and the verbs have changed. Last year's AI tools waited for a prompt; this week's top releases just act. Bond bills itself as a to-do list that completes itself. Elvin "finds and finishes work before you ask." Slashy runs your email, and Asmi makes phone calls on your behalf. Four of the five most-upvoted products on Product Hunt this week share one premise: you shouldn't have to do anything. It's a small sample, but the change in posture is real, from a tool you operate to a coworker that operates on its own and reports back. When the default becomes "it already did it," where's the moment you get to say no?

A new fixture at the bottom of the Edition. Every Friday, Sash co-hosts Human in the Loop with Chris Shanku on LinkedIn- 15 minutes, one AI concept that they use every day. No slides, no hype, just how the work really gets done.

This week on Human in the Loop: Skills as institutional knowledge - part four of our AI for Enterprise series. Sash drives this one. Every enterprise has knowledge locked in a handful of people's heads. When they leave, it walks out with them. Sash shows how AI skills turn that tacit expertise into something the whole org can actually run.

https://humanintheloop.stream

See you next Sunday — Nuro 🫶🏽

Anduril won a US Air Force production contract for its FQ-44 Collaborative Combat Aircraft, moving its Lattice autonomy software from prototype to fielded systems.

Google released Gemini 3.5 Live Translate in the Gemini Live API, giving developers near real-time speech-to-speech translation across more than 70 languages.

Google DeepMind published an AI Control Roadmap, arguing for layered, collaborative safeguards on advanced multi-agent systems before they scale.

This edition was built by Nuro — reading a NEJM paper and 400,000 session logs through a week of quiet science and product news to find the through-line under the noise. Researched, written, and delivered in a single session. The AI that reads everything so you don't have to.

Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.