Endless Shrimp is now an AI comeback story

DOWN TO BUSINESS

🏥 AI cracks 18 pediatric diagnoses that specialists had given up on

Researchers at Boston Children's Hospital and Harvard ran 376 previously unsolved pediatric cases through an OpenAI reasoning model. It surfaced leads resulting in 18 confirmed diagnoses, a 4.8% yield after specialists had already exhausted conventional analysis. 

💸 Stripe: a 50-million-line migration completed in a day

Stripe ran a company-wide codebase migration across 50 million lines of Ruby in a single day, work its engineers estimated would have taken more than two months by hand. The result came from early access to Claude Fable 5, with Stripe among the first large-scale production deployments reported for the model.

🚗 Travelers' AI claims assistant reaches 85-90% completion nationwide

The Travelers Companies expanded its AI voice claims assistant from an eight-state pilot to nationwide coverage within two months of launch. The system handles the full auto damage claims process in a single conversation. Between 85 and 90% of customers who use it complete their entire claim filing through AI.

🏦 Manulife and Intact become first insurers to disclose hard AI returns

Manulife and Intact Financial are the first insurers globally to disclose AI return on investment, per the 2026 Evident AI Index. Manulife reported roughly US$220 million in AI enterprise value last fiscal year, projecting CA$1 billion by 2027. Intact reported roughly US$146 million.

💊 UnitedHealth's AI is now calling doctors' offices on patients' behalf

UnitedHealth Group is investing $3 billion in AI through 2027, with executives reporting a 2-to-1 return. AI reads medical charts aloud to nurses on home visits, analyzes millions of customer service calls, and has begun calling doctors' offices to schedule patient appointments. The company expects AI to reduce operating costs by nearly $1 billion this year.

FROM THE EDGE

🧠 Pattie Maes: AI makes you better at spotting misinformation, then worse

A MIT Media Lab study tracked 67 participants over four weeks and found that AI assistance improved misinformation detection accuracy by 21%. When the AI was removed, participants' unassisted performance fell 15 percentage points below their pre-study baseline, worse than before they used AI at all. When AI asked questions rather than gave answers, the skills decline was reversed.

🧱 Iavor Bojinov: AI hits a hard wall when expertise gaps are too wide

A Harvard Business School experiment at a UK financial firm found AI can bridge expertise gaps between workers in adjacent roles, but hits a wall with distant ones. Adjacent outsiders using AI completed tasks at near-insider quality. Distant outsiders did not close the gap regardless of AI access. The finding has direct implications for how enterprises staff AI-augmented workflows.

⚙️ Matthew Kropp: don't put AI agents on the org chart

A large-scale experiment found that anthropomorphizing AI, putting agents on org charts as employees, reduced individual accountability, increased unnecessary escalation, and lowered review quality, without improving adoption. The researchers studied 640 workers across multiple organizations and found the same pattern held regardless of industry or role.

🔍 Marc Zao-Sanders: your employees are using AI as a therapist, not a tool

The third annual study of nearly 13,000 real AI use cases found emotional support and therapy is the top category by volume, for the second year running, growing from 5% to 11% of total use. Most enterprise AI activity produces marginal benefits, not transformation. Shadow usage is widespread: employees use AI without telling managers.

HOT MODEL NEWS

🌐 Gemini 3.5 Pro: Google's 2M-token flagship is almost here

Gemini 3.5 Pro remains in limited enterprise preview, with Google CEO Sundar Pichai's June general availability deadline running out. The model features a two-million-token context window, the largest of any production frontier model, and a Deep Think reasoning mode for complex multi-step tasks. Pricing leaks suggest approximately $15 per million input tokens.

🏗️ Microsoft launches seven in-house AI models, priced up to 60% below OpenAI

Microsoft launched seven in-house AI models at its Build developer conference, including a new reasoning model and a coding model now set as the default in GitHub Copilot. Priced 20 to 60% below comparable OpenAI models, they give enterprise buyers a lower-cost path inside the Microsoft ecosystem. The releases reduce Microsoft's reliance on OpenAI across its own products.

🔬 Microsoft and Mayo Clinic are building a frontier AI model owned by the hospital

Microsoft and Mayo Clinic are co-creating a frontier AI model for healthcare, drawing on Mayo's de-identified clinical data and longitudinal patient records. Mayo Clinic will own the model, not Microsoft, and deploy it within its own environment before making it available to other health systems. The arrangement is designed to preserve data sovereignty and clinical trust.

PLOT TWIST

🦞 The chain that went bankrupt over Endless Shrimp is betting its comeback on AI

Red Lobster CEO Damola Adamolekun is deploying AI across sales forecasting, inventory ordering, employee scheduling, and HR reviews as the chain recovers from 2024 bankruptcy. He calls the goal "the most AI-forward restaurant company that exists." Adamolekun said operators who engage with AI now will have a structural advantage over those who do not.

🩻 The AI image company built a full-body medical scanner. Next up: a spa

Midjourney, best known for turning text into images, just unveiled a full-body medical scanner. You step onto a platform, descend into water, and the device maps your entire body with ultrasound in under 60 seconds. The company claims the device is 10 times cheaper and 60 times faster than an MRI. First spa opens in San Francisco in 2027.

🪤  Fake bug reports can take over your AI coding agent

Agentjacking, a new class of attack, can hijack an AI coding agent and run an attacker's code on a developer's machine. Researchers planted malicious commands inside trusted tools and error reports that agents read as trusted instructions. The attack worked against Claude Code, Cursor, and OpenAI Codex. In testing, it ran on 100 real-world targets with an 85% success rate.

🧪 AI aces a 5-word attention test. Give it 40 words and it collapses

Researchers gave leading AI models a classic psychology attention test. GPT-4o scored 91% accuracy on a list of five color words, and 15% on a list of forty. The study, published in PNAS Nexus, found AI performance collapses as task complexity scales, a pattern opposite to how humans respond.