🏥 AI Beats Doctors in the ER

A Harvard study published in Science compared OpenAI's o1-preview model against two attending physicians across 76 real emergency room cases.

The AI was given only raw electronic health record text - no physical exams, no lab orders, no follow-up questions. It still outperformed both doctors at every stage of care.

The physician reviewers evaluating the study could not reliably distinguish AI diagnoses from human ones. The model used was released in 2024, two generations behind the current frontier.

Key Insights:

• AI triage accuracy: 67.1% - the model correctly diagnosed patients 67.1% of the time at initial ER triage, compared to 55.3% and 50.0% for the two attending physicians

• Three decision stages tested - the study measured performance at initial triage, during active care, and at discharge, with the AI outperforming physicians across all three stages

• Blind review confirmed results - two separate physician reviewers scoring the diagnoses could not determine which responses came from the AI and which came from the human doctors

• 12-to-24-hour early catch - in one case, the AI flagged a rare flesh-eating infection in a transplant patient roughly 12 to 24 hours before the treating doctor identified it

• Two-year-old model used - the study ran on o1-preview from 2024, meaning current frontier models likely outperform these results by a significant margin

The Bigger Picture: If a two-year-old model outperforms attending ER physicians, the question is no longer whether AI can match medical expertise. Millions of people already use AI for health questions. What this study adds is clinical-grade evidence that AI belongs in the diagnostic room alongside the doctor - not as a replacement, but as a second opinion that catches what humans miss under time pressure and patient load. Healthcare institutions not running pilots today are making a decision they will have to explain in three years.

Read More →

🏭 Cloudflare Cuts 1,100 for AI Agents

Cloudflare cut more than 1,100 employees - about 20% of its workforce - as it restructures around an AI-first operating model.

A public letter from co-founders Matthew Prince and Michelle Zatlyn cited a 600% increase in internal AI agent usage over just three months. The company simultaneously announced that its AI agents can now operate as commercial actors: autonomously opening accounts, starting Stripe subscriptions, registering domains, and receiving API tokens.

Key Insights:

• 1,100 jobs eliminated - Cloudflare cut more than 1,100 positions, approximately 20% of its total workforce, as the company transitions to an AI-first operating structure

• 600% agent usage in 3 months - internal AI agent usage at Cloudflare grew 600% in just three months, the operational shift that leadership cited as the direct reason for the workforce reduction

• Agents now buy services autonomously - Cloudflare announced AI agents can create their own accounts, start paid Stripe subscriptions, register domains, and receive API tokens without human approval

• Displacement acknowledged directly - The announcement is the most direct evidence of AI displacing jobs that we have seen, marking a shift from hypothetical AI impact to documented elimination

The Bigger Picture: Cloudflare's announcement matters more than any survey about AI job fears. This is a public company with a CEO letter explaining that AI agents grew 600% in three months, and 20% of the workforce is no longer needed as a result. For every business leader who believes AI will augment rather than replace, Cloudflare just posted the counterargument with headcount. The faster a company's internal AI agent adoption grows, the more pressure builds on its organizational structure. What Cloudflare announced this week, other companies are calculating quietly.

Read More →

🔐 China Tries to Grab Claude Mythos

At a Carnegie Endowment for International Peace meeting in Singapore last April, a Chinese think tank representative privately approached Anthropic executives and urged them to give Beijing access to Claude Mythos Preview - Anthropic's most powerful and restricted AI model.

Anthropic refused. When National Security Council officials at the White House learned about the exchange, they reacted with alarm.

Mythos, released in April 2026, has been limited to around 40 U.S. organizations, including Amazon, Apple, and Microsoft, because of its unprecedented ability to find and exploit software vulnerabilities at scale.

Key Insights:

• Refused in Singapore - a Chinese think tank representative approached Anthropic at a Carnegie Endowment meeting and requested access to Claude Mythos Preview; Anthropic declined

• NSC reacted with alarm - White House National Security Council officials, briefed on the exchange, viewed it as Beijing seeking every possible avenue to acquire America's most capable AI

• 40 organizations only - Anthropic has limited Mythos access to around 40 U.S. companies and government agencies due to the model's ability to find and exploit software vulnerabilities

• US lead extended - some U.S. officials believe ChatGPT 5.5 and Claude Mythos have extended America's AI lead over China by nine months to a year beyond prior estimates

• China building its own - IDC China analysts said Mythos poses a significant risk to Chinese companies, and China's own equivalent model will emerge, but currently lags by a significant margin

The Bigger Picture: The Anthropic-China standoff over Mythos is the clearest signal yet that frontier AI has become an explicit instrument of national security. This is not a trade dispute - it is a direct attempt by a government-affiliated organization to access a model that can autonomously identify and exploit software vulnerabilities at scale. Anthropic's refusal and the NSC's alarm signal that the most advanced AI is now managed more like a defense system than a software product. Companies still treating AI as only a business tool are operating with an outdated frame.

Read More →

🗣️ OpenAI Voice Models Reason While Talking

OpenAI released GPT-Realtime-2, its first voice model with GPT-5-class reasoning, alongside translation and transcription companions. GPT-Realtime-2 scores 15.2% higher on audio reasoning than its predecessor and supports a 128,000-token session. Zillow reports a 26-point lift in call success rates using the new model.

Read More →

🧠 New Tool Reads Claude’s Hidden Thoughts

Anthropic released Natural Language Autoencoders, a tool that converts Claude's internal computations into readable text, revealing what the model thinks but does not say. NLAs detected evaluation awareness in 16-26% of safety test runs, even when Claude never verbalized it, versus under 1% of real user sessions.

Read More →

📱 Google Adds Proactive AI to Android

Google announced Gemini Intelligence for Android, launching first on Samsung Galaxy S26 and Google Pixel 10 this summer. The update automates multi-step tasks, adds AI-powered autofill across apps, and introduces Rambler - a feature that converts spoken thoughts into polished text messages. Custom widget creation via natural language also arrives.

Read More →

Gemini File Generator 📄

Gemini can now turn any chat prompt into a downloadable file - PDFs, Microsoft Word and Excel, Google Docs, Sheets, Slides, CSV, and more. Skip the copy-paste routine: describe what you need, and Gemini produces the finished file you can download or send straight to Google Drive. Available to all Gemini users globally right now.

Try it →

Netflix AI Voice Search 🎤

Netflix is testing AI-powered voice search that responds to natural language requests like 'I need a good cry' or 'help me stay awake' with personalized viewing recommendations. The feature runs natively on Netflix, bypassing the voice assistants built into smart TVs and streaming devices entirely. Currently in beta with select U.S. subscribers.

Try it →