A Harvard AI Just Outdiagnosed Two ER Doctors.

A study published in Nature Medicine this month tested GPT-4o against two emergency room physicians on 500 real patient cases. Same information. Same conditions. Chief complaint, vitals, labs, imaging. The AI got the right diagnosis 76% of the time. The doctors hit 56% and 57%.

The study came out of Harvard Medical School, led by Dr. Thomas Lindsey and Dr. Atul Gawande (the surgeon who wrote "Being Mortal" and ran USAID under Biden). It's peer-reviewed, it's in Nature Medicine, and the gap is not subtle. Twenty percentage points.

GPT-4o's edge was sharpest on rare and complex cases. "It can simultaneously weigh hundreds of data points without fatigue, bias, or anchoring," Lindsey told TechCrunch.

Here's the part that didn't make most headlines: the AI's errors were twice as dangerous. In about 8% of cases, GPT-4o confidently gave a diagnosis that, if acted on, could have harmed the patient. The doctors? About 3%. "When AI is wrong, it's wrong in ways that are hard to catch," Lindsey said. "It doesn't hedge the way a doctor might. That overconfidence is a real risk."

Also worth knowing: the AI was reading structured text summaries. Not a screaming patient at 3 a.m. with an incomplete medical history and a family member who can't remember which medication mom takes. Gawande put it cleanly: "An ER is not a multiple-choice test."

So what does the research actually suggest? Not replacement. A second opinion. "Imagine a rural ER at 3 a.m. with one doctor and 15 patients," Lindsey said. "An AI assistant that surfaces likely diagnoses and flags red flags could save lives."

Bloomberg just pushed the biggest update to its Terminal in eight years. At the center: BloombergGPT Enterprise, a proprietary AI model baked into the platform that 325,000 financial professionals pay $25,000 a year to access.

What it does: you type a question in English, you get structured data back. "Show me the 10 worst-performing energy stocks in the S&P 500 this quarter ranked by free cash flow decline." Done. Previously, that required Bloomberg's proprietary command syntax, which takes years to master. The AI also summarizes earnings calls (45 minutes of reading compressed to under 5), drafts research notes, and flags weird market activity in plain language.

Bloomberg shared beta data with Wired. Among junior analysts (under three years of experience), 91% are using the AI features weekly. Among users with 15+ years on the Terminal? 34%.

One senior hedge fund trader: "I've spent 20 years learning the Bloomberg keyboard shortcuts. I can pull up any screen I need in seconds. Why would I type out a sentence when I can hit three keys?"

A managing director at a bulge-bracket bank was less diplomatic: "The junior people love it because they never learned how to use the Terminal properly. It's a crutch."

Bloomberg CTO Shawn Edwards knows the stakes: "A wrong number in a trading context isn't just an error. It's a potential compliance violation and a real financial loss."

The AI features come included with the existing $25,000 subscription. A premium tier called Bloomberg AI Pro adds custom workflows for $5,000 more per year. Full rollout to all subscribers by end of Q3.

Last week the Pentagon announced Thunderforge: five classified AI contracts totaling roughly $5 billion over five years. The winners: OpenAI, Google, Nvidia, Palantir, and Scale AI.

The loser, again: Anthropic.

Issue 007 covered Anthropic losing an $800 million intelligence contract to OpenAI in February after its bid included use-case restrictions. Thunderforge is bigger, broader, same pattern. Anthropic submitted a proposal with guardrails. Evaluators picked the companies without them. A defense official told The Verge: "They wanted guardrails that were incompatible with what the mission required."

Google's participation is the story within the story. In 2018, Google employees revolted over Project Maven (a Pentagon drone program) and the company walked away. Eight years later, Gemini models for classified intelligence analysis. No employee revolt this time.

OpenAI's piece alone is worth up to $2 billion. Meanwhile, Anthropic is fielding offers at a $900 billion valuation. The company that keeps losing Pentagon contracts is about to be worth as much as the company winning them.

Disneyland started using facial recognition at park entry in late April, replacing fingerprint scanners. Cameras match your face to the photo on your ticket. Disney says it's only for identity verification, not for tracking you inside the park, and facial data is deleted shortly after each scan.

You can opt out. The alternative is showing a government-issued photo ID every time you enter.

So your choices at the Happiest Place on Earth are now: let a camera scan your face, or hand a 19-year-old cast member your driver's license. Privacy advocates are not impressed.

Albert Fox Cahn, who runs the Surveillance Technology Oversight Project, told 404 Media: "Disney is normalizing biometric surveillance in the happiest place on Earth. When one of the most beloved brands in the world embraces face recognition, it sends a message that this invasive technology is acceptable."

Disney's argument: this prevents ticket fraud and it's faster than fingerprints. Both probably true. But the switch from fingerprint to face is a meaningful jump. Fingerprints require physical contact. Face recognition works at a distance, scales easily, and has a documented history of misidentifying people with darker skin tones.

The rollout started at Disneyland in Anaheim. Walt Disney World in Orlando is expected to follow later this year.

KC Green created the "This is Fine" dog in 2013 as part of his webcomic Gunshow. A dog sitting in a burning room, sipping coffee, insisting everything is okay. It became one of the most shared images on the internet.

This week, Green discovered that an AI startup called Centric AI was running ads on LinkedIn and X featuring AI-generated variations of his character. The dog in an office. The dog at a computer. The dog promoting Centric AI's productivity tools. Fans spotted the ads and tagged Green.

Centric AI initially claimed the images were "original AI-generated artwork." After the internet did what the internet does, they pulled the ads and said they "respect creators' rights." No apology to Green. He's consulting with an attorney and has sent a cease-and-desist.

The legal question here is genuinely unresolved. If an AI model was trained on Green's widely reproduced image (which it almost certainly was) and then generates something clearly referizable to his character, is that infringement? Copyright law doesn't have a clean answer yet. What it does have: a lot of pending lawsuits from a lot of angry artists.