- Uncovering AI
- Posts
- 🏥 AI just beat your ER doctor (with receipts)
🏥 AI just beat your ER doctor (with receipts)
PLUS: OpenAI's phone has zero apps, Anthropic AI just spooked Wall Street in one afternoon, and AI-assisted cyberattacks are alarmingly quick
My fellow AI explorers
We had a Harvard study drop in Science, the journal, finding that AI just out-diagnosed emergency room doctors. We had Anthropic waltz into Wall Street and spook half the financial data industry in a single press event. And we have Google Chrome quietly downloading a 4GB AI model onto your device without asking… and then reinstalling it when you delete it.
In today’s edition:
🏥 Harvard drops a bomb: AI beats ER doctors in real diagnostic tests
🤖 OpenAI's phone is real, and it has no apps
🏦 Anthropic just scared FactSet and Moody's in the same afternoon
Your ads ran overnight. Nobody was watching. Except Viktor.
One brand built 30+ landing pages through Viktor without a single developer.
Each page mapped to a specific ad group. All deployed within hours. Viktor wrote the code and shipped every one from a Slack message.
That same team has Viktor monitoring ad accounts across the portfolio and posting performance briefs before the day starts. One colleague. Always on. Across every account.
ChatGPT
According to analyst Ming-Chi Kuo, who's broken nearly every Apple hardware story of the last decade, OpenAI is actively developing a smartphone designed from scratch around AI agents. Not a ChatGPT app bolted onto Android. A whole new device where agents replace apps entirely.
The reported details:
Custom chip developed with MediaTek and Qualcomm, assembled by Luxshare
Mass production potentially targeting mid-to-late 2027 (accelerated from earlier estimates)
The device maintains "full real-time state"; it continuously captures context: location, activity, communication, environment
Lighter tasks run on-device; complex inference offloaded to cloud
Security features include pKVM and inline hashing
Here’s the concept: instead of opening Instagram, you say what you want. Instead of switching between your calendar, email, and Maps, an agent orchestrates all three silently in the background and surfaces the result. No home screen. No app grid. Just intent and execution.
Qualcomm's shares surged up to 13% in premarket trading after the report. The market clearly believes this is real… and something people will want.
There's a brutal competitive angle here too. Apple and Google both gatekeep what apps can access at the system level. By owning the hardware, OpenAI sidesteps that entirely. It gets direct access to everything. That's not a feature, that's a power move.
The honest counterpoint: This is still vaporware until it isn't. OpenAI has never manufactured a consumer product. And the graveyard of "AI-first hardware" is crowded. Remember the Humane AI Pin? That said, OpenAI has Jony Ive's firm on a separate hardware project, a war chest from their recent fundraise, and the largest consumer AI user base on the planet. If anyone can pull this off, it's them. But betting on timelines in hardware is a fool's game.
Prediction: The real disruption isn't the phone itself: it's the business model. If OpenAI controls the OS-level interaction layer, every app company that currently pays Apple 30% suddenly has a new gatekeeper to negotiate with. And this one's smarter.
AI In Education
A landmark study published in Science this week by researchers at Harvard Medical School and Beth Israel Deaconess Medical Center handed OpenAI's o1 model the same electronic health records that real ER physicians received. The result? The AI outperformed them.
Here's what the numbers actually said:
OpenAI's o1 hit 67% diagnostic accuracy at initial triage vs 50-55% for the two attending physicians
Given richer patient history, the AI climbed to 82% accuracy (physicians: 70-79%)
On longer-term treatment planning across five clinical case studies, AI scored 89% vs 34% for 46 doctors using conventional tools like Google
Here's the part that should stop you mid-scroll: the AI wasn't given clean, pre-processed data. The researchers deliberately used the messy, sparse real-world electronic records that the actual physicians saw. No clean-up. No curation. The same noise that makes ER medicine so brutally hard.
Lead author Arjun Manrai put it plainly in the press release: "We tested the AI model against virtually every benchmark, and it eclipsed both prior models and our physician baselines."
Now, before the pitchforks come out, the study is careful not to crown AI as a replacement for doctors. The models only worked with text. No visual cues, no patient distress signals, no bedside manner, no physical examination. Emergency physician Kristen Panthagani, who wasn't involved in the study, was quick to criticize the study’s methodology, and also point out that diagnosing isn't exactly the primary job of an ER doctor: keeping people alive while you figure it out is.
But here's the part I can't stop thinking about. One in five U.S. physicians was already quietly consulting an LLM for second opinions as of last month. That number is going up. Fast.
Prediction: Within three years, AI diagnostic assistants will be as standard in hospital EHRs as spell-check is in Word. The real debate won't be "should we use this." It'll be about liability, audits, and which model gets the contract. The question for founders: who builds the wrapper that makes this deployable in a regulated environment?
Claude
On Tuesday, Anthropic held an invite-only financial services briefing in New York. Jamie Dimon was in the room, announcing 10 pre-built AI agents specifically for banks, insurers, and asset managers. By the time the event was over, the market had sent a message: FactSet fell 8.1%, Morningstar dropped over 3%, and S&P Global and Moody's both took sharp hits.
The 10 agents cover:
Deal-side work: Pitch builder, meeting preparer, valuation reviewer, market researcher, model builder, earnings reviewer
Operations: General ledger reconciler, month-end closer, statement auditor, KYC screener
They plug directly into Claude for Microsoft 365, so they live inside the Excel, PowerPoint, and Word your analysts already use. A new Claude for Excel add-in can build financial models from filings, audit formulas, and run sensitivity analyses with minimal manual input.
The data partnerships are what make this serious: Moody's, Dun & Bradstreet, IBISWorld, Verisk, and a handful of others are now feeding live data directly into Claude workflows.
There's a telling moment buried in the announcement. Anthropic's own Claude Opus 4.7 scored 64.37% on the Vals AI Finance Agent benchmark, which, yes, means a failure rate that would get a human analyst fired. Anthropic's answer? Users "stay firmly in the loop — reviewing, iterating on, and approving Claude's work before it goes to a client."
That's the right answer. It's also the only answer regulators will accept for now. But the trajectory is clear.
Prediction: FactSet and Bloomberg Terminal are watching their moats get narrower every quarter. The real threat isn't that AI replaces the analyst. It's that AI makes a $25k/year Bloomberg subscription feel optional for most mid-market use cases. That's an existential pricing problem for the data incumbents, not a features problem.
30-Second AI Play
Get Your Own AI Second Opinion on Medical Records
You don't need to wait for your hospital to deploy this. You can already run a version of it today.
Step-by-step:
Open Claude or ChatGPT (o1 or Claude Opus are your best options here for reasoning tasks).
Request your medical records. In the US, you're legally entitled to them under HIPAA. Most hospital systems now offer digital export through a patient portal.
Copy the relevant sections. This works best with structured notes: chief complaint, vitals, labs, nurse triage notes. Skip imaging reports unless they include the radiologist's written findings.
Prompt the model clearly: "I'm a patient trying to understand my own health records. Based on these notes, what patterns do you see? What questions should I be asking my physician?"
Ask specifically for a differential: "What conditions could explain these symptoms together?"
Bring the output to your next appointment, not as a diagnosis, but as a conversation starter.
Important caveat: Do not use AI output as a diagnosis or treatment decision. This is for informed patient advocacy: helping you ask better questions, not replacing your doctor's judgment. And some doctors really don’t want to see AI-generated output, so if you know your doctor well and think they’re in that category, maybe just come prepared with these more informed questions.
The best health outcomes happen when patients and doctors both show up prepared.
Other Relevant AI News!
🔒 Google Chrome has been silently downloading a 4GB AI model called Gemini Nano onto user devices without consent or notification, and if you delete it, Chrome reinstalls it automatically. Security researcher Alexander Hanff says the practice likely violates EU privacy law under ePrivacy Directive Article 5(3), and at Chrome's scale, the mass deployment generates an estimated 640,000 tonnes CO2-equivalent.
💊 AI-assisted cyberattacks are reaching a tipping point. Mandiant's M-Trends 2026 report found that 28.3% of CVEs are now exploited within 24 hours of disclosure, and time-to-exploit has dropped from 700+ days in 2020 to just 44 days in 2025. AI is compressing the attack window faster than defenders can patch.
🏦 A Brazilian AI legal tech startup called Enter, which automates litigation for companies including Airbnb, raised $100M led by Founders Fund at a $1.2B valuation. It’s quiet proof that AI vertical SaaS in regulated industries is finally getting institutional backing.
Golden Nuggets
🏥 The AI vs doctor debate just became empirical, not hypothetical. A peer-reviewed study in Science is a different league from benchmarks and demos. Medicine will never look the same.
📱 The smartphone war is being re-fought from scratch. OpenAI's agent-first phone isn't a feature update. It's a bet that the entire app-based computing paradigm is obsolete.
🏦 Anthropic is targeting the rent-seekers of finance. When FactSet drops 8% on an AI announcement, that's the market pricing in a future where expensive data middlemen become optional.
Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)
Until our next AI rendezvous,
Anthony | Founder of Uncover AI