😺 Microsoft AI scored 83% where doctors got 36%
So apparently, researchers at MIT left 13 AI agents alone on a simulated island for 30 hours with no assignment and a gentle nudge to stay curious. No tasks, no deadlines, no Slack pings. Honestly, that's everyone's dream Friday, except the island is fake and the coworkers are all Claude.
Here's the weird part. The agents started inventing their own challenges, the way kids turn a playground into a game nobody asked for. Later, when everyone got real tasks, the agents with that open-ended island time generally beat the control agents who skipped it. According to the preprint, aimless exploration may teach strategies that carry over to real work.
Quick caveat: it's an early paper, and the agents' instructions encouraged curiosity, so calling it spontaneous play is a stretch. Still, if your boss ever questions your 2 p.m. wander around the office, you now have a citation.
Here’s what happened in AI today:
🐱 Microsoft's diagnostic AI outscored doctors in a simulation.
📰 OpenAI is reportedly seeking $30B at a $1.25T valuation.
📰 Barclays plans 5,000 job cuts, citing AI.
🍪 Canva launched Craft, its first in-house image model.
🎓 Forecast your real launch date with ChatGPT.
…and a whole lot more that you can read about here.
We're rounding up Black Friday and Cyber Monday deals for a special edition of The Neuron, and your brand can grab a sponsored spot in front of 700,000 subscribers. Reach out for details!
😺 Microsoft's AI Beat Doctors on Hard Medical Cases. The Fine Print Matters.
Picture a case that stumped the hospital: odd symptoms, no obvious cause. Microsoft AI says its new system, MAI-DxO, solved 83% of a set of hard cases from the New England Journal of Medicine. The doctors in the same test got 36%.
MAI-DxO (short for Microsoft AI Diagnostic Orchestrator) works like a panel of virtual doctors. Each one proposes diagnoses, picks which questions to ask or tests to order, and challenges the group's emerging answer, over several rounds. It also weighs what each test would cost, so it can't just order everything.
The test matters as much as the score. Medical exams hand you the whole case up front, and real doctors never get that. So Microsoft built SDBench (Sequential Diagnosis Benchmark), where the player starts with limited information and decides what to ask or order next, one step at a time. It covers 304 hard NEJM cases plus more typical emergency-room cases.
Here's how the scores shook out:
Hard NEJM cases: MAI-DxO hit 83%, versus 36% for physicians with textbooks and internet access and 22% without.
Typical ER cases: MAI-DxO hit 80%, versus 47% for physicians.
Cost: Microsoft says it reached the right diagnosis at roughly half to a third of the cost in the NEJM work, with fewer tests.
Now the fine print, which is where our eyebrows went up. TechCrunch's write-up flags a few things worth knowing before you picture a robot doctor:
The doctors were primary care physicians and general internists working alone, not the specialists and teams who usually handle cases this hard.
It was a simulation with no real patients, and Microsoft says real-world performance is still to be tested.
Most of the gain comes from the AI models underneath. GPT-5 alone scored 78.9%, and MAI-DxO added about four points on top.
Microsoft's star diagnostic panel does its best work on OpenAI's models. Family dinner must be fun.
Still, the panel idea is the part worth stealing. Getting several AI agents to argue before they answer beat a single answer, and that trick will probably show up anywhere the stakes are high, like law and finance.
What comes next is the real test. Microsoft is teaming up with Beth Israel Deaconess Medical Center in Boston for a study in emergency and urgent care. Patients get standard care, consenting patients also talk to the system, and researchers compare its recommendations with the human doctors'. Microsoft AI CEO Mustafa Suleyman is pitching all this as a step toward medical superintelligence, and Microsoft plans a healthcare model with Mayo Clinic next year. There's no pricing or availability for MAI-DxO itself yet.
The number to watch comes from Boston: how often the panel's call matches a real ER doctor's, with real patients in the room.
Read our deeper analysis of Microsoft’s Superintelligence breakthrough.
FROM OUR PARTNERS
Every AI waits for you to ask. Demi doesn't.
Demi sees your screen. It spots your open loops. It pulls context from apps you use, catches what's about to slip, and gets it done before you ask.
Here's how:
Never lose a deal to silence. Demi spots the cold leads, researches them, and drafts follow-ups.
Never drop a promise from a call. Demi turns it into action items, then does them.
Never build Monday's report again. Meta and Google numbers are ready before you ask.
Neuron readers get early access + $100 credits - Code NRN100.
🎓 AI Skill of the Day: Forecast Your Real Launch Date With a Monte Carlo Simulation
Every project plan has a finish date, and most of them are a polite fiction. Plans add up each task's most likely duration, but real tasks run far later than they ever run early.
We tested a five-task launch: design then build, legal review then content, and QA at the end. The most likely estimates add up to 18 days. Simulated 10,000 times, it hit 18 days or less only about 10% of the time. The median finish was around 21 days, and 9 times out of 10 it was done by roughly 24 to 25.
That 18-day date is a plan, not a forecast. A Monte Carlo simulation (rolling the dice 10,000 times and counting) turns one date into odds, and ChatGPT can run it from a spreadsheet in minutes.
It also shows where to spend your worry. Build and design drove most of the swing, while legal review barely moved the date. You can stop losing sleep over the lawyers. Probably.
Try it:
Build a simple sheet with one row per task and these columns: Task, Best case, Most likely, Worst case (in working days), and Depends on. Ask the people doing the work for all three numbers, because the worst case is where the truth hides. Our example: Design 3/5/13, Build 8/10/20, Legal review 4/6/15, Content 5/7/12, and QA and launch prep 2/3/6.
Upload the sheet to ChatGPT's data analysis tool and paste the prompt below, filling in your current planned finish. Using Gemini instead? Make sure code execution is on and that it shows you the code it ran, because a description of a simulation is not a simulation.
Read the odds, not the average. Plan around the P80 (the date you hit 8 times out of 10), then aim your risk mitigation at the top two tasks in the sensitivity chart.
You are a project risk analyst. I've uploaded a project schedule with columns: Task, Best case, Most likely, Worst case (all in working days), and Depends on.
Run a Monte Carlo schedule simulation:
- Model each task's duration with a PERT-beta distribution built from its three estimates.
- Respect the Depends on column and calculate the project finish for every run. The critical path can change from run to run.
- Run 10,000 simulations.
Then give me:
- The P50, P80, and P90 finish in working days, plus the chance of finishing by my current planned finish of [X] days.
- A bar chart ranking which tasks' durations drive the finish date the most.
- The percentage of runs where each path was the critical one.
- A 30-second plain-English summary I can paste into a stakeholder update.
Before running anything, list every assumption you're making (distribution, task independence, working days, no resource limits). Show me the Python code you used so I can check it. If any row looks wrong, for example a most likely value outside the best-to-worst range, flag it instead of silently fixing it.Two cautions. The default setup treats tasks as independent, so if the same overloaded team owns three of them, the real risk is higher than the simulation says. And the odds are only as good as the estimates, so skim the code and assumptions before you quote any number to your boss.
Have a specific skill you want to learn? Request it here.
FROM OUR PARTNERS
Disconnected systems and manual reconciliation can create hidden costs across tax, finance, and IT.
On October 27, 2026 at 1:00 PM ET, TechRepublic and Sovos present Connected Compliance, a 30-minute virtual fireside chat examining how compliance friction affects working capital and operations, and where connected data and AI can help address inefficiencies.
🍪 Treats to Try
*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!
*Got an AI project living rent-free in your notes app? The Neuron Academy has practical courses on prompting, research, and building with AI. Give that idea something to do →
What Is Data Management? A Guide to Systems, Processes, and Tools. Learn more
1Password for Agents gives your AI agent a digital key to log in for you, and you approve each sign-in on your device, so the agent never sees your password.
Canva Craft is Canva's first in-house image model, which plans a design before rendering it and handles legible text, 2K output, and 30+ aspect ratios.
DeepCell builds linked documents, spreadsheets, and presentations in one project file, so your numbers, narrative, and slides update together.
Radaric reads your RSS, Telegram, X, and YouTube feeds, learns from what you approve or reject, and groups duplicate coverage into one reading queue.
ElevenLabs Realtime Translation translates live speech across 90+ languages with expressive audio, captions, and consistent voices, handy for meetings, events, or support calls.
OpenAI's Decisions API sorts, ranks, and routes batches of up to 400 questions with short answers, and OpenAI says it runs up to 10x faster than using GPT-6 Luna through the Responses API.
📰 Around the Horn
Claude Code made auto mode the default on Aug 14. In Anthropic’s test of 1,053 people, a planted dangerous command slipped past humans 86% of the time, versus 11% for automated checks, and humans got worse in longer sessions. It reduces risk without erasing it, so review high-stakes changes yourself.
The U.S. Treasury proposed restricting certain investments above $50,000 in AI, supercomputing, and biotech involving China, Russia, Iran, North Korea, and Cuba, pending public comment.
OpenAI was reportedly working on a $30B raise at a $1.25T valuation, with Coatue and UAE investors among the possible backers.
DeepSeek reportedly raised about $15B, with CATL and Tencent among its backers, though the company hasn't confirmed the round.
Morgan Stanley, Bank of America, JPMorgan Chase, and Wells Fargo were blocking third-party AI agents from customer accounts, saying they can't reliably tell an agent from a person.
Anthropic and Theorem said Claude proved security properties across 123 open-source libraries and found hidden bugs, though only for narrow cases like small RISC-V binaries.
Barclays announced plans to cut about 5,000 roles, roughly 5% of its workforce, and cited AI among the tools behind its £1.1B savings target by 2027.
Want absolutely EVERYTHING that happened in AI? Click here!
UPCOMING EVENT: The Neuron IRL in San Francisco
The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.
It’s free, but space is limited. Save your spot!
💡 Intelligent Insights
The Guardian argues that the AI we were promised would save us time is doing the opposite and costing people time instead.
Bill Gates told The Verge's podcast to brace for roughly 20 years of turbulent AI disruption.
TechRadar argues Amazon's $75 to $100 Bee wearable, which records your conversations, can turn casual remarks into shopping prompts and raises real privacy questions, including for bystanders.
Boltzbit CEO Yichuan Zhang argues AI wrapper startups are fragile because they rent someone else's model, pointing to Jasper's revenue falling from about $120M to $55M after ChatGPT added native writing features.
Atos VP Pete Evans argues companies should pick agentic AI projects by measurable value, noting one procurement process that looked like four steps on paper turned out to be about ten.
CGI's Amar Aswatha argues hourly billing no longer fits AI-era tech work and pitches hybrid contracts that mix a base fee, usage charges, and outcome bonuses.
New from The Neuron: AI Explained
Click the image above to watch the video!
New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube
A Cat’s Commentary
 | That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today! |
Btw: We just launched a robotics newsletter! Sign up for it here.
P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.