Sponsored by
Reading time: 6 minutes
🗞️ In this edition
Microsoft's AI Beat Doctors 83% To 36%
Sponsored: HubSpot
OpenAI And Anthropic Gave Their Agent Code Away
Korea's AI Buildout Needs $350 Billion For Phase One
🔌 MCP CORNER Three servers worth wiring up this week.
⚡ YOUR MOVE For people building the life they want, now.
📢 What else is happening
Microsoft put its AI against practising physicians and published the score. Four rivals handed their agent code to a foundation none of them controls. And one Korean buildout needs a number no bank has written yet.
Keep reading.
Source: Gizmodo
WHAT HAPPENED
Microsoft published a study in the journal Science on Thursday putting its diagnostic system against practising generalist physicians on hundreds of difficult published case reports. The system does not answer in one pass. Several models question the case, order tests and argue their way to a diagnosis, and it scored more than twice what the doctors did. An earlier version of the work went up as a preprint last year and has now been through peer review.
WHY IT MATTERS
Labs have spent two years claiming medical ability in blog posts and demos nobody outside could check. A journal publication puts the claim where other researchers get to pull it apart, and that changes what a hospital buying team can be shown. The harder question is not whether the thing is accurate on written cases, but who carries the liability the first time it is confident and wrong.
OUR TAKE
The obvious read is that AI has overtaken doctors at diagnosis. We think the result is narrower than that, and the giveaway is which doctors were in the comparison. Generalists working from written case reports, with no patient in the room and no colleague to ask, are not who a hard case reaches in practice. The system beat the conditions, not the profession.
Watch the Boston hospital trial Microsoft says begins within a couple of months. If it reports accuracy anywhere near the paper's on live patients, the comparison was fair and the gap is real. If it does not start, or the numbers collapse on real intake, the study measured a test and not a clinic.
From Our Sponsor:
Finally, a Monday that doesn’t feel like one.
Imagine a Saturday where you're not catching up on the CRM you didn't touch all week.
HubSpot's new self-updating CRM doesn't wait for you to log back in. Friday's calls are already logged. The deal that needs a nudge Monday morning is already flagged. The automation you set up once keeps running without you thinking about it. None of it needed you over the weekend.
Customers are already seeing it pay off: 45% more deals closed. Finally, you start Monday caught up, not behind.
Source: Linux Foundation
Two companies that call each other the main threat donated on the same morning.
Anthropic, OpenAI, Hugging Face and Alibaba have each handed an agent project to the Agentic AI Foundation, the vendor neutral body that already looks after the connection standard most assistants use. The donations cover how an agent loads its instructions, how it talks to any provider's model, how it gets trained and where it runs. The foundation now holds ten projects and is less than a year old.
When you do click a story in here, what makes you?
Source: Seoul Economic Daily
And that is the first third of the plan.
SK Telecom's data centre arm is in funding talks for the first slice of a buildout it wants to reach fifteen gigawatts in South Korea, starting with clusters in the south east of the country. Bloomberg reported the talks first. The company says it wants its first outside money by early next year, and the earliest sites will not be running until the back end of the decade.
🔌 MCP CORNER
THREE SERVERS WORTH WIRING UP THIS WEEK.
An MCP server plugs your assistant into another app, so it stops describing the work and starts doing it.
Notion MCP: Opens your own workspace to your assistant, so the doc gets updated instead of described.
Cloudflare MCP Server: Lets your assistant work on your own edge setup, so the fix happens in the chat window.
Metabase MCP: Hands your assistant your real dashboards, so you ask the question instead of writing the query.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🔎 Anthropic pointed Claude at 123 open source libraries, turned their fuzz tests into mathematical specifications and generated roughly 180 million lines of proof code, surfacing bugs that had survived years of automated testing. The proofs only cover single threaded RISC-V builds, which is not most production systems, but the method is the part worth copying. → See what it found
🩹 GitLab has patched a flaw scored the full 10.0 that lets an attacker run code on a self managed server without logging in, and it has sat in every build since 2014. Twelve fixes shipped together, and there is no public exploit yet. That gap is the window you have. → Check your version
🛒 Walmart has opened an MCP gateway plus a separate profile for each agent, so an approved shopping agent can search real inventory and buy inside a spending cap you set, while Amazon is still shutting agents out. The developer beta is invitation only and the profiles land within the month. → Get on the beta list
🔑 1Password passed $500 million in recurring revenue and says monthly active non human identities on its developer tools grew 500% in a year. It also shipped something that gives an agent access scoped to one task, without the credential ever entering the model's context. → Lock your agent's keys
A Chinese Crew Ran An Open Agent Tool At 2,250 Tenants – Microsoft says Storm-0940 took stolen ConnectWise application credentials and a freely available agent based penetration testing framework and tried to sign in to more than 2,250 Microsoft 365 tenants, at times at 2,900 attempts a minute, in some cases going from picking a target to privileged cloud access in under three hours.
Nebius Borrowed $3.5 Billion For 364 Days – The cloud provider signed a secured facility with Citi to pay for data centre build and chip purchases, and its own revenue chief said capacity prices are climbing faster than the company expected and that its 2027 capacity would sell out if it came online.
Rivian Bought A World Model Startup – The carmaker acquired Third Dimension, a 15 person British team whose software turns images and video into usable 3D worlds, to simulate the rare driving conditions it would otherwise have to wait for, with no price disclosed.
Gamma: Turns your rough outline into a deck that looks like somebody designed it.
Ermine.ai: Transcribes your audio inside the browser, so the recording never leaves your machine.
Notability: Keeps your handwriting, your typing and the recording of the meeting in one place.
Jasper AI: Writes the marketing copy in your brand's voice instead of a generic one.
Three stories about moving a claim somewhere it can be checked. Microsoft took a diagnostic score out of a preprint and put it in a journal where rivals get to attack it. Four labs took their agent plumbing out of their own repositories and handed it somewhere none of them can quietly change it. And one Korean operator put a price on the first slice of a fifteen gigawatt ambition, which is how everyone else finds out whether it was ever real.
Put one of your own claims somewhere it can be checked.
Interested in featuring your services with us? Email us at [email protected]
