😸 Microsoft AI beats doctors 🤖

PLUS: Claude could power Siri?!

Welcome, humans.

Y’know how Apple just gets dumped on for Siri… not really working? Well, as it turns out, they've put Craig Federighi (Mr. iOS himself) and Mike Rockwell (the Vision Pro mastermind) in charge of making Siri finally work. And right now, Craig and Mike are considering outsourcing Siri's brain to either ChatGPT or Claude.

According to Bloomberg analyst Mark Gurman, Apple is asking both Anthropic and OpenAI to train models on Apple's cloud infrastructure. Like many devs out there, it seems that they’re learning towards going with Claude.

Some people see Apple as no longer being a risk taker, but an asset manager (a cash cow as it’s called), with billions in its war chest just sitting around.

OpenAI and Anthropic have spent years (and billions) perfecting their models, while Apple has been... making thinner iPads?

So why stop at importing the AI, Apple? If Meta has taught us anything, it’s if you can’t beat ‘em, buy ‘em!

P.S: We had a few last minute ad slots open up in July, but they’re on a first come first serve basis. If you want to reach 550,000+ AI professionals, fill out this quick form!

Advertise in The Neuron here.

Here’s what you need to know about AI today:

Microsoft's AI diagnosed 85% of cases that stumped expert doctors.
UK entry-level jobs dropped 33% since ChatGPT launch.
Gartner predicted 40% of agent projects cancelled by 2027.
UC Davis created brain-to-speech chip.

Microsoft's AI Just Diagnosed Cases That Stumped Expert Doctors—With 85% Accuracy

Microsoft just made a medical AI claim so extraordinary, it makes those “AI passes medical exam” headlines look like child's play.

Microsoft claims its new AI tool can diagnose complex health conditions better than actual doctors.

Here’s what happened:

Microsoft announced it is pursuing the path to medical superintelligence with a new system leading the way.
The system, MAI-DxO (Microsoft AI Diagnostic Orchestrator), correctly diagnosed 85.5% of the most complex medical cases from the New England Journal of Medicine.
For context, a group of experienced physicians only solved 20% of the same cases.

The breakthrough comes from teaching AI to think like a panel of doctors rather than a single expert. MAI-DxO simulates multiple physicians with different approaches collaborating on each case. It can:

Ask follow-up questions like a real doctor would.
Order specific tests based on symptoms.
Work within budget constraints (because infinite testing isn't realistic).
Verify its own reasoning before making final diagnoses.

This shows Microsoft’s model = smartest + cheapest.

Now get this: The 304 cases Microsoft tested weren't your typical “patient has a cough” scenarios. These were the medical mysteries that end up in NEJM—the cases that require multiple specialists and often take weeks to solve.

Why this matters: The AI didn't just beat doctors at diagnosis. It did it cheaper. U.S. healthcare spending is approaching 20% of GDP, with up to 25% wasted on unnecessary tests and procedures. An AI that diagnoses better while ordering fewer tests could fundamentally reshape healthcare economics.

Imagine every small-town GP having access to the collective knowledge of the world's best specialists. The path from research to your doctor's office will be long and carefully regulated. But make no mistake: this is what the beginning of medical transformation looks like.

Prompt Tip of the Day

Phil Scmid wrote a great blog breaking down “context engineering” and what it means beyond your typical prompt. The TL;DR = It's about building systems that give AI all the right information, tools, and format at the right time.

Think of context as everything your AI sees: instructions, conversation history, retrieved documents, available tools, even output format requirements.

Credit: Phil Schmid

Phil's example shows the difference between a “cheap demo” agent that only sees a basic request, versus a “magical” agent that has access to your calendar, past emails, contact info, and scheduling tools.

Instead of a robotic “What time works for you?” response, you get: “Tomorrow's packed on my end, back-to-back all day. Thursday AM free if that works? Sent an invite, lmk if it works.”

The magic isn't in the prompt. It's in the prep work that makes the prompt unnecessary.

Here’s how to put this practice: Before you prompt “help me schedule a meeting,” first gather your calendar, the other person's availability preferences, and past meeting patterns, then let the AI suggest optimal times. Try using a tool like n8n or OpenAI’s new connectors system to pool this data all in one place for your AI to access.

Want more like this? Check out all of our Prompt Tips of the Day from June here.

P.S: Completely new to AI? Start here!

Treats To Try.

*Asterisk = from our partners. Advertise in The Neuron here.

*Luma AI turns your text into videos and lets you completely restyle any video's background, characters, or setting in post-production.
Conquest Planning helps financial advisors create comprehensive financial financial plans 90% faster for any wealth level ($110M raised). (raised $80M).
Glean builds assistants and agents that work with your company knowledge to automate tasks, helping employees find information, create content, and automate workflows
Jotform’s new Presentation Agents makes your slides self-presenting with AI narration and live Q&A.
Cursor, the most popular AI coding startup, now works on web and mobile so you can start tasks with their background agents anywhere and review them in your editor.
BeanBook scans your coffee bags to automatically log roaster details and brewing notes, automatically identifying roaster details, origin, and process so you never forget which beans you loved. (ioS only rn).
Songscription automatically transcribes your audio or MIDI files into downloadable sheet music in PDF, MIDI, or MusicXML format for any instrument (based on this paper—read more).

See our top 51 AI Tools for Business here!

Around the Horn.

Read the full report here

The U.S. Senate will limit its proposed “10 year ban” on state-wide AI regulation to only five years, while state laws regulating deceptive practices, child safety, child abuse material, and identity misuse get specific protections.
Google announced 30+ free education AI tools at ISTE 2025, launched Gemini in Classroom (an education-focused version of its AI app), and expanded its Google Vids access to all Education users.
OpenAI is… taking the week off?? You know what, that’s A-OK with us…
Gartner predicted more than 40% of all “agentic” AI projects (so, workflows where AI makes the decisions in stead of humans) will be cancelled by 2027 because of high costs, inadequate risk, and business value that’s still TBD.
New data showed entry level jobs in the UK have dropped by one-third since the introduction of ChatGPT in November of 2022.

For the latest AI deep dives, check out our Explainer articles here!

AI In The Real World

China hosted the first 3 vs 3 autonomous robot football match, featuring bot ballers from Booster Robotics (here’s clips of the game w/ Chinese TikTokers reacting to them. Our fave? “Not bad…the robots are at the same level as China’s national football team” Boom roasted!).
TomTom announced it's cutting 300 jobs as it pivots to embrace AI in its mapping technology, joining the growing list of traditional tech companies restructuring for the AI era.
UK retailers are using AI-powered surveillance systems to stop the rise of organized criminal shoplifters.
Spotify hosted a band called The Velvet Sundown that got over 500K streams across two albums before anyone realized the band was AI (and NGL, its top song Dust on the Wind kiiiinda slaps??).
- Redditors are pretty sure it was made with Suno; also, Rick Beato’s analysis is an awesome breakdown of how to diagnose AI music w/ ProLogic.
Amazon, Google, McGraw Hill, Microsoft, NVIDIA, and OpenAI are among the 60+ companies making a commitment to provide AI tools to K-12 students over the next four years as part of the White House’s AI Education Pledge, which included 60+ companies like (for more on how AI is being applied in education, check out our deep dive from yesterday!).
A team of researchers at UC Davis created a new brain-to-voice chip that “instantly” translates brain signals into sounds (including intonation, emphasis, and even singing) without requiring ground-truth speech—essentially restoring intelligible, expressive communication (paper).

Under the Hood

Baidu open-sourced its ERNIE 4.5 family of models, which provides multimodal AI processing of text and images across model sizes from 0.3B to 424B parameters, while Tencent released Hunyuan-A13B, which delivers high-performance language reasoning using just 13B active parameters with 256K (~192K words) context support.
NeMo Retriever from NVIDIA extracts data from your PDFs and visual documents 15X faster with 50% better accuracy, helping you build applications that can instantly search through and retrieve specific information from your document libraries—try it here.
FineWeb2 from HuggingFace gives you a massive multilingual dataset with 1,000+ languages to train your language models, sourced from 20TB of cleaned web data.
Claude Code’s new hooks feature lets you automatically run custom commands when Claude Code edits files, runs scripts, or sends notifications.
Speaking of context engineering, Simon Willison recommends Drew Breunig’s writing on why long context fails and how to fix it, with plenty of technical tips to address agentic AI’s current limitations. His key tactics include:
- Keep tool count under 30 (3x better accuracy).
- Split complex tasks across separate AI threads.
- Use pruning tools like Provence to cut 95% of irrelevant content.
- Give AI a “scratchpad” (like how Claude works) for notes outside main context.
Our favorite insight: Most agent failures aren't model failures: they're context failures. The magic happens when you engineer the right context, not when you find the perfect prompt.