Microsoft steps towards ‘medical superintelligence’

Good morning, AI enthusiasts. Microsoft’s latest AI is not here to exchange your doctor — but it surely might just grow to be their most useful colleague.

With over 4x higher accuracy than human physicians and CEO Mustafa Suleyman calling it a “big step towards medical superintelligence,” this recent diagnostic tool is more proof that AI could possibly be healthcare’s strongest ally yet.

In today’s AI rundown:

• Microsoft’s ‘step towards medical superintelligence’

• Baidu’s open-source ERNIE 4.5 to rival DeepSeek

• Automate YouTube content creation with AI

• Chai Discovery’s AI designs working antibodies

• 4 recent AI tools & 4 job opportunities

LATEST DEVELOPMENTS

MICROSOFT

🏥 Microsoft’s ‘step towards medical superintelligence’

The Rundown: Microsoft just introduced the MAI Diagnostic Orchestrator, an AI system that achieves 4x higher diagnosis results than experienced doctors on some of medication’s most difficult cases, marking a “step towards medical superintelligence.”

The small print:

• MAI-DxO simulates a virtual medical team, with specialized AI agents handling hypothesis generation, test selection, and value monitoring.

• Researchers created SDBench, a benchmark with 304 complex cases — with MAI-DxO, paired with OpenAI’s o3, achieving the very best accuracy in testing.

• The MAI/o3 pairing solved 85.5% of cases appropriately, with a gaggle of physicians with 5-20 years of experience averaging just 20%.

• The AI system also resulted in cost savings over human doctors, spending $2,397 per case in comparison with a median of $2,963 for physicians.

Why it matters: A step towards “medical superintelligence” is a robust statement, but MAI’s numbers in comparison with physicians are truly jaw-dropping. Plus, ordering fewer unnecessary tests and nailing tough diagnoses directly addresses healthcare’s current paradox: the over-treatment of easy cases and under-diagnosis of complex ones.

TOGETHER WITH BASETEN

🚀 The platform for mission-critical inference

The Rundown: Baseten’s inference platform brings enterprise-grade performance and reliability to each stage of your model’s lifecycle. Launch custom, fine-tuned, or any open-source model and scale without limits across any cloud, including your personal VPC.

With Baseten, you possibly can:

• Construct with frontier Model APIs which can be purpose-built for production

• Scale Dedicated Deployments without limits across different cloud providers and regions

• Train and fine-tune any model on any dataset and deploy checkpoints seamlessly to production

Start doing inference now on infrastructure built for mission-critical workloads.

BAIDU

🤖 Baidu’s open-source ERNIE 4.5 to rival DeepSeek

The Rundown: Chinese giant Baidu just open-sourced 10 versions of its recent ERNIE 4.5 model family, including a 424B parameter multimodal system that outperforms DeepSeek V3 on key benchmarks despite being half the scale.

The small print:

• The models range from tiny 300M parameter versions to massive 424B systems, all available under Apache 2.0 licensing on Hugging Face.

• A “Heterogeneous” training architecture allows text and vision capabilities to bolster one another relatively than compete for resources for increased efficiency.

• Baidu’s largest model beats DeepSeek V3 on 22/28 benchmarks, while its variants also compete with o1, GPT 4.1, and Qwen 3 across quite a lot of tasks.

• The discharge marks Baidu’s first move into open-source models, coming only a 12 months after its CEO appeared against the route prior to the launch of DeepSeek.

Why it matters: While the U.S. has its AI rivalry between OpenAI, Google, and Anthropic, an intense one can be brewing between China’s giants. Baidu, ByteDance, Alibaba, Tencent, DeepSeek, and others proceed to push the pace of model releases — and the standard of offerings that make up the highest of the open-source charts.

AI TRAINING

🎬 Automate YouTube content creation with AI

The Rundown: On this tutorial, you’ll learn use Google’s Gemini AI to investigate videos and routinely generate titles, descriptions, tags, and chapters — streamlining your YouTube content workflow.

Step-by-step:

1. Go to Google Gemini, click the attachment button, and upload your video file

2. Generate titles with: “Analyze this video and supply 10 compelling YouTube titles”

3. Ask it to offer you an in depth video description with a hook, summary, and call-to-action.

4. Request it for tags as comma-separated values, plus chapter timestamps

Pro tip: Save these prompts as templates and customize the tone (skilled, casual, energetic) to match your brand voice.

PRESENTED BY EXPLORIUM

🧠 Higher business data for smarter AI agents

The Rundown: Explorium is the all‑in‑one B2B data engine that fuels your AI agents and workflows with live, actionable data right where you’re employed. Power your AI workflows with higher data, including 150M+ company profiles and 800M+ professionals.

Integrate the Explorium MCP seamlessly together with your favorite tools:

• n8n, Zapier, Make users – Automate data enrichment and custom workflows at scale

• ChatGPT & Claude Teams Subscribers – Enhance AI Agents with accurate, up-to-date company and speak to data right in your chat

• Connect external data enrichment to boost GTM operations – Lead Enrichment Flow, Deep Research, Outreach Personalization, and E2E AI SDR Teams.

Join to your free API Key – start using the Explorium MCP Integration today.

AI & DRUG DISCOVERY

🧬 Chai Discovery’s AI designs working antibodies

The Rundown: OpenAI-backed Chai Discovery announced Chai-2, an AI capable of making functional antibodies for drug development, — nearing a 20% hit rate that marks a 100x improvement over methods that typically achieve lower than 0.1%.

The small print:

• The model designed antibodies against 52 different disease targets, finding successful treatments for half of them by testing just 20 candidates each.

• Traditional antibody discovery requires screening tens of millions of candidates over months or years, with Chai-2 delivering leads to just two weeks.

• Chai-2 works “from scratch,” creating completely recent designs just by a goal’s structure while not having any pre-existing examples.

• Chai researchers said the system is like “Photoshop for proteins,” letting scientists specify exactly where antibodies should attach to disease targets.

Why it matters: AI’s role in drug discovery and molecular-level design is ramping up in an enormous way. With high R&D costs often resulting in firms ignoring treatments for rare diseases, AI models may also help cut the time and expense of finding recent antibodies, giving patients access to precision medicines tailored to their conditions.

QUICK HITS

🛠️ Trending AI Tools

• 🧠 Hunyuan-A13B – Tencent’s recent open-source hybrid reasoning model

• 🎇 Qwen VLo – Alibaba’s 4o-style model for image generation and editing

• 🎼 Songscription – Turning audio clips immediately into sheet music

• 🍎 NotebookLM – Now available via Gemini in Classroom for educators

💼 AI Job Opportunities

• 📞 The Rundown – Account Manager

• 🧠 OpenAI – Head of Pro Subscriber Community

• ⚡ Figure AI – Electrical Engineer Intern

• 📊 Meta – Data Scientist

📰 Every little thing else in AI today

RAISE Summit in Paris, July 8-9 — All things AI. Join SambaNova at booth #9, snag an invitation to an exclusive soirée, and catch CEO Rodrigo Liang’s keynote on Open Source AI.*

Mark Zuckerberg introduced “Meta Superintelligence Labs” to employees, with Alexandr Wang and Nat Friedman leading 11 hires from OpenAI, Google, and Anthropic.

Apple is reportedly considering leveraging AI from Anthropic and OpenAI for the revamped Siri over in-house models, in keeping with a brand new report from Bloomberg.

The Mayo Clinic unveiled StateViewer, an AI tool that analyzes brain scans to assist discover nine various kinds of dementia at 2x the speed and 3x the accuracy.

Cursor launched recent apps for mobile and browser, allowing users to administer and monitor agents via natural language outside of its IDE.

Google announced Gemini in Classroom, a collection of AI features and tools for educators for tasks like lesson planning, NotebookLM access, and student performance analytics.

*Sponsored Listing

COMMUNITY

🎥 Watch our latest workshop

Try our latest workshop on construct advanced n8n agents that react, research, and automate like pros, with Dr. Alvaro Cintas, The Rundown’s AI professor

Watch it here. Not a member? Join The Rundown University on a 14-day free trial.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team