Microsoft’s AI diagnostic bet, OpenAI’s $30 billion round, and banks vs. agents

Compact Conversations for 2026-10-22: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Microsoft says its AI diagnostic system beats doctors

Microsoft published research in Science showing its Diagnostic Orchestrator, powered by GPT-5, achieved 83% diagnostic accuracy on NEJM cases versus 36% for doctors using references. The system uses multiple AI agents that reason through symptoms and order tests like a panel of experts.

Why it matters: This is one of the strongest claims yet for AI in clinical diagnosis, backed by peer-reviewed research. But Microsoft acknowledges it needs real-world validation before deployment.

Source: The Verge

Number to Know: OpenAI nears $30 billion funding round at $1.25 trillion valuation

OpenAI is closing in on a funding round of up to $30 billion at a $1.25 trillion pre-money valuation, led by UAE investors and Coatue. The round is positioned as a bridge ahead of a potential IPO, with signatures targeted by end of next week.

Why it matters: The valuation is slightly lower than the $1.4 trillion figure reported yesterday, and the involvement of UAE investors signals continued global appetite for AI compute deals.

Source: CNBC

The Feed

AI safety group METR finds flaws in OpenAI’s internal monitoring

Reuters reports that the AI safety organization METR has identified flaws in OpenAI’s internal monitoring system. Details are limited at this time.

Why it matters: As AI systems become more capable, the robustness of internal safety monitoring is critical for enterprise adoption and trust.

Source: Reuters

Anthropic donates Agent Skills open standard to Agentic AI Foundation

Anthropic is donating its Agent Skills standard to the Agentic AI Foundation, with over 70 platforms already integrated. This follows its donation of the Model Context Protocol to the same foundation in 2025.

Why it matters: Open standards for agent workflows reduce fragmentation and make it easier for enterprises to build interoperable AI agent systems.

Source: Anthropic

Major US banks block AI agents using customer credentials

Morgan Stanley, JPMorgan, Bank of America, and Citi are detecting and blocking AI agents that log into accounts using customer credentials. Most plan to limit agents to read-only access.

Why it matters: This is a concrete signal that financial institutions are drawing hard lines on agentic access, with implications for how AI agents are designed and deployed in regulated industries.

Source: The Decoder

Tesla begins early Optimus production but acknowledges major gaps

Elon Musk announced early production of the Optimus humanoid robot but listed unfinished work: hands, safety, reliability, and AI. Tesla did not disclose how many robots it has produced.

Why it matters: Production doesn’t mean readiness. The gap between manufacturing and useful autonomy remains wide, a reminder for anyone evaluating humanoid robot timelines.

Source: The Verge

One Thing to Try

A plain-language set of instructions you paste into custom instructions or system prompts to make LLMs flag uncertainty instead of guessing. Built from real failures, modular, and editable. Start with one or two rules from your own recent failures.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Today’s lead is Microsoft, and a striking claim. The company published research in the journal Science on Thursday saying its AI diagnostic system can diagnose medical conditions up to four times more accurately than human doctors. The system is called the Diagnostic Orchestrator, and it runs on OpenAI’s GPT-5. Instead of one model answering a question, it works like a panel of medical experts, with multiple AI agents reasoning through symptoms and histories, deciding which tests to order, and then arriving at a diagnosis.

Host B: [curious] And the numbers behind that claim are specific. On cases drawn from hundreds of New England Journal of Medicine reports, the system hit around 83 percent diagnostic accuracy, compared with roughly 36 percent for generalist physicians using references, and 22 percent without. On emergency department cases from a large Beth Israel Deaconess database in Boston, it scored 80 percent against an average of 47 percent for doctors. Microsoft built its own benchmark called SDBench and plans to open-source it.

Host A: [thoughtful] Worth noting the limits the company itself flags: this was tested only on past medical cases, not live patients, and Suleyman says it needs clinical validation. Microsoft plans to start hospital testing with Beth Israel Deaconess in the next couple of months.

Host B: One number to know today: 30 billion dollars. CNBC reports OpenAI is closing in on a funding round of up to 30 billion dollars at a 1.25 trillion dollar pre-money valuation, led by investors from the United Arab Emirates alongside Coatue. We mentioned yesterday OpenAI was courting investors at a higher 1.4 trillion figure, so the target has come down a bit. The round is being positioned as an opportunistic bridge ahead of a potential IPO, and OpenAI told investors it aims to secure signatures by the end of next week. CNBC also reports OpenAI is negotiating a separate deal for more compute capacity from the UAE.

Host A: In other news, Reuters reports that the AI safety group METR has found flaws in OpenAI’s internal monitoring system. That is the extent of what Reuters is reporting so far, and we’ll link the story in the show notes.

Host B: [conversational] Next up, Anthropic. The company says it is donating Agent Skills, its open standard for giving AI agents specialized knowledge and workflows, to the Agentic AI Foundation. Anthropic says more than 70 platforms have already integrated the standard, and this follows its donation of the Model Context Protocol to the same foundation last year.

Host A: [with a small lift] From The Decoder, via Axios reporting: several major US banks, including Morgan Stanley, JPMorgan, Bank of America, and Citi, are detecting and blocking AI agents that log into customer accounts using customer credentials. That violates bank policies requiring account holders to log in personally, and some agents appear to act without direct instructions. Morgan Stanley says it has not seen any rogue agents on its platforms, and most banks plan to limit agents to read-only access.

Host B: [lighter] And finally, Tesla. On its earnings call, Elon Musk said early production of the Optimus humanoid robot has begun on the high-volume line in Fremont, but he listed what still needs work: the hands, safety, reliability, and the AI itself. Tesla did not say how many robots it has produced.

Host A: One Thing to Try today comes from a prompt engineering community project called the Confidence Compact. The idea behind it: AI assistants are wrong in exactly the same confident voice they use when they’re right, so you can’t tell which answer you’re getting. The author kept a running log of every failure that actually mattered, turned each one into a plain-language rule, kept the rules that held up, and cut the ones that didn’t. The result is a short set of instructions you paste into your custom instructions or system prompt, things like asking the model to flag uncertainty instead of guessing. It’s modular and editable, so start with just one or two rules from your own recent failures and see if the answers change.

Host B: [with a small lift] And that’s a low-effort experiment with an obvious payoff. We’ll have the link in the show notes.

Host A: That’s Compact Conversations for Thursday. More AI news tomorrow. Until then, happy prompting.