Skip to content

Vol. I · No. 20Friday, October 23, 2026dailyprompt.si

The Daily Prompt

Superintelligence, delivered daily

No. 20 ·

Microsoft AI out-diagnoses general doctors on hard cases in a Science study

A Microsoft system more than doubled generalist doctors' accuracy on tricky medical puzzles in a simulation, while the industry handed its agent 'skills' format to a neutral foundation and security experts warned that AI-found bugs will keep coming.

Top news

  1. Microsoft's AI diagnoses hard cases twice as often as generalist doctors

    In a Science paper published Oct. 22, Microsoft's MAI-DxO, which has several AI agents act like a panel of doctors, correctly diagnosed 83% of complex New England Journal of Medicine cases. Primary care doctors and internists working alone got 36%. The test was a simulation, the doctors were not specialists, and TechCrunch notes most of the gain comes from the underlying models rather than Microsoft's orchestration.

    Source: TechCrunch

  2. Agent Skills and Open Responses become neutral, shared standards

    The Agentic AI Foundation, part of the Linux Foundation, added four projects on Oct. 22: Anthropic's Agent Skills, OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox. Skills are folders of instructions and scripts an agent loads when needed; Open Responses is a shared way to send requests to different model providers. The foundation now hosts 10 projects, including MCP and AGENTS.md.

    Source: Linux Foundation

  3. Experts warn AI-found software bugs will cause more outages

    Seven security professionals told Reuters that AI-driven bug discovery will likely bring more disruptions like this week's, when critical PHP flaws pushed governments and the BBC to take pages offline. VulnCheck says AI helped double the yearly number of discovered flaws. Some experts, including curl's lead developer, say many AI-found bugs are hard to exploit.

    Source: Reuters via ThePrint

  4. OpenAI seeks up to $30 billion at a $1.25 trillion valuation

    Sources told CNBC that OpenAI is raising up to $30 billion led by Coatue with UAE investors, at $1.25 trillion before the new money. That is below the $1.4 trillion Bloomberg reported a month ago, and terms are still being negotiated. OpenAI declined to comment.

    Source: TechCrunch

  5. Outside testers found gaps in OpenAI's system for watching its own agents

    Staff from METR and Redwood Research spent three weeks inside OpenAI this summer stress-testing the monitor that watches internal AI agents for misbehavior. They found holes, including one that could keep hundreds of agent actions unmonitored and two settings that switched monitoring off. OpenAI says it patched the flaws affecting production; the testers say they have not checked that claim.

    Source: Redwood Research

New tools in this edition

  • Ideogram Vector 1.0Design

    Saves the manual tracing step for logos and illustrations, but API use is billed per SVG (from $0.10 to vectorize an image) and complex artwork may still need cleanup.

  • Theorem OSS VerifyCoding

    Offers a stronger guarantee than fuzzing alone at no cost, but projects must be accepted, builds must target static riscv64, and proofs can take days.

  • Canva CraftImages

    The easiest way for non-designers to produce polished graphics, but its template-first approach can make work look generic, and AI use is capped by plan.

  • NirvaWearables

    Not listed in the directory.