Microsoft AI Beats Doctors at Diagnosis & OpenAI Nears $1.25 Trillion Valuation - AI News (Oct 24, 2026)

An AI system just solved some of medicine's toughest diagnostic puzzles more than twice as often as experienced doctors, and it ordered fewer tests while doing it. We'll get into what that really means in a moment. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 24th, 2026. We've got big money, custom chips, hidden watermarks, and a few updates on stories we've been tracking. Let's get going.

Microsoft AI Beats Doctors at Diagnosis

Let's start with that diagnosis story. Microsoft has published results in the journal Science for its Microsoft AI Diagnostic Orchestrator. On notoriously difficult case studies from the New England Journal of Medicine, the system reached about 83 percent accuracy. Generalist physicians managed around 36 percent with reference materials, and roughly 22 percent without. It also leaned on fewer tests and procedures, which hints at lower costs. The best results came when it ran on OpenAI's GPT-5. Important caveats, though: these are curated test cases, not live patients. Microsoft plans trials at a Boston hospital soon, positions the tool as an assistant rather than a replacement, and the open question of who's accountable when an AI-assisted diagnosis goes wrong hasn't gone away.

OpenAI Nears $1.25 Trillion Valuation

Now to the money. OpenAI is reportedly close to raising up to 30 billion dollars at a valuation of 1.25 trillion dollars before the new cash comes in. Investors from the United Arab Emirates and Coatue are leading, and the round is framed as a bridge to a possible IPO, which OpenAI has already filed for confidentially. Notably, the deal may come with a separate agreement for more computing capacity from the UAE, a reminder that in this race, data centers matter as much as dollars.

Flapping Airplanes Valuation Hype Questioned

At the other end of the scale, a widely shared post says startup Flapping Airplanes may be raising at a 5 billion dollar valuation, up from 1.5 billion only months ago. The author urges caution. The pitch has drifted from radically sample-efficient training toward broader GPU and inference efficiency claims, and private valuations can be driven by star hires, secrecy and investor momentum. The author does note that a real breakthrough in learning from less data would matter a lot, both for AI timelines and for safety. For now, though, the verdict is: interesting, but show us the evidence.

OpenAI Jalapeño Custom Inference Chip

Staying with OpenAI, hardware VP Richard Ho gave an interview about Jalapeño, the company's custom inference chip. The team started from a blank slate, tightly pairing memory with dedicated compute cores. That makes programming a bit harder but squeezes out better efficiency and lower latency for large language models. Ho also said AI tools helped speed up the design work, and that OpenAI is already planning future generations. The takeaway: OpenAI increasingly sees silicon as a core part of its strategy, not just something it buys.

Microsoft Decision-1 Compact Model

On the model front, Microsoft released Decision-1, a compact nine-billion-parameter model built not for chatting but for picking between predefined options in a single pass, ideal for routing, classification and verification jobs. Eyebrows went up because it's post-trained from Alibaba's Qwen, though Microsoft says it will later rebase it on its own and OpenAI models. It's API-only for now, and its impressive speed and benchmark claims haven't been independently checked yet.

Gemini 3.5 Pro Never Shipped

An update on Gemini. Sundar Pichai explained why a standalone Gemini 3.5 Pro never appeared. Google chose to push agentic and coding improvements through the smaller Flash models first, since they allowed faster iteration. By the time those gains were ready to carry over, Pro's pretraining had already wrapped. A rare glimpse at how roadmaps bend mid-flight.

Perplexity JOLT Teacher-Student Training

A few research highlights. Perplexity introduced JOLT, a training method built on the idea that a brilliant teacher isn't automatically a good one. It trains teacher and student models together, so guidance matches what the student can actually absorb, and reports gains across math, coding and agent benchmarks. NVIDIA Research showed Joints, Camera, Action!, which generates video alongside matching 3D motion for all kinds of animated rigs, not just human skeletons, and can also convert video back into rig motion. And a hobbyist project called WorldTUI trained a small model to pretend to be a computer terminal, with no real shell behind it. It handles simple commands convincingly, stumbles on complex ones, and runs right in your browser, a curious peek at learned interfaces.

NVIDIA JCA Video and 3D Motion

Then there's the Enron Brain, a project that turned roughly 45,000 Enron emails into an AI-built knowledge base where every fact is tied to a source email and a date, so outdated claims drop away. A spot-check found 93 percent of sampled facts fully supported. The point: AI can organize messy corporate knowledge, as long as everything stays traceable.

WorldTUI Neural Terminal Simulator

Speaking of traceability, Anthropic's latest Claude models now embed invisible watermarks in their text. There are no hidden characters; instead, word choices are subtly nudged so detectors can later flag the text as AI-generated. That applies across the API, Claude Code and the desktop app, and it's largely driven by the EU AI Act. OpenAI and Google have similar efforts. Practically, anything you draft with these tools is becoming more identifiable, whether you notice or not.

Enron Emails Become AI Knowledge Base

On the safety side, an update to the AI monitoring story we've followed. METR, working with a Redwood Research staffer, spent three weeks red-teaming OpenAI's internal system for catching misbehaving agents. They found gaps, including traffic that wasn't eligible for monitoring and ways agents could potentially launch unmonitored instances. OpenAI says the issues are patched, though METR hasn't verified that. The lesson: oversight systems need to be built for adversaries, not just accidents.

Claude Outputs Carry Invisible Watermarks

Anthropic also handed its Agent Skills standard, a simple format for packaging know-how for AI agents, to the Linux Foundation's Agentic AI Foundation. It's already used by more than seventy platforms, including ChatGPT and GitHub Copilot, and will now be maintained in a vendor-neutral way.

METR Red-Teams OpenAI Monitoring

On policy, an essay argues that a go-it-alone American slowdown in frontier AI wouldn't work, since talent and know-how would simply move elsewhere. Its proposal is a verifiable US-China agreement centered on compute, the most measurable input. The author admits that's only a temporary brake, but argues even buying time is worth it.

Anthropic Donates Agent Skills Standard

And finally, a quick update on the Samsung TV saga. The researcher who rooted the set and found AI features listening for sounds like a barking dog or a crying baby says OpenAI has now suspended their account following the investigation. No explanation yet, but it's added fuel to debates over device privacy and who really controls the hardware we buy.

That's it for today's edition of The Automated Daily. From trillion-dollar valuations to watermarks you can't see, it's clear the industry is racing ahead while trying to keep things accountable. Links to all the stories are in the episode notes. I'm TrendTeller, thanks for listening, and see you tomorrow.

Microsoft AI Beats Doctors at Diagnosis & OpenAI Nears $1.25 Trillion Valuation - AI News (Oct 24, 2026)

Our Sponsors

Today's AI News Topics