Poached Nobel Laureates, Secret OpenAI Models, and the Five Eyes AI Cyber Alert

Yesterday in AI  |  June 23, 2026

Poached Nobel Laureates, Secret OpenAI Models, and the Five Eyes AI Cyber Alert

The battle for elite AI talent and algorithmic dominance just reached a staggering new peak. Today's episode breaks down the massive shockwave running through Google DeepMind as Nobel Prize-winning chemist John Jumper walks out the door to join Anthropic, compounding a brutal week of high-profile departures.
 
Plus, we unpack the unprecedented joint public warning from the Five Eyes intelligence alliance, putting a concrete timeline of mere months on AI-accelerated offensive cyber threats. We explore the quiet corporate shift as US executives deploy open-weight Chinese models like GLM-5.2 to smash design benchmarks, dive into the leaked demos of OpenAI’s unannounced GPT-5.6 Pro model crushing 3D design tasks in the wild, examine Nvidia's new standardized safety certification architecture for physical humanoid robotics, and close on a deeply moving look at how OpenAI's o3 model is compressing rare childhood disease diagnostic timelines from seven years down to weeks.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, I'm back. This is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Tuesday, June 23rd, and the Talent Wars just claimed a Nobel Prize winner. Plus, five global intelligence agencies just put a concrete timeline on an AI cyber threat nobody wanted to admit was this close, and developers are leaking demos of a secret open AI model that's already beating the competition in the wild. Let's get into it. Google DeepMind is having a rough week. John Jumper, who co-led AlphaFold and won the Nobel Prize in Chemistry for it, just announced he's walking out the door after nine years to join Anthropic. If you're not familiar with Alpha Fold, it's the model that cracked one of biology's absolute hardest problems, predicting how proteins fold into 3D shapes. It's a riddle that scientists spent 50 years failing to solve, and AlphaFold did it so well that it earned Jumper and Demis Hasabas a Nobel Prize back in 2024. It stands as probably the single biggest scientific achievement in the history of AI. And now Anthropic has poached him. The timing here is brutal for Google. Just days before Jumper's exit, Gemini co-lead Noam Shazir abandoned DeepMind for OpenAI. That's two marquee researchers jumping ship to two of DeepMind's hottest rivals in a single week. Deepmind spent a decade building the premier AI research lab on Earth, but in 2026, it keeps losing its crown jewels to the very companies it inspired. Jumper is taking a quick breather before starting in Anthropic, which perfectly lines up ahead of Anthropic's science event on June 30th. DeepMind's ultimate competitive edge has always been pure science, and that edge just got a whole lot narrower. While Anthropic is winning the war for elite scientific talent, a rare global intelligence alliance is warning that the software these scientists are building is about to weaponize the digital world. The Five Eyes Intelligence Alliance, comprising the US, UK, Canada, Australia, and New Zealand, just issued a rare joint public warning, stating that frontier AI models are actively accelerating offensive cyber capabilities, and they didn't give a timeline of years or decades. They said the window before these AI-enabled cyberattacks scale significantly is measured in months. This is the most direct, coordinated public alert we've ever seen from the intelligence community. They aren't talking about theoretical sci-fi doomsday scenarios. They're looking at active, immediate risk. The statement dropped alongside reports confirming that recent U.S. guidance was what triggered the global suspension of anthropics fable models for foreign nationals. It reveals a clear downstream pattern from the Commerce Department. If a tech lab can't isolate foreign nationals from accessing advanced cyber intelligence in real time, the government will force them to yank the model offline entirely. For any IT leader running critical workflows on third-party frontier AI, model recall is no longer a hypothetical compliance risk. If Washington can flip a switch and pull your software, multi-model strategies and locally hosted open source alternatives are no longer just technical preferences. They are core risk management. And speaking of deploying open source alternatives to hedge your bets, a growing number of American executives are quietly looking to China for their production workflows. A striking New York Times deal book piece reveals that U.S. companies are increasingly bypassing domestic tech giants to deploy cheaper Chinese open source models directly into scale production. The economics are hard to argue with. Chinese models have become entirely good enough to handle real workflows when paired with proper internal tooling and fine-tuning. A concrete example of this dropped in yesterday's AI Breakfast newsletter. GLM 5.2, a brand new open weight model out of China, is actively beating GPT 5.5 on design and UI tasks, while sitting neck and neck with Claude Opus 4.8 on complex long horizon engineering benchmarks. Now, you can't run this casually on a standard laptop. It demands a massive hardware footprint like 8H200 GPUs packed with a terabyte of aggregate VRAM, or a heavily quantized setup on a high-end Mac Studio, but the broader trajectory completely scrambles Washington's export control strategy. Choking off physical semiconductors only works if frontier capabilities are bottlenecked by hardware. Once elite models are freely distributed globally as open weights, the strategic battle shifts entirely to the governance layer, the deployment stack, and who controls the fine-tuning data. But while China plays the OpenWait Volume game, OpenAI is preparing to drop a massive counterpunch to reclaim absolute algorithmic dominance. The rumor mill is spinning out of control as developers and leakers share demos showing that OpenAI's unreleased GPT 5.6 Pro is already quietly running inside Select ChatGPT Pro accounts. Early testers are showing off the model spinning up OneShot 3D games and complex sim-style world builds out of a single HTML file, reportedly edging out Anthropic's Fable on single pass 3D design. OpenAI is keeping quiet, but insiders whisper a public launch is targeted for this Thursday, June 25th. Simultaneously, Sam Altman took the stage to directly challenge his loudest critics, specifically targeting Meta's Jan Lacun, who has spent years insisting that LLMs are a technological dead end. Altman doubled down on the scaling hypothesis, arguing that aggressively piling on more compute, data, and parameters continues to yield massive capability gains. To back it up, he pointed to a recent milestone where a model successfully generated genuine new knowledge by disproving a long-standing mathematical conjecture. Anthropics Dario Amade even joined him on stage to back up the claim. Altman conceded that LLMs still struggle with complex judgment and long-term planning, but his core thesis was clear. The people betting against scaling are consistently wrong. To prove the point, OpenAI is also rolling out GPTBD. This bidirectional voice architecture is hitting mobile users right now, allowing the model to interrupt you mid-sentence, track background conversations in real time, and organically correct its own verbal slip-ups. The transition from clunky, turn-based prompting to low-latency conversational AI is moving faster than the market has internalized. As these commercial capabilities accelerate, European regulators are realizing they have no choice but to give enterprise teams a bit of breathing room to figure it out. European lawmakers have reached a provisional agreement on what they're calling the AI omnibus, a legislative package that significantly delays the compliance deadlines for the EU AI Act. High-risk AI systems deployed in critical areas like hiring, education, and credit scoring have been granted an extension until December 2027. If those high-risk systems are embedded inside regulated physical products, like medical devices, the hard compliance deadline pushes out even further to August 2028. For enterprise teams building across Europe, this provides a massive runway to get governance frameworks aligned. However, don't confuse this delay with regulatory weakness. Brussels is tightening the screws on immediate harms. Strict, non-negotiable bans on AI-generated non-consensual intimate imagery and child exploitation material are bypassing the extensions and kicking in this December. The timeline became realistic, but the regulatory trajectory remains completely locked. And while Europe works to regulate digital software compliance, NVIDIA is stepping in to build the definitive safety architecture for physical autonomous robots. Nvidia just launched Halo's for Robotics, a full-stack safety system designed for the physical AI units navigating warehouses, factory floors, and public spaces. The ecosystem is incredibly thorough. It packages a dedicated operating system called Halo's OS, a camera-based exterior perception layer called the Outside in Safety Blueprint, and a fully accredited third-party inspection lab to give robot makers a formalized safety certification before deployment. Agility Robotics is the first out of the gate to adopt it, integrating HALOS into their digit humanoid robots. You might remember Digit from earlier this year when Figure AI's competing hardware clocked 200 consecutive hours of live warehouse operations. This matters far beyond the logistics sector because certification bottlenecks have historically been the ultimate graveyard for physical AI deployment, even when the mechanical hardware works perfectly. Standardizing a certified safety stack completely shifts that calculation, while handily ensuring that the entire software layer of the robotics industry gets tethered to NVIDIA infrastructure. But when you strip away the corporate talent wars, the geopolitical chip bands, and the warehouse automation metrics, you find the real reason this technological leap matters. OpenAI's O3 reasoning model is quietly being deployed to help families unlock diagnoses for incredibly rare childhood disease conditions that have eluded teams of human doctors for years, and sometimes for the majority of a child's life. The model pulls off a feat a single physician physically cannot match. It exhaustively cross-references massive medical literature across thousands of global case studies to surface rare genetic anomalies matching a specific hyper-niche symptom presentation. For rare conditions affecting one in 100,000 people, the average diagnostic delay is an agonizing seven years. That is seven years of a family watching a child suffer while the medical clock runs down. By compressing that multi-year diagnostic nightmare into a matter of weeks, this software is actively saving lives. We spend an immense amount of time tracking the competitive bloodbath of this industry. Who poached which researcher, which model built the 3D game faster, how Brussels is moving deadlines. But instances like this are a stark reminder of what happens when this raw computational capability is pointed at families who have completely run out of options. And that's it. If you have any feedback about this show, you can email Mike at yesterdayNai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I will see you tomorrow.