Apple Finally Shows Up...With Their Own Models: WWDC, AI Ownership, and the Week Washington Changed the Rules

Yesterday in AI — Weekend Recap | Tuesday, June 9, 2026

Apple Finally Shows Up...With Their Own Models: WWDC, AI Ownership, and the Week Washington Changed the Rules

Apple had a lot to prove at WWDC 2026, and for the first time in two years, it delivered. The new Siri is a standalone app, running on Apple's own Foundation Models (five of them, built with training help from Google Gemini but containing zero Google code). Visual intelligence and systemwide control are baked in. It's the most competitive Apple AI product yet, reaching iPhone 15 Pro, iPhone 16, and later. That's the headline. But today's episode goes deeper.

We cover the Trump administration's move to negotiate an equity stake in OpenAI, and potentially the broader AI industry, through a "Public Wealth Fund" designed to give every American a financial stake in the AI boom. We look at the joint letter that brought Sam Altman and Dario Amodei to the same table to ask Congress for bioweapons guardrails. We break down what Wall Street quietly did to its junior analyst hiring classes. And we get into the $400 million bet on "physical AGI" and the AI agent systems that now fix their own mistakes.

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 0:05

Hi folks, 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 9th, and Apple finally showed up to the AI party. Everyone said Google was driving. The models are entirely Apple's. We've also got government taking ownership stakes in AI companies, Wall Street quietly gutting its junior analyst classes, and a set of competing AI labs that apparently still fear bioterrorism enough to write Congress together. Let's get into it. The single biggest story yesterday had to be WWDC, Apple's annual developer conference. After two years of promises and a Siri that remained stubbornly mediocre despite everything happening around it, Apple stepped on stage and delivered. The new Siri is a standalone app. It has visual intelligence. It can see what's on your screen or in front of your camera and act on it. It has system-wide control, so it can actually reach into your apps and do things rather than just talking to you about doing them. And the whole thing runs on Apple Foundation models, Apple's own AI trained with help from Google Gemini, but running without a line of Google code. Sit with that for a second. For months the narrative was that Apple had surrendered and Gemini would be running under the hood of the new Siri. What actually happened? Apple used Gemini to help train their models through a process called distillation. Think of it like using a better student's notes to study than writing your own exam. The finished models are pure Apple. When you talk to Siri, you never touch a line of Google code. Craig Federigi confirmed it on stage. Apple Foundation models run on Apple Hardware or Apple's private cloud compute servers, and your data never reaches Google. Apple built five foundation models for this new architecture, two on-device models that run privately on your iPhone, and three server-based models that handle heavier tasks in private cloud compute. The most powerful server tier runs on Google Cloud and NVIDIA GPU infrastructure, though it stays private cloud compute certified. Third-party AI integrations are coming as opt-in extensions. ChatGPT was already part of the ecosystem from last year, and Apple's new extensions framework is designed to support others over time. There's also a new App Store category for third-party AI agents, software that takes actions on your behalf, not just answers questions, all running inside Apple's privacy rules. One important caveat: Apple Intelligence requires an iPhone 15 Pro, iPhone 16, or newer. Not every iPhone in the wild gets this, but the newest hardware across hundreds of millions of devices does. Two years ago, Apple overpromised and underdelivered. This time the features look baked, using Gemini to train their own models filled the capability gap Apple couldn't close alone, and the distribution is enormous. Whether Siri becomes as capable as ChatGPT or Claude in daily use is still an open question, but the architecture is now competitive for the first time. Apple spent Monday making one thing clear. The AI on your iPhone is Apple's, built on Apple's terms. Washington is working on a different kind of ownership. OpenAI and the Trump administration are in active negotiations over the U.S. government taking an equity stake in OpenAI. We're talking 1-5%, routed into what they're calling a public wealth fund, designed to give ordinary Americans a direct financial stake in the AI economy. Axios reported the concept doesn't stop at OpenAI. The White House is thinking about this more broadly. As AI companies create trillions in value, the government could negotiate equity positions that flow back to citizens, rather than concentrating in venture capital and institutional investors. Sam Altman met with Trump officials directly. Separately, on June 5, Trump signed National Security Presidential Memorandum 11, directing the military and intelligence agencies to move faster on AI adoption across war fighting and intelligence operations. Both moves signal the same thing. Washington is done watching from the sidelines. The government wants ownership commercially and deployment aggressively on the defense side. Six months ago, the DC conversation was mostly about regulation and safety guardrails. Now it's about equity stakes and battlefield acceleration. Whether that's smart governance or a concerning entanglement of government power and private AI companies, is worth debating. What's clear, the relationship just changed character. That's Washington moving toward AI. On June 3rd, the AI companies went the other direction. OpenAI and Anthropic signed the same letter. More than 70 AI leaders, including Sam Altman and Dario Amode, put their names on an open letter to Congress calling for mandatory tracking and screening of all synthetic DNA and RNA purchases. The coalition, organized around a group called Screen DNA, wants something like a pharmacy system for genetic material. Every order logged, every buyer verified, every transaction auditable before the material ships. The reason this is happening now is simple. Advanced AI dramatically compresses what a bad actor needs to know and how fast they can move from I want to engineer a dangerous pathogen to actually having one. The letter cites a 2017 case where Canadian researchers spent $100,000 on mail-order synthetic DNA and reconstructed horsepox, a close cousin of smallpox from scratch. That was with 2017 tools. The worry is obvious. The ask is specific. Pharmacy-style record keeping for synthetic DNA and RNA. Mandatory, nationwide. And the argument is partly deterrent. Awareness of traceability itself deters misuse. When you have to sign your name, behavior changes. What makes this notable beyond the policy ask is who signed it. These are companies fighting for the same enterprise customers, competing for the same talent, and outright hostile to each other in the market. When they show up together, pay attention to what they're actually afraid of. The screen DNA letter is about AI making dangerous expertise easier to reach. Wall Street is dealing with the same logic, just applied to spreadsheets instead of biology. Major banks are cutting their junior analyst and associate intake classes by up to two-thirds. Not over a decade, in one or two recruiting cycles, the work that used to fill a first-year analyst's 80-hour week, financial models, deal presentations, valuation analyses, formatted decks, AI tools now do at a fraction of the time and cost. Banks aren't replacing that work with a different kind of entry-level job. They're just not hiring for it. This matters beyond finance for a specific reason. The analyst program at a bank was one of the most reliable paths into upper middle class careers for college graduates without a technical background. A two-year program that launched 20 to 30 year careers in private equity, venture capital, and corporate strategy. If that on-ramp compresses by 66%, the downstream effects on professional class mobility are significant, and largely invisible in the standard job loss metrics. We've tracked AI workforce displacement throughout this show, but most of those cases involved layoffs of existing employees. What's different here is new hiring collapsing. Companies aren't firing people and replacing them with AI. They're just not replacing the people they would have hired. Harder to track, harder to regulate, and probably more significant long term. From the finance floor to the factory floor, a startup called Generalist, backed by Nvidia, just raised $400 million at a $2 billion valuation to build what they call physical AGI. The idea is a robot brain that generalizes across tasks the way a human can. You show it something new, it figures it out. You don't have to retrain it for every environment or object type. Generalist joins physical intelligence, rota AI, and skilled AI in chasing the same prize. And a company called Luma just launched the Open Physical AI Lab, a collaborative initiative arguing that generalization is too hard and too important for any single company to solve, so the foundational work should be built in the open. Luma's CEO put the stakes plainly. If generalization works and robots can operate across work, business, and home environments, they're going to become the means of production. The humanoid robot deployments we've been tracking, Figure AI's warehouse runs, Gatsby's Home Cleaning app, are purpose-built for specific tasks. Physical AGI is the layer above all of that. Crack generalization and every purpose-built application gets dramatically cheaper and faster to deploy. Finally, a team of OpenAI engineers embedded at enterprise customer sites have been working on agents that can learn what they don't know yet, and they have something running in real-world deployments. The project shows AI agents that learn from their own mistakes in real time. The system reviews its own outputs, identifies failure patterns, and rewrites its internal workflow to perform better on the next run without a human stepping in to retrain it. Anthropic's June 4th post was about who's actually writing the code. Claude now authors more than 80% of Anthropic's production code. This OpenAI project works in live enterprise workflows, agents reviewing their own outputs, finding where they failed, rewriting the process. Both are pointing the same direction, AI iterating faster than any human work cycle. Right now, when an AI agent makes a mistake in a business workflow, a human has to notice it, diagnose it, and either retrain it from scratch or rewrite the instructions it's working from. If agents can identify and correct their own failure modes, that cycle collapses. Improvement measured in hours rather than weeks. Six months from now, this might be just how software works. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterday inai.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 moment 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'll see you tomorrow.