The 1 Billion User Milestone, On-Device AI Upgrades, and OpenAI's $7B Buyback

Yesterday in AI  |  13 August 2026

The 1 Billion User Milestone, On-Device AI Upgrades, and OpenAI's $7B Buyback

Consumer AI scale, hardware privacy indicators, and model security flaws intersected this week. This episode breaks down Google Gemini hitting 1 billion monthly active users as Google launches the Pixel 11 powered by the 2nm Tensor G6 chip for on-device execution and a "HiLight" privacy indicator.

We examine Apple's iOS 27 beta "Apple Reference Image" photo provenance system, analyze a zero-click Zoom vulnerability alongside security researchers decoding 315,000 "scrambled" AI reasoning logs, cover OpenAI COO Brad Lightcap's departure and its $7B employee stock buyback at an $852B valuation, and look at Google's AMIE diagnostic medical AI matching primary care doctors in video trials.

Plus: A special announcement regarding our upcoming interview with Bianca Baumann, VP of Learning Solutions & Innovation at Ardent Learning!

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 and welcome back to another edition of Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Thursday, August 13th, and AI just quietly crossed the line from cool new thing to the thing that's already in your pocket, your camera roll, and your work chat. A billion people, a new phone that runs your errands, and a growing pile of fine print about whether you can still trust what you see. Let's get into it. Let's start with a number that's hard to hold in your head. On Tuesday, Google said its Gemini app crossed 1 billion monthly users. Not Gemini sprinkled across search and Android and your thermostat, the actual app, the one you open on purpose. Google says it's the fastest anything has grown in the company's history, and this is a company that makes just a few popular things. A couple of details stuck with me. About 63% of those people talk to Gemini out loud instead of typing. Roughly 100 million of them are on iPhones, which has to sting a little over in Cupertino, and the thing spits out more than 150 million images a day, which explains why your group chat is now 40% real photos and 60% of your cousin turning the family dog into a Renaissance oil painting. One more thing that hints at where this is heading. About one in five Gemini live sessions now involves the camera or screen sharing, not just talking. People are starting to point the app at the actual world and ask it what they're looking at. Here's the honest read. A billion users is real, but it's a product-specific scoreboard, and Google is very good at making a number look enormous. It doesn't mean a billion people love Gemini or even remember downloading it. What it does mean is that the consumer AI race stopped being about who has the clever demo. It's about who's already on the phone in your hand. And once you're at a billion, every voice command feeds the next version. Which brings us to the phone Google would very much like that to happen on. Because on Wednesday, Google got on a stage in New York and showed off the Pixel 11, the 11 Pro, and the Pro XL. There's a new chip called the TensorG6 built on a 2 nanometer process, which is a fancy way of saying the transistors are packed so tightly they're basically whispering to each other. Google says it runs AI in the phone itself up to 3.5 times faster while sipping less battery. That on the phone itself part is worth slowing down for. Instead of shipping your request off to a data center in Nevada and waiting for an answer, the phone does the AI work right there, offline. Faster, more private, and it keeps working when your signal doesn't. The pitch is that Gemini can now take multi-step actions across more than 40 apps. There's less here's a search result and more book it, text them, and pull up the ticket. My favorite bit is a tiny LED. The Prophones get something called highlight, a little light that glows when Gemini is listening, thinking, or answering. Google clearly noticed that my phone's assistant is always awake makes people a little twitchy, so they gave the microphone a nightlight. It's a small thing. It's also the whole game right now, which is companies scrambling to prove the AI isn't doing something behind your back. Pre-order start at $899 and climbed to $1,299. The phones ship August 20th, and yes, that's a hundred bucks more than last year. Because, of course it is. That little honesty light is a good setup for what Apple's cooking, because Apple is worried about the opposite problem. Not the AI inside your phone, but the fakes coming to it. Buried in the iOS 27 beta this week, people found code for something called Apple Reference Image. The idea is that when you shoot a photo in a new reference mode, the iPhone stamps it with Provenance data. Basically a sealed receipt that says this came off a real camera sensor at a real moment. Tap a badge later and it checks with Apple's private cloud compute to verify, without Apple actually rifling through your photos. It's off by default, and you have to switch it on yourself. Think of it like a tamper-proof seal on a bottle of pills. It doesn't stop anyone from making a fake. It just gives you a fast way to check the real one. In a world where anybody can conjure up a convincing photo of a flood, a bomb crater, or a politician doing something they never did, prove this actually happened is turning into a real feature. It's the same instinct behind Claude quietly watermarking its text earlier this week and the EU's new labeling rules. Everybody is suddenly in the business of proving what's real. Picture the everyday version, the photo you send an insurance adjuster after a fender bender, the listing photos on a marketplace app, the image a news outlet runs during a breaking story. A verifiable this is real stamp actually means something in all of those. The catch, as always, is that a seal only helps if people bother to check it. And let's be honest, most people won't or won't remember. Checking under the hood matters though, because this week was a rough one for the tools you already trust. Start with Zoom. Researchers at a firm called A Security found a nasty flaw in Zoom's annotation feature, the little drawing tool that lets you scribble on a shared screen. Turns out that Digital Sharpie could be used to run code on someone else's machine mid-meeting. No clicking a bad link, no warning, nothing. They call that zero click, which is about the scariest phrase in security because it means you did everything right and still got hit. Here's the part that should make you sit up. The researchers say they built a working version of this attack in about a day, using fewer than 20 prompts on ordinary, publicly available AI models. This is the kind of thing that used to take an elite team months. Zoom pushed out a patch on Tuesday, so go update your app tonight. I mean it. Now the twist, and this one's almost poetic, a separate group of researchers went after the AI models themselves. When a chatbot reasons, it generates a private train of thought, and the providers hand that back to your app in an encrypted, scrambled form that you're not supposed to read. These researchers found they could take that scrambled reasoning, replay it into a weaker version of the same model, talk that weaker one into transcribing it, and outpop the hidden thinking in plain English. In one sweep of roughly 6,700 public AI sessions, they decoded more than 315,000 of these supposedly private reasoning blocks and pulled out real passwords and API keys. The secret credentials software uses to log into other software, sitting inside those blocks. So the AI's inner monologue, the stuff labeled Don't Look at This, was quietly leaking secrets. The providers say they've patched it, but it's a good reminder that hidden and safe are not the same word. Speaking of things happening behind the curtain, let's talk about the biggest name in AI and a very expensive week of housekeeping. OpenAI had two moves land back to back. First, on Tuesday, Brad Lightcap, the company's chief operating officer since 2018, said he's leaving to start something new. Lightcap is one of the people who turned OpenAI from a research lab into an actual business, and he's the latest in a string of senior folks heading for the exit, following names like Fidisemo, Kevin Weill, and Bill Peebles, all out the door in recent months, all as the company gears up for a possible IPO, its debut on the stock market. The second move is the eye-popping one. The day before, OpenAI finished buying back about $7 billion worth of stock from its own employees. This is called a tender offer, which is a formal, hey, want to cash out some shares? at a set price, in this case pegged to the company's $852 billion valuation. So without ever going public, OpenAI just turned another wave of paper millionaires into actual millionaires. Why spend $7 billion doing that? Two reasons, and they're both about people. When your competitors are dangling giant pay packages, letting your staff take some money off the table keeps them from walking, and it buys OpenAI room to stay private a while longer instead of rushing onto the stock market. It's a strange kind of flex. Spending a small country's budget so your engineers don't quit, and you don't have to answer to Wall Street just yet. The AI boom isn't only models and benchmarks. A lot of it is a very high-stakes game of keeping the right people in their chairs. Let me leave you with a better use of all this, because it wasn't all leaks and billion dollar cash grabs this week. Google spent its week showing off new phones and a billion users, but the thing that actually stuck with me came out of its research lab. It's called Amy, and it's an AI built to be a doctor you talk to over video. We've had AI symptom checkers for years, the kind where you type my throat hurts and it calmly informs you that you might be dying. Amy is a big step past that. It runs on Google's Gemini models. It watches and listens over a live video call, and it can look at you, read your expression, notice how you're moving, and even walk you through a basic self-exam, all while reasoning through what might be wrong. Google ran a real test on it. They had Amy handle 300 video visits with trained patient actors, then had 30 board-certified primary care doctors grade the results. It scored on par with, and on some measures better than, the human doctors on the stuff that counts, taking a proper history, landing on the right diagnosis, and suggesting a sensible plan. The patients even said they preferred how Amy assessed and explained their condition. Though, and I love this detail, the human doctors still won on bedside manor, the warmth and the rapport. So the machine out explained them, and the people outcomforted them. Now the honest part, because you know it's coming, this is a research study. You can't book an appointment with it yet. Actors aren't real patients, a controlled test isn't a chaotic clinic, and Google itself is waving the caution flag here. But think about who this actually helps. Somebody hours from the nearest doctor or stuck on a months-long wait list or just too anxious to pick up the phone. A calm, competent second opinion at 2 a.m. isn't nothing. After a whole show about AI's cracks and leaks, it's a nice change to go out on the version that might actually help you feel a little better. And that's the show. But stick around for a second because I have a quick announcement. I'm doing a special interview on Friday with my guest Bianca Baumann, VP of Learning Solutions and Innovation at Ardent Learning, for an extended episode that will come out on Monday. We're going to discuss why so many AI deployments stall out after the initial launch and possible solutions on how to fix that. It's a topic I really want to dig into. I hope you'll enjoy it. Now, if you have any feedback for me, email Mike at yesterdayNai.news or connect with me on LinkedIn, X, or Blue Sky. If you enjoy Yesterday and AI, please take a minute to rate and review the podcast wherever you listen, and maybe share it with a friend.