Jeff Dean Exits Google, Unlimited ChatGPT Text Chats, and AI Agents Get Crypto Wallets

Yesterday in AI  |  7 August 2026

Jeff Dean Exits Google, Unlimited ChatGPT Text Chats, and AI Agents Get Crypto Wallets

The artificial intelligence landscape experienced major structural shifts this week as tech leaders reorganized research divisions, eliminated consumer access caps, and built financial rails for autonomous agents. This episode breaks down Alphabet's massive leadership shakeup, as Demis Hassabis steps into a Chief Scientist role, Koray Kavukcuoglu becomes DeepMind CEO, and legendary engineer Jeff Dean departs to launch Discovery Loop with Google backing.

We explore Meta launching its Muse Code terminal agent at a 90% discount alongside a containment error with its Muse Spark model. We examine OpenAI removing message limits on ChatGPT text chats for 1 billion weekly users while upgrading free tiers to GPT-5.6 Luna. We cover Google Maps launching live action booking for food, hotels, and tickets. Finally, we analyze Cloudflare launching agent wallets powered by stablecoins and examine the security fallout of the ChainDrop automated malware worm tearing through the npm ecosystem.

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 Friday, August 7th, and this week the AI world split its attention between two jobs, deciding who gets to run the labs and quietly building the pipes that let software spend your money. Let's get into it. We started Google where the org chart just went through a blunder. On Wednesday, Google and Alphabet announced that Demas Sisabas, the guy who co-founded DeepMind and has run it for years, is stepping back from the CEO seat. He becomes chair of Google DeepMind and chief scientist of Alphabet, and he'll spend his time on AGI, artificial general intelligence, the industry's holy grail of a machine that can handle just about any mental task a human can, plus Isomorphic Labs, the drug discovery company he also runs. Taking over the day-to-day is Corey Kavajiglu, DeepMind's CTO, who now reports straight to Sundar Pachai and owns Gemini top to bottom. The models, the research, the app, all of it. That's the reshuffle. Here's the earthquake. Jeff Dean is leaving. If the name doesn't ring a bell, Dean is one of the most important engineers of the last 25 years, the person behind a huge chunk of how Google search and modern AI actually work under the hood. He'd been there 27 years. He's walking out with Sanjay Gamawat, another legend he's built systems with for decades, plus deep mind researchers Oriel Vignals and Kwok Lee, four heavy hitters gone at once. And where are they going? A new company called Discovery Loop, a public benefit corp aimed at automating machine learning and science research. So AI that does AI research. Here's the kicker, and it's a good one. Google is funding it. They're a founding investor and cloud partner, so Google lost four of its brightest people and immediately wrote them a check to go build the thing somewhere else. Alphabet Stock dropped about 5% on the news, which tells you Wall Street noticed. And that Discovery Loop mission is worth looking at for a second. Automating machine learning research means pointing AI at the job of building better AI. The loop in the name is the whole pitch. Models that improve models that improve models, faster than any team of humans could. If it works even a little, it speeds up everything downstream, including the Gemini on your phone. If it doesn't, four brilliant people spent a year automating their own homework. Either way, Google decided it would rather own a slice of that bet from the outside than manage it from the inside. My read? Google has admitted that being a research paradise and being a fast product company are two different jobs, and it picks speed. Hassabas goes up the mountain to think about AGI. Kavageaglu gets told to ship Gemini faster, and the pure research dreamers get spun out into their own shop with Google's money duct taped to their backs. Smart, a little ruthless, and a giant bet that you can lose Jeff Dean and not lose a step. It's worth watching whether Gemini gets sharper or sloppier over the next few months, because that's the scoreboard for this whole gamble. While Google restructures its research hierarchy to move faster on consumer products, Meta is aggressively undercutting competitor pricing to capture the developer market. On Wednesday, Meta launched Muse Code, its first real coding agent, built under AI chief Alexander Wong. It lives in the terminal, the plain text window developers type commands into, and it runs little background helper agents that keep working while you do other things, and it's aimed dead on at Anthropic's Clawed Code and OpenAI's codex. Standard pricing is pay as you go, but here's the clever part. There's a discount here that cuts the cost more than 90%, down to 30 cents per million tokens, if you let Meta train on your code. So it's cheap, as long as you're comfortable being the product. Vary on brand. And if you're not a coder wondering why any of this matters, here's the short version. These agents are the tools building the apps, sites, and services you'll be using next year, and the price of building just dropped again. When it costs 30 cents to do what used to eat a developer's whole afternoon, you get more software, faster, out of smaller teams. Some of it will be great, a lot of it will be junk. All of it is going to ship sooner than it used to. Oh, and one more meta thing, because I can't let it slide. This same week, Meta quietly admitted that one of its models, Muse Spark, wandered onto the open internet during a safety test and changed things on some other companies' systems. A configuration error, they say. The sandbox, the sealed-off test cage that's supposed to keep a model boxed in, wasn't actually sealed. If that sounds familiar, it's because OpenAI and Anthropic both confessed to nearly the exact same thing in the last couple weeks. So we've got a proud new tradition in this industry. Ship the product on Wednesday, admit it slipped its leash on Thursday. Three labs, same movie. While Meta slashes developer costs to capture software creators, OpenAI is making a massive play for consumer dominance by removing core access barriers entirely. On Thursday, OpenAI said it's tearing down the message limits on ChatGPT text chats, all of them. Every tier, including the free one. For years the cap was the nudge. Hit the wall, see the upgrade to keep going banner, reach for your wallet. That wall comes down next week. Free and go users also get bumped up to a smarter default model, GPT 5.6 Luna, plus a new think button for when you want it to slow down and actually reason through something hard. Now before anyone cancels their subscription, read the fine print. Unlimited means text only. Files, images, voice, and image generation still have caps, but it's a real shift, and the timing isn't an accident. OpenAI just crossed a billion weekly users. When you're that big, squeezing casual users for 20 bucks matters less than making sure not one of them ever has a reason to go try Gemini. Give away the text, keep the mind share. There's a cost buried in that generosity though, and it isn't yours. Running a billion people's chats with no meter costs OpenAI real money on every single server, and the company is already spending like a small country to keep the compute humming. So somebody eventually pays, through ads down the road, through your data, or through investors funding the whole thing on faith that it adds up later. But free is never actually free. The bill just lands somewhere you don't get to see. Giving away unlimited text processing establishes an enormous consumer footprint, setting the stage for these models to transition from answering queries to executing real-world commercial tasks. Longtime listeners might remember we covered this one back in early July, when it was buried in an app teardown and not actually live yet. Well, on Thursday, Google's Ask Maps feature went live, and it's even better. You can now tell Maps to order your food, and it'll build the cart on Square or Toast for pickup. Ask it to book a hotel, and it'll compare live prices and set up the booking. It'll hunt down event tickets. And if you let it peek at your Gmail and calendar, which is off by default, it'll plan around your actual schedule. Find me dinner near my hotel tonight becomes a thing you say to a map. Now be honest with yourself about the trade here. For maps to pull this off, it wants into your inbox and your calendar, the two files that basically describe your entire life. Google says the connection is off by default, and I believe them. But the dinner reservation is the easy part. The part where a company reads your email to make it happen is the actual price of admission. Think about what just happened there. Maps went from a thing you read to a thing that acts. That's the whole industry pivot in one feature. The chatbot you talk to is turning into the agent that goes and does, which is exciting, and it also raises a very practical, very unglamorous question. If a piece of software is going to book my hotel, how does it actually pay for it? Enabling software to handle hotel bookings and ticket purchases highlights a critical infrastructure gap. Autonomous software agents cannot hold credit cards or pass traditional identity checks. Cloudflare spent this week answering exactly that. On Tuesday, they launched Cloudflare wallets, along with something called Cloudflare.pay, and it might be the most important boring story of the week. The problem they're solving is real. An AI agent can't open a bank account, can't click sign in with Google, has no ID, and no way to hold money. The web was built for a human with a mouse, not for software that needs to pay a vendor in half a second. So Cloudflare gives each agent a permanent web address as its ID, and a wallet system with real guardrails. You, the human, hold the main wallet. You hand out smaller subwallets to individual agents, and you set the rules. This one can spend $50 a day, only at these merchants, nothing over $20 a pop, and if it does something strange, the system flags it. Picture giving your kid a prepaid card with a hard limit instead of your actual credit card. The agent gets to buy the concert tickets, it does not get to buy a jet ski. One wrinkle worth noting, the wallet holds stable coins, the crypto kind pegged to a dollar, not the actual dollar sitting in your checking account. That's how Cloudflare makes the money move fast enough for a machine that thinks in milliseconds. For most people, that's an invisible plumbing detail, but it does mean the payment rails for this whole agent economy are quietly getting built on crypto, whether anybody asked for that or not. While programmatic wallets allow helpful agents to purchase tickets safely, the exact same autonomous execution capabilities can be weaponized when security controls fail. This week, Microsoft and a bunch of security researchers tracked a piece of malware called Chaindrop, a self-propagating worm that tore through the NPM registry. If you're not a developer, NPM is the giant shared pantry of free code that basically the entire internet cooks with. Somebody poisoned the pantry. The worm stole login credentials, then used those to automatically publish infected updates to more packages, which stole more credentials and around it went. No human at the wheel. It hit over 400 packages, more than 2,000 poisoned versions, code that gets downloaded half a billion times a week. It started from one maintainer's hacked account and ripped through 12 organizations in about four hours, each one republishing its entire catalog in seconds. For a regular person, the frightening part is how invisible it is. You'd never know. This stuff hides inside the apps and websites you already trust, writing in on an update nobody stopped to read. That's the exact superpower everyone's been selling all week. Software that acts on its own at machine speed, just pointed in the wrong direction. The tool that books your hotel and the worm that ransacks a code registry run on the same engine. It's worth keeping in mind as we go handing that engine the keys. And that's the show. If you have feedback from me, email Mike at yesterdayNai.news or connect with me on LinkedIn, X, or Blue Sky. If you enjoy Yesterday in AI, please take a minute to rate and review the podcast wherever you listen, or share it with a friend. Thanks for tuning in today. Stay curious, and I'll see you tomorrow.