Tech Brew Ride Home - Big Google AI Shakeup

Episode Date: August 5, 2026

Google shook up its AI leadership, kicking Demis Hassabis upstairs while Jeff Dean and Sanjay Ghemawat left to found Discovery Loop. SpaceX's first earnings spooked investors, a UK-tested AI agent went rogue, and Disney let TikTok fans into Disney+. Links Google just announced a major shakeup of its top AI leadership (The Verge) SpaceX reports Q2 revenue up 92% YoY to $7.8B, vs. $6.81B est., AI operating loss of $1.26B, vs. $2.39B est., says capex in Q3 and Q4 will remain similar to Q2 (Bloomberg) An AI agent went rogue during UK safety tests, creating fake identities and launching social engineering attacks unprompted (The Decoder) Disney announces a global deal with TikTok to bring "thoughtfully curated" fan-created short-form videos based on Disney's IP to Disney+'s vertical Verts feed (The New York Times) Subscribe to the ad-free feed.

Welcome to the TechBoo ride home for Wednesday, August 5th, 2026. I'm Brian McCullough today. Google shook up its AI leadership kicking Demas Hasabas upstairs while Jeff Dean and Sanjay Gamwat left to found a startup. SpaceX's first earnings as a public company spooked investors. Now a UK tested AI agent went rogue as well, and Disney is letting TikTok into Disney Plus. Here's what you miss today in the world of tech. This is a woe headline from Google DeepMind, quoting the very very Google is making some significant AI leadership changes, including a major shift for Google DeepMind leader Demas Hasabas. Hasabas will become the chair of Google DeepMind and the chief scientist at Alphabet. CEO Sundar Pichai announced on Wednesday, Hasabas will continue to lead Alphabet's isomorphic labs, which aims to use AI to develop drugs. Coray Kavukulu, formerly DeepMind's CTO, is stepping up to become Google's SVP of deep mind and will report to Pichai. Kavuchkulow will remain Google's chief AI architect as well. As you've heard me say, many times, I've always believed the number one application of AI should be to

improve human health, Hasabas says in a message to staff, it's time for AI to prove its unequivocal value to the world, and what better way to demonstrate that than to help finally cure diseases like cancer. Hasabas says he will continue to work closely with Sundar on strategic and global AGI matters and to advise Kavukulu and other DeepMind leads. In another big change, veteran Googler Jeff Dean, the chief scientist of Google DeepMind and Google's 30th employee, is leaving with Google fellow Sanjay Gamowat to form an AI-focused public benefit corporation called Discovery Loop. Google will be a founding investor in the new company, Pichai says. Our mission is straightforward. We are building AI solutions that can automatically solve

important problems in machine learning, science and engineering discovery loop says on its website, by advancing the pace at which we conduct engineering and scientific discovery, we can bring the benefits of science and technology to the world much faster. Ultimately, our goal is to build AI systems that act as a deeply positive empowering force for humanity, delivering technology solutions that improve people's lives on a global scale, end quote. Discovery Loop is apparently backed by seed funding from radical ventures, Costa Ventures, and, of course, as I mentioned, Google has a stake in the company as well. More on this quoting Wired. On July 25, Jeff Dean spoke to 6,000 would-be founders at Y Combinator's Startup School at San Francisco's

Chase Center. Dean is a legendary engineer at Google. He is co-founder of Google Brain, chief scientist of Google DeepMind and Google Research, and technical co-lead on the company's flagship AI model Gemini. With his frequent collaborator, Sanjay Gemmwat, he's written significant parts of the company's computing and search infrastructure, a running joke in geekdom is a list of surreal facts about his technical prowess. Jeff Dean's resume lists the things he hasn't done. It's shorter that way. At Chase Center, the interviewer asks him what areas startup founders should think about tackling. Dean said that the most important thing would be to pick something you're excited about and would be useful in the world. He then shared his own interest, an automated version of the scientific method. You propose an experiment, you implement, you implement, what you need to run the experiment, you evaluate the experiment, and you get results from that,

he said. Dean said that running thousands of those automated loops could unlock groundbreaking advances in multiple domains like science, biology, chip design, and even making improved AI models. Unbeknownst to the crowd that day, Dean was talking about his own secret startup, which he had just organized. Today, it's official. After almost 27 years, Dean is leaving Google, along with Gammat and two other top-tier AI scientists to found a company called Discovery Loop. Even though Google will take a stake in the new company, it's a devastating blow to the search giant as it attempts to keep pace in the frantic AI model competition. It's bad enough that Dean and Gemmowat, who were among the company's first hires, are leaving. That's Mick Jagger

and Keith Richards, ditching the Rolling Stones to start a new band. Worse, the other co-founders are also supernovas in the AI firmament. Oriole Vinyals, VP of Research at DeepMind and a technical lead for Gemini and Kwak Lee, a founder of Google Brain and the key scientist behind Google's AutoML Zero, a project that uses machine learning algorithms to autonomously build AI products. In an interview I did with the co-founders just before launch, Dean tells me that the idea came up only a few weeks ago. He was quickly joined by the other three who are not only longtime collaborators, but friends who vacationed together. We were working on different things, but we were all starting to see the possibility of AI being able to automate scientific and

engineering loops, he says. The Quartet worked up a roadmap where in discovery loop would be its own first customer. The autonomous experiment loops they create initially will be directed toward improving the company's advanced machine learning algorithms. That improved software will then become the core of an AI system that can tackle problems in subsequent domains. I'm very excited about the automating machine learning, says Lee. It might be that we will discover a different transformer architecture. From there, discovery loop will set out to make superhuman advances in a variety of fields, including chip design, biology, drug discovery, and material design. As we create the core intelligence and the discovery loop fundamentals, will be able to

quickly generalize to other domains, says Vinyals, one of the things that will be obviously very focused on is how these models come up with new ideas to try. That's not something that currently they're super strong at. At first, the team envisions that those ideas will be co-developed with humans, but their goal is deeply automating the process. If it works out, they contend, it's possible that a few dedicated individuals may use Discovery Loop to out-invent the world's largest research organizations. Presumably, those small teams will be customers of Discovery Loop. The startup may also make its own loops to seek and monetize breakthroughs. They put together a simple pitch deck and set out to talk to VCs about funding their company, which they set up as a

public benefit corporation. We just put a few slides with our backgrounds and a rough sketch of how we think of our approach and how it will be successful in a bunch of different ways, says Dean, but their presence spoke louder than any PowerPoint could. With this team, I wouldn't need to know what they were doing before I back them, says Venture capitalist Vinod Kosla, who met with them in his Sand Hill Road office on a Saturday, so no one would get a hint of Google's upcoming loss. It's the ultimate superstar team. Kostla also love their idea.

Humans have been using AI to do research, not AI to be a researcher, he says. The fundamental thing in Discovery Loop is that AI is the researcher, end quote. So you want some tea leaves. I've got some for you to read, quoting Peter Liu. The magnitude of this is hard to understate. The Gemini leadership left, Nome left earlier, Demis is no longer head of deep mind. More talent will be leaving. The best thing Google has left is TPUs and a stake in these new companies, end quote. Dan Shipper says something that lines up with that thing we did about a week or so ago, about the two routes to AGI and how Demis Hasabis comes down definitively on one side. quote, in order to be competitive today, Google needs to catch up on frontier coding. Demise believes different fundamental research directions like world models are more important

to his long-term goal, even if they're less important competitively today, end quote. Here's G. Moskowitz, quote, big moves at Google Deep Mind. Early Reid is, this is Demis being kicked upstairs and allowed to work on his special projects. Also, Jeff Dean is stepping away. I wonder if any of this is related to Google's decisions on the Department of War. contract, end quote. And here's Tenebrus. Demise is ousted as DeepMind CEO and Jeff Dean plus Sanjay are leaving Google to start

a NeoLab. They're all being very careful to frame these as positive shifts, but there's no way in hell Demise would have accepted this willingly, and there's no way Sundar happily accepted Jeff doing this as a totally independent new PBC rather than a bet under Alphabet. Tough to see an interpretation other than Jeff losing confidence in working on AGI under Google. very bad day for Alphabet overall, end quote. SpaceX reported earnings for the first time as a public company yesterday. How'd it go? Quoting Bloomberg. SpaceX shares fell after the Elon Muskled company disclosed higher than expected spending on its artificial intelligence business,

spoiling an inaugural earnings release that broadly exceeded Wall Street estimates. The rocket, satellite, and AI conglomerates capital expenditures soared to $28.5 billion in the first half of this year, more than four times it's spending a year earlier. That outlay is nearly as much as the amount raised in the biggest initial public offering ever before SpaceX's record listing in June. Musk is steering the vast majority of SpaceX's expenditures toward AI, a business that is growing rapidly, though not at the rate of all that spending. The company generated $3.39 billion in revenue from artificial intelligence during the first six months of the year, less than half, what its connectivity business centered around the satellite internet service Starlight,

link brought in during that span. SpaceX shares dropped as much as 13% shortly after the start of regular trading Wednesday. The stock has been volatile since its debut, first trading up 49% in a matter of days, only to fall below the IPO price within just over a month. The pace of SpaceX's spending caught some analysts off guard. Capital expenditures likely will jump to around $65 billion next year, Alexander Potter, a Piper Sandler analyst with the equivalent of a hold rating on the stock said in a note, that's about $17 billion more than he was anticipating. Potter wrote that unpredictable spending on AI and excessive CAPEX are among the risks to SpaceX's stock reaching his $140 price target, along with the release of hundreds of millions of locked-up shares.

SpaceX's revenue rose to $7.81 billion in the second quarter, topping the 6.81 billion average estimate among analysts surveyed by Bloomberg. The company posted a loss of $0.9 per share, also beating estimates, end quote. All I can do is want to sing that it's happened again song from all the soccer chance, quoting Decoder. During routine cybersecurity testing by the British AI Safety Institute and AI agent went rogue on the open internet. Without being told to do so, it created fake identities, tried to slip malicious code into an open source project, and targeted real people and organizations with social engineering attacks. AISI says this is the first time risks tied to AI autonomy and deception have appeared so clearly in the real world without specific

prompting. The incident caused no actual harm. It took place between July 25th and July 28th, out of the 122 test runs across seven models 10 showed problematic behavior. AISI recorded 19 unauthorized actions in total with 17 attributed to Anthropics Mythos 5 and 2 to OpenAI's GPT 5.6 sole. The models ran without the safety restrictions used in commercial products, so the test conditions don't reflect normal public use. But the results reveal what these models are capable of when those safeguards are stripped away. According to AISI, the agent tried to inject malicious code into a public open source project on GitHub to win approval from human reviewers. It researched the project's maintainers and created several fake online identities that it

then used in a coordinated deception. One fake account claimed to have reviewed the code and found no malware, while another thanked it for the supposedly independent review. The goal was to convince the real maintainer that the code was safe. When someone publicly questioned the proposed code change, the agent edited its earlier activity to make it look harmless and considered returning under a new identity. It used the Tor network to bypass GitHub network restrictions, but a human maintainer spotted the malicious code and rejected it. The agent also reached out to real people directly sending messages and files through an online file transfer service in an attempt to trick them or their AI coding tools into running malicious code.

Some of those messages contain malicious payloads. AISI documented two other types of behavior as well. The agent planted malicious instructions where automated AI systems could find and execute them, a technique known as prompt injection. One agent even posted public GitHub messages offering to collaborate with other agents being tested at the same time, explaining how they could reuse the accounts and artifacts it had left behind. Later agents found and used them.

According to AISI, the agent was simply pursuing its assigned goal of solving the cybersecurity task, just as in the open AI case. This is the classic alignment problem in AI, which philosopher Nick Bostrom once illustrated with his paperclip thought experiment, how do you make an AI system pursue a worthy goal without taking harmful steps to get there? The deception wasn't triggered by a specific instruction, AISI said. It emerged as a byproduct of the agent relentlessly chasing its objective. This kind of goal-driven deception had been largely theoretical until recently.

In some runs, the task was configured incorrectly, leading the agent to conclude that no solution existed within the stated limits. AISI says there are good reasons to think that nearly impossible tasks push models toward more creative boundary-crossing solutions, but the agents sometimes behaved this way even when the task was set up correctly. Other tests show that all frontier models are prone to this, with newer open AI models particularly likely to deceive. OpenAI co-founder and ChatGPT co-creator John Schulman thinks the answer to solving these issues may lie in how models are trained after pre-training. He points to

chunky post-training, a phenomenon where models develop distinct behavioral patterns tied to certain tasks types. When a model recognizes a cybersecurity evaluation as similar to reinforcement learning tasks where completion is the only reward, it pursues the objective single-mindedly. Safety behavior learned in other contexts don't carry over. Training on Capture the Flag, cybersecurity challenges could make this response especially likely during cyber tests, end quote. Finally today, SORA might have been a failure, but it seems it's given Hollywood some ideas. Maybe it's okay to let fans create art around your IP, quoting the times. Disney said on Wednesday that it had reached a global deal with TikTok to allow fan-created short-form video based on Disney characters and franchises to appear inside the Disney Plus streaming service.

The videos will be tightly vetted by Disney, thoughtfully curated in its words, and will appear in Verts, a vertical video feed that debuted inside Disney Plus this spring. Verts is an attempt by Disney to give people a reason to open Disney Plus, even when they don't have a specific movie or show in mind. The hope is that a few minutes of scrolling will turn into an hour of streaming. So far, Verts has featured only Disney produced clips from the company's movies and television shows. For TikTok, the deal is a chance to associate itself with a trusted family brand, as the

social media platform pursues its own user growth. Under the deal, TikTok will offer certain creators access to video clips related to hundreds of films and TV shows from Disney's library, including assets from the Pixar, Marvel FX, and Lucasfilm brands. A Disney spokeswoman declined to identify any titles, end quote. Well, new house, new studio, but hopefully this will sound like the same old sound quality. Talk to you tomorrow.