Tech Brew Ride Home - Tue. 07/08 – Zuck Poaches AI Talent From Apple

Episode Date: July 8, 2025

How much further behind can Apple get in AI now that Zuck is poaching from them as well? OpenAI has been forced to batten down the hatches, quite literally. A fully licensed AI video model. And back to Apple. They heard your complaints. They’re pumping the brakes on Liquid Glass a bit. Sponsors: AGNTCY.ORG Links: Apple Loses Top AI Models Executive to Meta’s Hiring Spree (Bloomberg) OpenAI’s Stock Compensation Reflect Steep Costs of Talent Wars (The Information) OpenAI clamps down on security after foreign spying threats (Financial Times) Bluesky can really keep up with the news now that it has activity notifications (The Verge) This New AI Tool Wants to Work With Filmmakers—Not Replace Them (Time) iOS 26 beta 3 dials back Liquid Glass (TechCrunch) Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the Tech meme right home for Tuesday, July 8th, 2025. I'm Brian McCullough today. How much further behind can Apple get an AI now that Zuck is poaching from them as well? Open AI has been forced to batten down the hatches quite literally, a fully licensed AI video model, and back to Apple. They heard your complaints. They're pumping the brakes on liquid glass a bit. Here's what you miss today in the world of tech. Well, here we go again. Mark German is reporting that, Rowling Pang, the head, of Apple's Foundation Models AI team is leaving to join Meta, which offered a pay package worth tens of millions of dollars per year. Quote, at Apple Pang had been running a roughly 100-person team

responsible for the company's large language models, which underpin Apple Intelligence and other AI features on the company's devices. In June, Apple announced that those models would be opened up to third-party developers for the first time, allowing for a range of new iPhone and iPad apps. but internally, the Foundation Models team has come under scrutiny from new leadership, which is exploring the use of third-party models, including from either OpenAI or Anthropic, to power a new version of Siri. Those internal discussions have soured some of the morale on the Foundation Models team, also known as AFM, in recent weeks.

While the company has explored a move to third-party solutions to power the AI and the new Siri, it has simultaneously been working on a new version of Siri based on, the models developed by Pang's group. Those models also power Apple Intelligence features that run on Apple devices, including email and web article summaries, Gen Moji, and priority notifications. The major departure, the most significant in Apple's AI ranks since the company started working on Apple Intelligence a few years ago, underscores the heightened competition for talent in the emerging space. Meta has been making offers to the world's top engineers worth many millions of dollars per year, significantly more than what the iPhone maker pays its engineers doing similar work.

departure could be the start of a string of exits from the AFM group, with several engineers telling colleagues they are planning to leave in the near future to meta or elsewhere, the people said. Tom Gunter, a top deputy to Pang, left Apple last month, Bloomberg reported at the time. The Foundation Models team reports to Daphne Luong, a top deputy to AI senior vice president, John Giann Andrea. Earlier this year, Giann Andrea was sidelined internally and saw Siri, robotics, core ML, and app-intense frameworks and other consumer product-related teams stripped from his command.

That came after a poor response to Apple intelligence and continued delays for new Siri features, including the ability to tap into user data to fulfill commands, end quote. So the summing up of the situation here is Apple is already behind an AI, and now the Mark Zuckerberg Money Gun is brain-draining them as well. It's probably not just the money on offer, though, because honestly, if you're a top AI talent, would you have confidence Apple is going in the right direction with AI at the moment? And also, as Emil Protolensky tweeted, meta has been turned down by so much AI talent that it's now poaching from Apple, end quote. As at Aaron P613 tweeted, at least Apple doesn't need to worry about him leaking Apple's AI strategy secrets since there isn't anything to leak, end quote.

And Aaron Rao on X, quote, Rowming and his team did really great work. The problem at Apple is cultural and structural around GPUs, data access, privacy slash secrecy, and fast launches to iterate on software. The only way to fix it is to buy Anthropic or Mistral plus perplexity and make big changes to the culture the team can operate in. Apple still has time, but the window is closing. It's unlikely they move with the boldness of Zuck, end quote.

Meanwhile, over at OpenAI and on the whole, are they scared tip? OpenAI apparently told investors stock compensation jumped over 5x in 2024 to $4.4 billion or 119% of its total revenue. It does expect that to drop to under 10% by 2030 because I guess they figure they'll be making a ton of money by then. But still, notable that they're having to dip into their scarce capital just to keep people happy. quoting the information. That proportion, a common way investors track stock compensation was projected to decline to a still high 45% of revenue this year.

But as revenue soars at the end of the decade, the chat chepti-maker expects stock-based compensation to fall to under 10%, according to projections seen by the information. OpenAI made those forecasts earlier this year before meta-platforms hired at least nine of its researchers prompting OpenAI chief research officer Mark Chen to indicate the chat GPT maker could get even more aggressive with its stock offers. OpenAI's stock expense, which hasn't previously been reported, shows the high-price AI labs are paying to attract and retain workers, particularly compared to past generations of tech companies. It also points to how much of the company employees could own after Open AIs for-profit unit, which is governed by a nonprofit,

converts to a public benefit corporation that can issue shares like most companies. Currently, employees receive profit units that give them a share in open-a-iase. Open AI's eventual profits rather than typical stock awards such as restricted stock units or options. After the restructuring, employee stakes are slated to convert into common shares. Open AI leaders have discussed a scenario in which employees will own roughly a third of the restructured company, according to a person who spoke to OpenAI executives. In that scenario, Microsoft would own another third, while investors and the nonprofit that governs OpenAI would share the remaining equity.

The talks are ongoing, and it couldn't be learned whether the company's leaders are still considering that scenario. Privately held companies rarely make stock compensation costs public unless they're on the cusp of an IPO, making it difficult to compare OpenAI's expense to other companies of a similar size. But the securities filings for well-known tech companies give a window into how much they were paying out as private companies, at least in the year or two, before they went public. Google's stock compensation expense amounted to 16% of the search engine company's revenue in

2003, a year before its initial public offering, Facebook's similar expense was roughly 6% in 2011 a year before its IPO. And Snowflake's stock compensation expense amounted to 30% of revenue for its 2020 fiscal year ending on January 31st ahead of its IPO roughly eight months later. Stock compensation costs at OpenAI are so high that they rival how much OpenAI spends on inference computing or running its chat GPT and its AI models. This year, the company has forecast spending about $6 billion on inference compute or just slightly more than what it expected to record for stock-based compensation. It also expects to spend another $1.5 billion in employee costs such as salaries this year, end quote. And then one more thing. The FT says OpenAI overhauled

its security procedures, adding biometric checks in its offices and isolating sensitive info to protect IPs such as model weights from espionage. Quote, the San Francisco-based startup had been bolstering its security efforts since last year, but the clampdown was accelerated after Chinese AI startup Deepseek released a rival model in January. OpenAI claimed that DeepSeek had improperly copied the California-based company's models using a technique known as distillation to release a rival AI system. It has since added security measures to guard against these tactics. DeepSeek has not commented on the claims. The episode prompted OpenAI to be much more rigorous, said one person close to its security team, who added that the company led by Sam Altman had been aggressively

expanding its security personnel and practices, including cybersecurity teams. A global AI arms race has led to greater concerns about attempts to steal the technology, which could threaten economic and national security. U.S. authorities warned tech startups last year that foreign adversaries, including China, had increased efforts to acquire their sensitive data. OpenAI insiders said the startup had been implementing stricter policies in its San Francisco offices since last summer to restrict staff access to crucial. information about technologies, such as its algorithms and new products. The policies known as

information tenting significantly reduced the number of people who could access the novel algorithms being developed, Insiders said. For example, when Open AI was developing its new 01 model last year, codenamed Strawberry internally, staff working on the project were told to check that other employees were also part of the strawberry tent before discussing it in communal office spaces. The strict approach made work difficult for some staff. It got very tight. You either had everything or nothing, one person said. They added that over time, more people are being read in on the things they need to be without being read in on others. The company now keeps a lot of its proprietary technology in isolated environments, meaning computer systems are kept offline and

separate from other networks, according to people familiar with the practices. It also had biometric checks in its office where individuals could only access certain rooms by scanning their fingerprints, they added, end quote. Blue Sky has rolled out activity notifications letting users get push notifications about new posts and replies from specific accounts as it doubles down on sports activity. Quoting the Verge. If you want to know every time the Verge or ESPN or one of your friends posts, you can just by toggling the bell icon on their profile page, along with an option to see notifications

just for new posts or with replies included too. It's the kind of feature I've gotten used to on other platforms, especially Twitter, where newsbreakers have been able to keep their audiences updated from minute to minute. making it easier to follow interesting topics or developing stories. That goes double for sports, and Blue Sky has called growing its presence in sports discussions a top priority. So far, ESPN reporter and Waj Hare Shams Charania hasn't brought his free agency coverage to the platform yet to go with X threads and the ESPN app.

But now that it has push notifications for specific accounts, maybe he will. Another small change is that it can also notify people if someone likes or repost something they have reposted. Of course, if your problem is that you're getting too many notifications, the updates that rolled out on Monday also have something for you. There's an explosion in the amount of granular controls available to decide if you'll get updates about reports, likes, new followers, or other activity, and if you want to receive them from activity by anyone, no one, or just people you follow. According to a blog post, the simple priority notifications toggle that has been available previously, has been migrated to the new setup so anyone who had it

turned on, we'll still get only notifications from people they follow, end quote. Moon Valley has released Mary. It's fully licensed AI video model available for 1499, 3499, or 1499 per month. 80% of its footage is from independent filmmakers. Quoting time, Moon Valley was founded by DeepMind researchers and has close ties with the film industry. The company owns AI Film Studio Astoria Film, which was co-founded. this year by filmmaker and actress Natasha Leone and filmmaker Bryn Muzer. Astoria has been advising Moon Valley on the development of its AI model, Mary, which is now available to filmmakers for three subscription tiers.

Mary may become AI's main entry point into Hollywood as it's being developed with the approval of filmmakers and trained on licensed data, theoretically allowing studios to avoid the ethical issues and copyright lawsuits that have plagued the AI industry. We have to make sure that we're building these tools the right way, with the filmmaker and the artist at the center of it, rather than trying to automate their job away. Naim Talukdar, Moon Valley's CEO and co-founder, tells time. Moon Valley has raised over $100 million from investors, including Kostla Ventures and Bessemer Venture Partners.

Astoria is using Mary for a new documentary about Carl Sagan to restore and tweak footage. Talukdar also says that Mary is being tested in pilot programs at over a dozen large studios, as well as by major advertising companies. Many other AI video models are black box systems. You type in a prompt and it generates a scene wholesale. If you try to tweak one variable in the scene, another may change, making it hard to maintain control over everything that's been filmed. Moon Valley aims to build tools that integrate into the filmmaking process,

much like CGI and special effects programs did in the past. Mary allows filmmakers to input storyboards or frames and then tweak them as they see fit, hypothetically giving filmmakers far more control over every detail, from objects to characters to motion to scene composition. It's this iterative process where you start with some input guidance and then you build up toward the scene that you want, which really isn't very different from how VFX workflows are today, to Luke Dar says. If you're an independent studio that doesn't necessarily have massive infrastructure, you can now, even in a small space, create and curate these scenes in a very granular way. All of the footage that the model is trained on is licensed from,

IP owners. About 80% of that footage, Tolukdar says, comes from a group of independent filmmakers and agencies that have amassed B-roll over the years. This approach means Mary is trained on roughly one-fifth the data as its competitors, like Google's V-O-3, Tolukdar says. But he claims that Moon Valley is overcoming this deficiency with better technology, created by alumni from DeepMind, Meta, and other top labs. The reality is, if we scraped data, our model would be more powerful, without a doubt, he says. But our inclination is that you don't necessarily have to be the number one model. You just need to be among the best. And I think this is the first generative fully licensed model where you don't have to compromise quality, end quote. Finally today, back to Apple not exactly firing on all

cylinders lately. In iOS 26 Beta 3, Apple dials back the degree of transparency in many UI elements following user complaints about its new liquid glass design language. Quoting TechCrunch. WWDC 2025 in June, the Tech Giant introduced its new design language known as Liquid Glass, which is inspired by the optical qualities of glass in the real world, including how it refracts light and its translucent nature. But the early version of the first developer beta of iOS 26 and the accompanying updates for Apple's other devices still left room for improvement in terms of usability, accessibility, and legibility.

Last month, Apple fixed some of the more prominent issues with liquid glass, like how it made the control center so transparent that the iPhone home screen, icons and widgets shown through, creating visual clutter and confusion. Monday's update sees Apple taking yet another step to dial things back from an overly glassy look in a number of key areas, while Beta 2 address problems with the control center, beta 3 shifts its focus to other areas of the mobile operating system like notifications and navigation within Apple's first-party apps like Apple Music. For instance, the navigation bar and Apple's streaming music app no longer sees the background

shining through a bit, opting for a more solid white. While the changes arguably make features easier to read, some users now complain that Apple has gone too far in the other direction, with a return to more of a frosted glass aesthetic. However, it's worth remembering these are just developer betas, early versions of the mobile operating system that won't be finalized until its public release this fall. The point of beta software is to allow Apple to collect feedback, find bugs, and address issues before the software rolls out more broadly. That means Apple could continue to tweak the liquid glass look and feel over the coming releases to find the sweet spot for the new glassy look within every app and screen, end quote. Nothing more for you today. Talk to you tomorrow.