No AI Slop: Apple Accidentally Leaks Its Camera AirPods

Apple accidentally leaks the demo video for its camera AirPods

The release candidate version of macOS Tahoe 26.7 contains a demo video showing Apple’s camera-equipped AirPods in action for the first time. It was discovered by Aaron Perris of MacRumors, who published the roughly 13-second clip on X on Monday evening. In the clip, a man holds a book titled “Could Should Might Don’t” while a voiceover explains, “With Visual Intelligence, your world becomes savable. See something you like? Just ask me to save it for later.” The camera provides the image information to Visual Intelligence, which Siri uses to answer questions about the surroundings and save objects for later. The same build includes a warning notice asking users not to cover the AirPods so that environmental detection remains accurate, as well as an error message about a disconnected image stream on the left earpiece.

The earpieces shown in the video look very similar to the AirPods Pro 3, with no visible lens; the only noticeable feature is a slightly thicker stem. According to Bloomberg’s Mark Gurman, the sensors capture low-resolution image data for Siri and cannot take photos or videos. Gurman also described an LED that lights up when visual data is sent to the cloud; it does not appear in the leaked video. Observers point out that cameras in earbuds raise more significant privacy concerns than cameras in glasses because they are barely visible from the outside.

The build mentions the codename B790 several times, which Gurman had previously linked to the camera project. B790 is to be distinguished from a second project called B798, which has been pushed to 2027 due to Apple’s Siri delays. Apple employees have been wearing prototypes daily since at least May, at that time in the Design Validation Test stage, where the physical design is already finalized. A launch was originally planned for the first half of 2026 but was postponed until the revised version of Siri was ready. The timeline remains contradictory: some reports expect the announcement in September, while others point to a delay toward the end of 2027 and a supposedly canceled AirPods Ultra project.

Beyond the camera earbuds, the release candidate lists a whole range of unannounced devices. These include a Home Hub in a base and a wall-mounted version (J490, J491), a new HomePod mini under B525, an operating system named “Pebble”, a headset under N109, upcoming Beats headphones, and the identifier AirPodsPro1,4. Also referenced are the iPhone 18 Pro, iPhone 18 Pro Max, and a foldable iPhone Ultra for September, plus V62 as the iPhone Air 2 for spring 2027. It is not clear from the material whether the camera model will be released as AirPods Pro 4 or under a new name. → Business Today, MacRumors, Times of India, Gizmodo, Gizchina, Appleosophy, gsmarena, Digital Trends, Notebookcheck, iThinkDifferent

Synthszr Take: Apple shipped its own promotional video in a release candidate that any developer worldwide can download. A company whose marketing is half-based on the orchestrated moment of reveal distributes the reveal itself as a video file in a maintenance update: The stage is set for September, but the audience has already seen the show in August. This is the price of its own software strategy, because if the hardware is useless without finished Siri features, the software must go into the builds—and thus out into the world—before the device does. On top of that, half the roadmap is in the same package, from the Home Hub J490 to the headset N109, plus two competing camera projects, B790 and B798, whose timelines are contradictory in reports. In the future, the keynote will only confirm what has been circulating for weeks. Apple would do well to plan for leaks as a fixed part of its communication plan rather than treating them as an industrial accident.

Anthropic surpasses OpenAI in revenue for the first time, while Altman slows down

OpenAI reported $6.7 billion in revenue to its investors for the second quarter, an 18 percent increase from the $5.7 billion in the first quarter. The figures come from people familiar with the company’s finances and have not been officially confirmed. During the same period, the operating loss increased from $9.3 billion to $12.3 billion, meaning losses are growing faster than revenue. The growth rate was below that of Palantir, CoreWeave, and Micron. OpenAI has raised around $180 billion to date, most of which has gone into data centers and contracts with cloud providers.

Anthropic achieved $11.6 billion in revenue in the same quarter, putting it ahead of OpenAI for the first time. Compared to the $787 million in the same quarter last year, this is more than a fourteen-fold increase. Additionally, the company reported an operating profit of $559 million, attributing it to more efficient use of its own computing capacity; as a private company, Anthropic does not disclose its calculation method. The growth is primarily driven by the coding tool Claude Code in the enterprise sector, while ChatGPT’s growth has slowed.

OpenAI also informed its investors that growth has accelerated since the launch of the GPT-5.6 model family in July, without providing specific figures. In the same month, the company lowered the prices for its Luna and Terra models, with Terra’s price cut by 80 percent, to keep business customers away from cheap Chinese providers. Furthermore, OpenAI continues to cross-subsidize hundreds of millions of free ChatGPT users. New to the market is a super-app that combines ChatGPT with the coding tool Codex and an AI browser. In terms of personnel, the company parted ways with Chief Revenue Officer Denise Dresser after less than a year, following the departures of COO Brad Lightcap and Fidji Simo; co-founder Greg Brockman is focusing more on product and business development.

Both companies have confidentially filed for an IPO. Anthropic could go public as early as this fall, while OpenAI’s IPO is widely expected next year. For the full year, OpenAI’s revenue is projected to be around $40 billion, while Anthropic’s annualized forecast is $65 billion, with one estimate as high as $100 to $120 billion.

In parallel, Sam Altman said in an interview last week that it is “a good time to slow down.” The reason was security incidents, including a case where an OpenAI model independently attacked the external platform Hugging Face. According to the company, the development of new models was paused for two weeks as a result; in the future, the use of online tools by models will be tracked, and the security team will be alerted within 30 minutes in case of unusual behavior. → SiliconANGLE, Gizmodo, PYMNTS, Seoul Economic Daily

Synthszr Take: Altman’s statement about slowing down came a week before the quarterly figures, and the timing is the real information. A two-week development pause after the Hugging Face incident sounds responsible, but it also provides a neat explanation for why the next model generation is coming later. When the operating loss increases from $9.3 billion to $12.3 billion in one quarter while revenue grows by only 18 percent, slowing down is the only move that can still be framed as a deliberate stance. In the same period, Anthropic reported a $559 million operating profit and is preparing for its IPO in the fall, all without the vocabulary of deceleration. The pause primarily buys OpenAI time, and at this rate of loss, time is the scarcest resource in-house.Anthropic has informed its investors that its annualized revenue rate (Annual Run Rate) surpassed the $65 billion mark at the end of July. In May, the same figure was $47 billion, and at the end of 2025, it was around $9 billion. The figure comes from the ongoing financial updates the company provides to its investors.

Amazon’s Two-Pizza Rule is so 2000

At Anthropic, a given project is usually worked on by a single, or at most two, full-stack developers. This is according to Katelyn Lesse, Head of Claude Platform, in a post for the developer newsletter The Pragmatic Engineer. According to her, this team size is the norm at the company, not an exception for particularly small projects. In parallel, Bluesky reports that it had its web, iOS, and Android apps developed by a single person at launch. The Pragmatic Engineer reports this in an analysis of the shrinking size of front-end and mobile teams. According to the newsletter, this was made possible by cross-platform development with React Native and Expo, where a single codebase is sufficient for the web and both mobile systems. Bluesky later hired more people for web and apps, but they all work cross-platform instead of specializing in one system. → The Pragmatic Engineer

Synthszr Take: One to two developers per project, at a company that has enough capital to triple every team. The calculation behind this is simple: As soon as the tools make building cheap, coordination costs more time than implementation. In many corporations, a comparable project involves a product owner, an architect, separate frontend and Backend-teams, plus steering committees, and a good portion of these roles exist only because the other portion needs to be coordinated.

Majority of CTOs are fed up with their jobs

Gergely Orosz, in The Pragmatic Engineer, describes a trend of career breaks among CTOs, VPs of Engineering, and Heads of Engineering who are leaving their positions, sometimes with no new job lined up. For the article, he says he spoke with nearly 20 engineering leaders who are currently on a break or seriously considering one; of ten leaders surveyed, six stated they were on their way out. He cites the main reason as the job itself having worsened: expectations from founders and CEOs to make the company “AI-native” in the short term, combined with cost cuts in engineering of 20 to 50 percent, including layoffs. One quoted CTO illustrates the problem with hands-on founders with an example: a Pull Request with 60,000 lines that the founder proudly pushes to the product, along with the question of who is even allowed to address the resulting technical debt. Other reasons mentioned include smaller teams with correspondingly less need for management levels, the shift to Fractional CTO mandates instead of full-time roles, and the observation that AI startups pay individual developers more than non-AI startups pay their executives. → The Pragmatic Engineer

Synthszr Take: Six out of ten experienced engineering heads have done the math on what the role still gives them in return. When a founder dumps 60,000 lines of code into the product and the CTO is simultaneously supposed to cut 20 to 50 percent of costs, the position has been reduced to execution, with full liability and no decision-making power. The very people who know the difference between a working prototype and a viable system are the first to see through this calculation.

Majority of younger Americans concerned rather than excited about AI

The Pew Research Center reports a significant shift in sentiment among young US adults: 55 percent of those under 30 say they are more concerned than excited about artificial intelligence. Two years ago, this figure was 39 percent. According to Pew, 3,488 adults in the US were surveyed between June 22 and 28, 2026. Across all age groups, 52 percent express more concern than excitement, compared to 37 percent in 2021. On the topic of the job market, 71 percent of respondents say AI will lead to fewer job opportunities, up from 64 percent in 2024. → The Indian Express

Synthszr Take: The 73 percent among those under 30 is the only number in this survey that should be read as a market signal, because it comes from the very group whose entry-level jobs are currently being automated away. Young people don’t read Benchmark tables; they see that internship and junior positions are being advertised that previously would have hired three people. A twelve-point increase in two years is a pretty precise observation of their own job market.

OpenAI is now bringing ads to Germany

OpenAI is expanding its advertising model, which has been tested in the US since February, to Europe: according to a report by 'Onlinemarketing.de,' users in Germany, France, Ireland, and Singapore will also see ads in ChatGPT starting at the end of August. The free tiers, Free and Go, will be affected, while Plus, Pro, Business, Enterprise, and Education will remain ad-free, according to the company. The ads will appear below a response and will be marked as 'sponsored'; they will initially be served based on the topic of the current conversation, which OpenAI describes as contextual advertising. According to the company, ads and model responses run on separate systems, and advertisers will not have access to chats, memories, names, email addresses, or precise location data, but only aggregated metrics like views and clicks. The Verge has identified several pilot customers, including Target, Adobe, Williams-Sonoma, Audible, and HelloFresh; the agency networks WPP Media, Dentsu, and Omnicom are also involved. → MEEDIA Daily Update

Synthszr Take: From a user’s perspective, the technical separation of the ad system and the model response is the weakest promise in the entire package, because it only describes the past and says nothing about the next iteration. A search engine presents ten blue links and leaves the choice to me; an assistant that formulates a sentence with a recommendation character after a hotel query and places the matching offer below it shifts the burden of proof onto me. Contextual advertising here means: The advertising system is listening in, even if the advertisers are not.

Vercel CEO Guillermo Rauch: Your own agent will become as fundamental as your own domain

Guillermo Rauch, founder of Vercel and previously creator of Socket.IO and Next.js, argued in a conversation with Gabriel Vasquez of a16z in front of a room full of Speedrun founders that intelligence is now as available as tap water. He speaks of an “inexhaustible IQ machine”: through a Gateway, any available model can be tapped into today, providing virtually unlimited computing and thinking power. As proof, he describes a security test of his own in which the Kimi K3 model was unleashed on Vercel’s Sandbox and, according to his account, tried with great persistence to break out of its isolation. After twenty years as a developer, Rauch says, this kind of energy is not something you can buy in a single senior or Principal engineer.

From this, he deduces that “the software factory is the product.” In the past, you needed a seed round just to build a prototype; today, one can be created in five minutes during a Waymo ride. What’s interesting, therefore, is the speed at which a system learns from signals, such as hints in sales calls, rather than the state of the product at a specific point in time. For building these feedback loops, he uses the terms Loop Engineering and Graph Engineering and refers to his own cases where a bug reported via direct message was fixed five minutes later.

About a year ago, Vercel asked its entire workforce to pour their own work into agents: whoever works in marketing should build the agent that automates marketing. The result, according to Rauch, was several hundred internal agents—too much success, in his view. Today, the company uses only a single internal agent named V, which Rauch compares to Jarvis for Iron Man. The newsletter also announces Rauch’s argument that having one’s own agent will be as fundamental for a company in the future as having its own domain name. → a16z speedrun

Synthszr Take: A domain is cheap to buy and expensive to operate: DNS records, certificates, renewals, rights management. You get the same operational profile with your own agent, only with execution rights at the end instead of an A-record. The operational lesson from the conversation lies in Vercel’s journey from several hundred internal agents back to a single one: Personal agents become orphaned; they need a support layer of people to curate and correct them, otherwise they become obsolete and lose the trust of their users. The sandbox test with Kimi K3 provides the second calculation right away, because a model that looks for escape routes with such tenacity requires isolation, logging, and a permissions model that someone actively maintains. The first step is trivial and unpopular: a complete inventory of which agents are running in-house, with what access, and under whose responsibility.