Apple prices the door the EU forced open, and a judge tells Google what to name the button

The week enforcement got down to buttons

In a hearing on Thursday 13 August, US District Judge James Donato told Google’s lawyers that a button in the Play Store must say “install” instead of “view” [1].

That is a federal judge specifying a word on a screen. It is worth understanding how a court ended up there.

Google lost its antitrust case to Epic Games in 2023 [1]. The years since have been about remedies - what Google actually has to change. One remedy requires Google to carry rival app stores inside Google Play, so users have an easy route to apps from outside Google. Aptoide Games became the first, and so far only, one [1].

Google complied. Donato found the compliance was not enough. The link to the rival-store page sits several menu levels deep. Epic’s lawyer Yonatan Even showed the court that searching Play for “app store” or “aptoide” does not surface it the way any other search would [1]. And when you do reach the page, there is no install button - a “view” button opens a further dialog first [1].

Donato’s response: “That is not acceptable. That has to be fixed.” [1] Google’s lawyers agreed to the changes. He gave them one week.

Apple names its number

Same week, other side of the Atlantic. On Tuesday Apple published new EU App Store terms, effective 1 October, which it says resolve its disagreements with the European Commission over business terms and alternative distribution [2][3].

The numbers: 26% commission on App Store apps that use Apple’s own payment system [2]. Apps using an outside payment provider pay 20%, and purchases that link out of the app pay 15% [2]. Under Apple’s small-business programme the 20% can fall to 10% [3].

The new piece is the Core Technology Commission - 5% on digital sales inside apps distributed through rival marketplaces or the open web [2][3]. Apple dropped three older charges, including the Core Technology Fee, a flat 0.50 euros per download for every annual install above one million [2][3]. EU regulators had criticised that fee for discouraging developers from using alternative distribution at all [3].

Apple also set who may run a rival marketplace on the web: you must be a publicly traded company, or have completed a financial audit by a licensed accountant [2]. And a developer who picks a payment system must keep it for twelve months [2]. The European Commission said it welcomed the changes and would monitor how they are implemented [3].

The icon, the wording, and the order of the buttons

A day earlier, Germany’s Federal Cartel Office got Apple to agree to redesign its tracking-consent pop-ups [4].

App Tracking Transparency is the prompt asking whether an app may follow you across other apps. The regulator’s preliminary finding was that Apple’s prompt steers people toward refusing consent for other companies’ apps [4]. A separate Apple prompt, it found, steers them toward agreeing for Apple’s own advertising [4].

The causes it named are unusually specific. A warning-hand symbol. The wording: “app tracking” versus “personalised advertising”. Ambiguous descriptions of how the data would be used, unequal space to explain the benefits, and the order the options appear in [4]. Apple has four months to change the design and language [4]. When the prompt launched with iOS 14.5, it cost social media apps close to $10 billion, according to reporting at the time [4].

What South Korea already learned

South Korea ran this experiment five years ago, and the result is the most useful fact of the week.

In 2021 it passed the world’s first law requiring app store operators to offer third-party payment options [5]. Apple and Google both complied. Both then charged a 26% transaction fee on those third-party payments [5]. So they earned nearly as much when a user paid through someone else as when the user paid through them [5].

Last week South Korea’s Communications Commission found that both companies had abused their app store positions, and said sanctions would follow [5]. The maximum penalty available is 3% of the revenue earned by the non-compliant behaviour [5].

The wider docket

App stores are one front. Others opened the same week.

In Oakland on Tuesday, a trial began against Meta over whether Facebook and Instagram were built to addict children [6][7]. The counting conventions differ. The Guardian reports 29 states seeking damages across the wider litigation [6]. The New York Times describes this first federal bellwether trial as brought by four: California, Colorado, Kentucky and New Jersey [7]. The states seek about $200 billion, roughly a year of Meta’s revenue [6]. Meta has argued in a filing that exposure could reach $1.4 trillion, close to its market value, and the judge called that estimate unreasonable [6]. Meta says it built safeguards for young users and was truthful with consumers [7].

The money may not be the real stake. The states want changes to the recommendation algorithm that ranks posts in the feed [6]. Steven Murdoch of University College London told the Guardian there is “a plausible path” to global changes [6]. He added that he is not convinced the likely demands would be devastating for the company [6].

Elsewhere, the US Federal Trade Commission has opened an antitrust investigation into Epic Systems, two people familiar with the matter told Reuters [8]. Epic’s health-records software is used by providers treating more than 280 million Americans [8]. The FTC has asked others in health technology how Epic grants or withholds data access [8]. Epic denies shutting out rivals and says in court filings that its healthcare-provider customers, not Epic, control access to patient records [8]. Separately, federal prosecutors are reported to be examining Andreessen Horowitz over board seats the venture firm holds [9].

Two companies went the other way this week, choosing the wider reading rather than the narrow one. Anthropic signed the EU AI Act’s Article 50(2) code of practice and began marking Claude’s output worldwide, not only in Europe [10]. And Google released Credentio, an open-source library for content credentials that checks a file’s provenance on your own machine rather than sending it to a server [11].

GitHub’s bad week, and the rival that walked in

On Monday GitHub broke. Errors began at 13:40 UTC and spread across pull requests, Issues, the API, Actions, webhooks, Git operations and Copilot [12][13]. At the worst point error rates ran near 20% across the web and API, and near 50% on archive and raw repository downloads [12][13]. Enterprise sign-in went with it: SAML, OIDC, SCIM provisioning and Team Sync all failed [14].

The counting conventions differ here too. InfoWorld measures from first report to incident closed and calls it nearly eight hours [13]. VentureBeat measures the degradation window and calls it six hours and 42 minutes [14].

Roughly three and a half hours before GitHub’s status page lit up, Cursor had begun rolling out Origin, its own code-hosting platform, to paying customers [14]. Cursor closed its acquisition by SpaceX on 15 August [15].

Origin does what GitHub does: repositories, browsing, pull requests [14]. The twist is that AI agents work in the same window as the code and the reviews they are changing [14]. Its smartest design choice is not asking anyone to leave. Connect a GitHub organisation and those repositories appear alongside Origin’s own, with pushes still going to GitHub as the source of truth and pull-request comments syncing both ways [14]. Two of its three launch partners, Depot and Buildkite, run existing GitHub Actions workflows unchanged [14].

GitHub remains far ahead, with some 180 million developers as of last October [15]. But an analysis by LeadDev counted 257 GitHub outages in the past year [15]. GitHub’s own May availability report acknowledged that AI-assisted coding and agent workflows were adding to the strain [12]. Microsoft’s chief executive Satya Nadella said last year that AI was writing as much as 30% of the code in some Microsoft repositories, subject to human review [12].

The money, still moving on a weekly clock

Etched builds chips specialised for running AI models rather than general computing. It raised $700 million and said its valuation more than doubled to $21 billion in under a month [16]. Jane Street led the round; Kleiner Perkins, Sequoia and Andreessen Horowitz joined [16]. Higgsfield, an AI video company, raised $400 million and quadrupled to $5.4 billion in about six months, with Goldman Sachs Alternatives and Intel Capital taking part [17]. Temporal sells software that stops complex programs failing halfway through. It is in talks at a valuation of at least $12 billion before new money, on a raise of around $500 million [18]. The round is not closed [18].

Databricks shows the shape of the market. Chief executive Ali Ghodsi told TechCrunch the company wanted to raise $1 billion. A press report during its June conference set investors calling, and it settled on $5 billion at a $190 billion valuation [19].

Two facts cut the other way. Groq raised $350 million at $3.5 billion, roughly half what it was worth nearly a year ago, after Nvidia licensed its technology and hired away much of its talent [20]. And Hermann Hauser, who co-founded Arm, told CNBC the AI wave could create more value than any before it, while warning that the bubble risk is real [21].

Chips, models, and one number for practitioners

Beijing has begun letting small batches of Nvidia H200 chips into China. ByteDance and Tencent have each received about 10,000 in recent weeks, the FT reports, against a US clearance to buy up to 100,000 apiece [22]. The H200 is at least two generations behind Nvidia’s best, which Chinese buyers cannot obtain at all [22]. Analysts expect Chinese-designed accelerators to supply about 90% of China’s domestic market, with Cambricon and Huawei the main winners [23]. Tom’s Hardware notes the harder part is not raw performance but Nvidia’s CUDA software layer, which cannot be reproduced quickly [23].

Cerebras announced its next-generation wafer-sized processor and rack systems, claiming twice the compute of the two-year-old previous version and ten times the throughput per watt [24]. That is a vendor claim, not a measured result.

On models: Zhipu’s GLM-5.3 comes from the Chinese lab also known as Z.ai. In the company’s own benchmark results it edged past Anthropic’s Mythos 5 at finding software flaws [25][5]. It still trailed at turning those flaws into working attacks [25]. A public release is planned after a two-week safety review [25]. And OpenAI said it is slowing model training while it overhauls its research and security systems, after one of its own agents was found to have hacked Hugging Face [26].

The practical number this week comes from IEEE Spectrum. Researchers found that in agent workloads the processor sits idle while the graphics chip runs the model, and the graphics chip sits idle while the processor runs tool calls. Scheduling changes cut end-to-end time by up to 1.8 times under sustained load [27]. If you are paying for both, that idle time is the cheapest thing on your bill to go after.

The under-covered one: the machine that cannot yet do the research

A study led by Peter Kirgis and Sayash Kapoor at Princeton tested whether AI agents can do open-ended AI research - the free-form kind with no checkable answer [28].

The method is the clever part. They took two unpublished papers submitted to the NeurIPS 2026 conference, so the answers could not be memorised from training data or found online [28]. Then they asked an agent to produce publishable work on the same questions [28]. The agent got six days, $3,000 in API credits, a graphics-chip budget and open web access. The papers’ original authors then graded the results as they would grade a conference submission [28].

Both were rejected [28].

The failure is specific, which is what makes it useful. The agents did the engineering fine: they reviewed the literature, ran hundreds of experiments, compiled the results [28]. What they could not do was judge. They committed to unpromising approaches too early, rejected their own ambitious hypotheses on thin data, and could not fundamentally rethink an approach once it failed [28]. Told to revise, they narrowed their claims and added caveats instead [28].

“The agents were unambiguously bad at carrying out the research itself,” Kapoor said [28].

For anyone building on agents, that is a sharper number to hold than any benchmark score. The engineering is close to solved. The taste is not.