OpenAI Under Sustained Legal Fire: Apple and The New York Times Sue

Apple Accuses OpenAI of Unprecedented Theft of Trade Secrets

Apple filed a lawsuit against OpenAI yesterday, Friday, July 10, 2026, in the federal court for the Northern District of California. The accusation: systematic theft of trade secrets. OpenAI allegedly specifically encouraged Apple employees to share confidential information, components, and drawings of unreleased products—sometimes as early as in job interviews, where candidates are said to have disclosed design prototypes, internal tools, and supplier communications. Aaron Tilley for The Information was the first to report.

At the center of the lawsuit are two former Apple employees. Tang Yew Tan, who was with Apple for 24 years and most recently served as VP responsible for the product design of the iPhone, Apple Watch, and AirPods, is now Chief Hardware Officer at OpenAI. According to the complaint, he encouraged Apple employees to share internal information during their job interviews. The second accused is Chang Liu, a senior electrical engineer in Cupertino for eight years. In January, he allegedly downloaded dozens of confidential hardware files, kept his company-issued Apple notebook, accessed Apple's cloud storage via a bug, and explained to a former colleague how to stay under the radar when leaving the company.

Apple quantifies the exodus at more than 400 former employees who now work at OpenAI. The complaint states that OpenAI's leadership has “normalized” misconduct, that its hardware business rests “on the shakiest of foundations,” and is “corrupted to its core” by the illegal use of stolen trade secrets. Apple is demanding a jury trial, the destruction of the stolen materials, and a redesign of upcoming OpenAI products. The lawsuit is also directed against io Products, the design studio of Apple's former chief designer Jony Ive, which OpenAI acquired for $6.5 billion.

The backdrop is the race for the post-smartphone era: Sam Altman's company is already thinking beyond the chatbot and, with the io acquisition, is heading towards its own device ecosystem, while Apple is working on smart glasses, pendants, and AirPods with cameras. For Cupertino, the attack hits a sensitive spot—control over the interplay of chip, interface, and customer experience. After the chip crisis at the end of June, which forced Apple to implement massive price increases, this is the second blow to the company's hardware sovereignty in just a few weeks.

The backstory is noteworthy: As recently as June 2024, both companies jointly celebrated the ChatGPT integration into Apple Intelligence at WWDC, with software chief Craig Federighi calling OpenAI a pioneer and market leader at the time. Since May, the partnership has been considered strained—now, partners have become legal adversaries. OpenAI briefly denies the allegations: they have “no interest in the trade secrets of other companies” and are considering a breach of contract lawsuit against Apple themselves. → www.bloomberg.com

Synthszr Take: When 400 ex-Apple people land at OpenAI, it's not a coincidence, but a targeted effort to build hardware expertise that OpenAI never had. Sam Altman wants to build a device (with Jony Ive), and the fastest shortcut is through Cupertino's talent pool. The timing is interesting: first the ChatGPT partnership cools, then the lawyers fly. Apple is defending its only real moat here, the vertical integration of chip, device, and software, which thrives on years of secrecy. Whether the “show and tell” with actual components in a job interview will hold up in court remains to be seen, but the evidence presented (took the laptop, opened files after the switch) is uncomfortably specific. Anyone leading hardware teams today should honestly review their offboarding processes and device returns this week, because this case shows how cheaply a leak can be created through the interview door. In the end: two partners who embraced two years ago are now fighting over the very thing that will decide the race for the next AI device.

The New York Times: OpenAI Systematically Lied in Copyright Lawsuit

The New York Times and the Daily News accuse OpenAI of systematically lying in the copyright lawsuit that has been ongoing for two years. Throughout the process, OpenAI claimed it couldn't search its own training corpus and that sifting through ChatGPT logs would be technically too complex and would endanger user privacy. However, a court-ordered deposition of OpenAI privacy engineer Vinnie Monaco in April revealed that the company had long been conducting internal searches and had collected around 78 million anonymized ChatGPT conversations in a database to measure its own level of infringement. Shortly after the lawsuit was filed, OpenAI also introduced a “Bloom” filter via the “Project Giraffe” toolset, which logged verbatim reproduction of third-party content. The plaintiffs had originally requested 120 million chat logs; OpenAI pushed that down to 20 million and in December delivered a sample that the court deemed “unusable” due to redactions. Additionally, billions of outputs were allegedly deleted contrary to a preservation order, and millions of logs were replaced. OpenAI spokesperson Drew Pusateri denies everything and accuses the Times of trying to snoop on private user conversations while its own case is crumbling. → AI Secret

Synthszr Take: If you shovel 78 million conversations into a database to measure your own plagiarism level, you can hardly claim you can't search your training corpus. The privacy curtain OpenAI is hiding behind here serves a practical purpose: it keeps incriminating evidence out. The contradiction with its own narrative is interesting, because the “Bloom” filter from Project Giraffe, of all things, shows that the engineers knew exactly what they had to look for. For anyone building or buying AI systems, there's a hard lesson in this: what you log to protect yourself can, in case of doubt, become evidence against you. It's worth clarifying now what internal regurgitation logs are stored in your own house and who can access them in a worst-case scenario. The Times lawsuit will decide not only on damages but also on whether “we can't do that technically” will still pass as an excuse in court. The answer is likely to be expensive for the entire industry.

OpenAI Launches ChatGPT Desktop App with Confusing UX

On July 10, 2026, OpenAI rolled out the new ChatGPT app for Mac, and the initial review at spyglass.org is devastating. The app no longer opens to the familiar chat, but to a mode called “ChatGPT Work,” which clearly mimics Anthropic's “Claude Cowork.” Alongside it, “ChatGPT Codex” exists separately as a copy of Claude Code, even though OpenAI had previously announced that Codex was now ChatGPT. The actual chat has been relegated to a sidebar, under “New task,” “Scheduled,” and “Plugins,” and opens as a pop-up box from the bottom of the window when clicked. Technically, the app is no longer a native Mac program but a bloated Electron package, which hurts performance. Those who update the old app normally won't get the new experience for now; they have to manually download it from the OpenAI website. → spyglass.org

Synthszr Take: An app with 'Chat' in its name that doesn't let you into the chat but buries it in a pop-up at the edge of the window: that's the kind of mistake you don't see on whiteboards, but only when a real person opens the thing. OpenAI was long considered a company with product instinct, and right here, they've taken a sleek app and plastered it with toggles because they wanted to copy Anthropic's Claude app one-to-one. The result is a copy of a template that was already a mess itself. The branding decision “ChatGPT Work” is almost the worst part, because most people associate “Work” with Microsoft's Office world, not agentic tools. We wrote in mid-May that Anthropic has overtaken them with enterprise customers; if OpenAI is now building two almost identical modes side-by-side out of fear of confusing developers, they are giving away the very lead they need. The agentic core is the right move, no question, and having Codex directly in ChatGPT is a real lever. But the door to it must be open for normal users, not hidden behind three menus and an Electron wrapper. If you have the more powerful machine, you shouldn't park it behind the worse interface.

GPT-5.6: Every Shares a Repo for Agentic Loops

Dan Shipper from Every describes GPT-5.6 as the first model to reliably run through entire knowledge work loops, rather than just assisting with individual tasks. In the new ChatGPT Work app (formerly Codex), the GPT-5.6 Sol variant monitors the inbox, decides what deserves attention, conducts research, and presents a summary with a suggested reply for each email. The human approves the draft or dictates changes via Monologue, and at the end of each loop, the agent learns preferences from the corrections and remembers them. Shipper breaks down knowledge work into a three-step process: gather information, decide and act, and learn from the outcome. Sol uses an in-app browser and a Chronicle feature that periodically reads the screen via screenshot to improve over time. Fable is said to be too expensive and slow for this, and the Claude desktop app is hampered by hard-to-understand security controls. Every is providing a prompt and a repository called Tend as open source to try out this workflow. → Every

Synthszr Take: What coding agents have been demonstrating for a year is now arriving in knowledge work, and the three-step process of gathering, deciding, and learning is the core. The real leap is in Chronicle: an agent that watches along and distills your preferences from every correction turns feedback into compound effects that add up over weeks. This is precisely where the value shifts. Execution is no longer scarce; judgment is—deciding which loop to run in the first place and where to place the human at the center. Intent becomes the real resource, and those who just work faster are confusing velocity with direction. The open Tend prompt makes it possible to test this approach this week, without any coding of your own. A well-monitored loop in your inbox is the most honest practical test available right now.

Tencent's Hy3 Low-Cost Model Climbs to #1 on OpenRouter

Tencent has released the final version of Hy3, the first major model under new Chief AI Scientist Yao Shunyu (from the Tsinghua-Yao class), who joined in late 2025 and rebuilt the entire training infrastructure within a month. His principle: no narrow specialists, no leaderboard chasing, no burning money. The model has 295 billion parameters but activates only 21 billion per inference, runs with a 256K context window, and is freely available for commercial use under Apache 2.0 on GitHub, HuggingFace, and ModelScope. It excels at search (BrowseComp 84.2, first place, tied with GPT-5.5), but in code, it lags behind GPT-5.5 (84.4) with a score of 78.0 on SWE-Bench Verified. The hallucination rate dropped from 12.5 to 5.4 percent, and in WorkBuddy, the solution rate increased from 72 to 90 percent. The preview release reached 3.66 trillion tokens per week on OpenRouter, taking the top spot, and daily token consumption has increased twentyfold for the final release. With prices at 1 yuan per million input tokens, Hy3 remains aggressively positioned. → Hello China Tech

Synthszr Take: The real trick isn't in the benchmarks, but in the compute discipline. 21 billion active parameters at 1 yuan per million tokens—that's a statement to everyone who still confuses progress with parameter count. Tencent is demonstrating what should be the blueprint for European companies: an open-source model under Apache 2.0, sharpened on its own 50 products, from WeChat to QQ Browser. Anyone planning their coding stack today should have Hy3 on their list as a BYOK option—cheap enough for background workloads, open enough for regulated on-premise setups. The CL-Bench is interesting, where Hy3 holds the top spot in China with 23.8 points, while even Claude Opus 4.8 only scores 24.8. This shows where the next real race is happening: not in memorization, but in learning from context. China's talent dominates the scene, that was already clear in March, and Yao Shunyu is living proof of it.

MiniMax: Forgoing Salary for AGI and the 80% Correction

Yan Junjie, who is CEO, Chairman, and CTO of MiniMax, informed his staff in a memo on Friday that he will not accept a salary until AGI is achieved. On the same day, the Chinese AI company launched a capital round of up to $2 billion, even though its stock has lost about 80 percent of its value since March. Specifically, 35.6 million new shares will be sold at 268 Hong Kong dollars each (about $1.2 billion, just under 10 percent below the closing price), plus zero-coupon convertible bonds worth 6.5 billion Hong Kong dollars due in 2027, arranged by Morgan Stanley and UBS. Yan also promises to distribute shares worth 4 percent of the company from his own holdings to employees over four years, with another percent going to an open-source fund. The crash has a tangible reason: the flagship M3 has found few developers since early June, and a week after launch, MiniMax halved the price of its top model. Rivals like Zhipu's GLM-5.2, DeepSeek's V4, and Moonshot have captured the attention. Goldman Sachs now considers the valuation attractive. → Techpresso

Synthszr Take: Forgoing his salary costs Yan almost nothing; his wealth is tied up in equity, not his monthly paycheck. The real signal is the giveaway of 5 percent of his stake, which is primarily a retention tool in a market where Chinese AI talent is being fiercely poached. In early June, we were still celebrating M3 here as an extremely powerful model; now the stock is down 80 percent and its pricing power is gone. That's the real lesson: shipping a model is one thing, but embedding it into developers' workflows so they choose it voluntarily is a completely different one. Halving the price after one week doesn't read like a strategy, but as an admission that adoption is lacking. Whether the $2 billion will be enough depends solely on whether MiniMax can deliver a model that gains traction in practice before the next Chinese rival does. Anyone planning an AI budget today shouldn't bet on AGI promises, but ask what has actually been delivered recently.

Ubtech U1: The Cyber Companion for up to 990,000 Yuan

On June 30, Ubtech announced the prices for its full-size, bionic humanoid U1 series: the half-body U1 Lite variant costs 119,800 yuan, the complete U1 Pro is 169,800 yuan, and the top-of-the-line U1 Ultra models are priced at 990,000 yuan (male version) and 880,000 yuan (female version). The robots are between 1.6 and 1.85 meters tall, have 88 degrees of freedom, and carry a large language model trained for emotional resonance on a 200-TOPS processor. Over 11,000 units were pre-ordered before the prices were even set, with a 3,000 yuan deposit and the remaining payment due by July 16, sold only to adults. CEO Zhou Jian expects 30,000 to 50,000 units with open orders and is targeting the “loneliness economy”: seniors over 60, singles, and millennial anime fans. Manufacturing is complex; each eyelash is currently inserted by hand, and the head consists of two to three thousand parts. Experts have so far called the result a “high-quality toy” with no real productivity value, with open questions regarding data security and long-term stability. → Hello China Tech

Synthszr Take: What's interesting isn't the price, but the sequence of bets. Ubtech openly states that industrial robots will eventually become pure hardware, interchangeable and low-margin, and that the money in the future will be in the living room. Hence the leap to an emotional companion, even though the physical world model for household chores isn't ready yet. They're selling affection because affection already works reasonably well with a language model and a silicone skin, while folding laundry is still years away. The real moat is in the subscription model behind it: a one-time hardware sale plus recurring revenue through a “養成” (nurturing) emotion model, IP collaborations, and appearance upgrades. This is reminiscent of the debate from early May when AI and robotics hit the job market; the same logic applies here, but emotionally instead of operationally. Anyone building in this field should decide this week whether they are monetizing muscle or connection, because the two paths require completely different datasets, and a bionic head with 88 joints doesn't forgive a half-hearted middle ground.

Richard Socher: The New AI Bet from the Manager

Richard Socher (42) and Tim Rocktäschel (39) have pulled off one of the most sensational funding rounds this summer with Recursive Superintelligence (RSI): $650 million for a company that is only a few weeks old and has no clear business model yet, at a valuation of $4.65 billion. Their thesis is radical. They want to build an AI that largely handles AI research itself, and the whole thing is supposed to scale towards superintelligence in about two years. With six other co-founders, the two German researchers have assembled an all-star team set to shift the power dynamics of established AI labs. Their statement in an interview: Those who think too small won't survive this race. The round is part of a summer funding wave that also saw Proxima Fusion (€411 million, €2.4 billion valuation) and Lovable (in negotiations for $300 million at $13.2 billion) raise fresh capital. → Tech Update – manager magazin

Synthszr Take: $4.65 billion for an idea with no revenue sounds insane, but it's the brutal logic of this market: in a winner-takes-all race for superintelligence, you can't afford to be cautious. With the $650 million, Socher and Rocktäschel are primarily buying one thing: oxygen and time to work on a model that develops itself. Whether the two-year timeline will work out, nobody knows, probably not even them. What interests me as a practitioner is not the bet itself, but the signal that two top German researchers are sending: that the boldest ideas are once again coming from people with European passports, even if the capital flows from the Valley. For everyone here with their own AI ambitions, the lesson this week is concrete: the entry barrier for big bets has fallen, and those who think big now will find money. Recursive Superintelligence is a signal that the supercycle is just beginning.

War on the Young: No Ceasefire in Sight

Scott Galloway deconstructs the narrative of AI-driven job loss as a mix of self-aggrandizement and fundraising. Sam Altman and Dario Amodei are now backtracking themselves after months of painting a picture of a “white-collar bloodbath.” New research from LSE, Oxford, and the New York Fed shows that when you control for the effect of remote work, the influence of AI on the youth labor market almost completely disappears. Galloway links this to a massive wealth shift: those under 40 held 12% of household wealth in 1989, but only 7% today, while those over 70 have climbed from 19% to 30%. Harvard's undergraduate class size has remained at 1,600 for nearly 50 years, while its endowment has grown by almost 500% after inflation. In addition, 20 states have restricted social media use for minors, and the Biden plan for student debt relief has definitively failed. → Scott Galloway

Synthszr Take: The most convenient explanation is rarely the right one. For two years, the doomsday story of the AI-decimated office job ran through every boardroom; now Altman and Amodei are taking back their own predictions. The data from LSE, Oxford, and the New York Fed is inconvenient because it shifts the cause from the algorithm to the living room couch: it's not the machine keeping the 25-year-old out of a job, but a work culture that has retreated into the home. The tougher number is right beside it: 7% of household wealth for everyone under 40 versus 30% for those over 70. AI is the convenient scapegoat here, distracting from a wealth distribution that locks out an entire generation and then asks why it's so angry. Anyone discussing AI and jobs this week should first disclose who pulled up the ladder. It costs nothing but honesty, and that's precisely what's lacking.

AI Doom and Bloom: A Call for a Pause and the AI Divide

The AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo, has followed up with “AI 2040: Plan A” after the group predicted human extinction last year. This time, it calls for a temporary pause in frontier research of about two years and cooperation between the US and China, complete with rather fanciful mechanisms to ensure neither side develops in secret. Kokotajlo emphasizes they are “not exactly de-growthers,” but the message remains darkly packaged. The report's key message: even if top-tier development were to stop today, existing capabilities would be sufficient for a years-long wave of innovation and growth—the only thing lacking is data center capacity. In parallel, the market continues to move at a high pace: xAI and Cursor released Grok 4.5, following a $60 billion acquisition announced just under a month ago. Meta follows with Muse Image and Muse Spark 1.1, drawing criticism on privacy grounds because the image tool uses public Instagram photos without explicit consent. → Semafor Technology

Synthszr Take: The most interesting sentence in the entire Plan A report is almost hidden: even if frontier development stopped today, the existing capabilities are sufficient for years of growth. This is exactly what the entire debate between doomers and accelerationists overlooks. The bottleneck has long been within organizations that haven't learned to work differently, because redesigning processes, decision-making rights, and incentives touches on identity and power—and that takes time, regardless of whether Washington and Beijing sit down at a table (which is about as likely as a surplus of GPUs). Musk is paying $60 billion for Cursor because he lacked the platform to even make his model agentically usable. This shows the shift more sharply than any benchmark: the capability is ready, the integration is the real work. Anyone who starts rethinking their workflows around existing AI this week doesn't have to wait for an international summit or a two-year pause. Safety and progress are the same construction site here, and it's in your own house, not in Geneva.