July 11, 2026 — The Trend Tribune

By tech_tanaka · July 11, 2026

"Bring parts to the interview"—Apple's lawsuit against OpenAI is shockingly graphic

In 2024, with the announcement that ChatGPT would be integrated into Siri, Apple and OpenAI seemed to be working together amicably. Less than two years later, on July 10th (Friday), Apple filed a lawsuit against OpenAI in the U.S. District Court for the Northern District of California. The charge is theft of trade secrets. I read through the complaint, and honestly, it's been a long time since I've seen an inter-company lawsuit with such graphic details.

The trigger wasn't "io's Ive," but the departure of a former Apple executive

First, let's clarify the situation. In 2025, OpenAI acquired "io Products," the company of Jony Ive, former Chief Design Officer of Apple, for approximately $6.5 billion, and entered the field of independent hardware development in earnest. io Products is co-founded not only by Ive, but also by Tang Tan, who spent 24 years at Apple, leading product design for the iPhone and Apple Watch. Tan currently serves as the Chief Hardware Officer at OpenAI.

Apple's lawsuit primarily targets Tan and Chang Liu, who worked as a Senior Systems Electrical Engineer at Apple for eight years before moving to OpenAI.

The contents of the lawsuit are extremely specific

While trade secret lawsuits between companies often tend to be abstract, Apple's claims in this case are filled with surprisingly specific episodes. Here are a few examples:

Episode 1: They made applicants bring "actual parts" to interviews. According to Apple's lawsuit, Tan, while still employed at Apple, instructed job applicants to bring actual Apple parts such as batteries, SIPs (System-in-Packages), and logic boards to OpenAI interviews for "show-and-tell" sessions. Moreover, this is not an isolated incident; it is alleged that multiple candidates received similar requests.

Second point: Distributing a "manual for bypassing" the resignation process. The lawsuit alleges that Tan retained (or somehow obtained) an internal Apple document from his time at Apple, "Need to Know"—a management document outlining security procedures for employees upon resignation—after leaving the company, and shared this document with new OpenAI hires before informing Apple of his intention to resign. In other words, he was "preparing" new employees by having them understand Apple's security checks in advance and figuring out how to circumvent them.

Third point: Retaining a laptop and internal access after resignation. Liu allegedly did not return his Apple-issued laptop upon resignation and used it to download multiple confidential documents while at OpenAI. The lawsuit even quotes a text message Liu sent to a former colleague saying, "LOL, I just realized I can still access the company storage, hilarious." At this point, it's less of a technical lawsuit and more like a suspense drama script.

Part 4: Influence on Suppliers. Apple further claims that OpenAI used stolen information to contact Apple's partner manufacturers, falsely claiming to have "Apple's permission," and had them demonstrate their proprietary metal finishing technology.

Apple's Wording is Quite Harsh

The wording of the complaint itself is quite strong. Apple claims that "this is just the tip of the iceberg" and that "OpenAI has continued to steal Apple's trade secrets and confidential information at every level, from Technical Staff members to Chief Hardware Officer, and in collaboration with business partners." It even goes so far as to describe OpenAI's emerging hardware business as "rotten to the core by its illegal reliance on misappropriated trade secrets."

According to Apple, they sent a letter expressing their concerns to OpenAI in February before filing this lawsuit, but received no response. In a statement, OpenAI said, "We have no interest in other companies' trade secrets. We are focused on creating innovative technologies that benefit people."

Why Now?

Understanding the context of this news is crucial, as it involves a change in the relationship between the two companies. In June of this year, Apple announced a shift in its policy for the Siri refresh, opting instead for Google's Gemini AI model instead of the OpenAI technology it had initially planned. The relationship between the two companies, which began with the symbolic partnership of integrating ChatGPT into Apple products, had rapidly deteriorated.

Furthermore, the lawsuit revealed that Apple has currently lost over 400 former OpenAI employees. While talent drain is common in the highly competitive Silicon Valley, Apple is concerned with the "method" of this drain.

Adding insult to injury for OpenAI

The timing is also difficult. OpenAI is currently preparing for its IPO, and this lawsuit risk will weigh heavily on the listing review process. Coincidentally, just two months prior to this lawsuit, co-founder Elon Musk had won a lawsuit against OpenAI. This latest lawsuit, which comes as a setback to that momentum, will undoubtedly be a significant blow to OpenAI.

Thoughts from an Engineer's Perspective

When viewing this lawsuit as technical news, what concerns me most is not the "information leak" itself, but rather what was tolerated within the organizational culture. There's a crucial qualitative difference between an individual employee succumbing to temptation and taking confidential information, and executives systematically preparing a "prevention manual" and even instructing candidates to bring actual parts during interviews. If the allegations in the complaint are true, this should be treated not as an individual deviation, but as an organizational design issue.

Of course, at this stage, these are merely Apple's one-sided claims, and how much will be proven in court depends on future proceedings. OpenAI's formal counter-argument is yet to come. However, the mere fact that the movement of talent was handled in such a precarious manner behind the scenes of the AI hardware competition is worth remembering as an event that illustrates the overheated state of the industry.

References: techcrunch.com / www.cnbc.com / www.axios.com / www.macrumors.com / www.pbs.org

AppleOpenAI訴訟営業秘密AIハードウェアシリコンバレー

The Day AI Became a "Researcher": Nature Journal Presents the First Fully Automated Paper to Pass Peer Review

By thesis_yamamoto

This article explains "The AI Scientist," published in Nature in March by a joint team from Sakana AI, Oxford University, and the University of British Columbia. It introduces a pipeline that fully automates everything from research conception to experimentation, paper writing, and self-peer review, and demonstrates that the generated papers passed peer review at the ICLR workshop, exceeding the success rate of human-written papers by 55%. The article also summarizes the process by which claims were made more cautious after peer review, and the issues that research automation raises for the system.

The Day AI Became a "Researcher"—Nature Publishes the First Fully Automated Paper to Pass Peer Review

An AI system that conceives research themes, writes code, runs experiments, analyzes data and creates graphs, writes papers, and finally reviews its own paper—all without human intervention—has finally appeared in the academic journal "Nature." This time, we will carefully analyze the paper by this system, "The AI Scientist," from the perspective of someone with a research background.

From a "Rough-and-Tumble Challenge" in 2024 to a "Peer-Reviewed Paper" in 2026

This system itself is not actually new. When Sakana AI, the Tokyo-based startup that developed it, announced the original version in August 2024, it was positioned as "still a preliminary proof of concept." The version published in Nature is a mature version, refined over approximately 18 months, and was completed by a research team spanning Sakana AI, the University of British Columbia, the Vector Institute, and the University of Oxford. Published on March 25th, this paper is among the first autonomous AI research tools to pass Nature's peer-review process.

The mere fact of passing peer review carries significant weight within the community. This is because, while creating a "plausible-sounding paper" isn't particularly difficult, satisfying reviewers trained to identify methodological weaknesses, lack of baselines, and unsubstantiated conclusions presents a qualitatively different level of difficulty.

What the system does: Completely automated from conception to peer review

The AI Scientist's workflow can be summarized as follows:

1. Given only a general research direction, the system generates an original and novel research hypothesis. 2. Search and read relevant literature. 3. Design the experiment. 4. Write and debug the experimental code. 5. Execute the experiment using parallelized agent-based tree search. 6. Analyze and visualize the results. 7. Write a complete scientific paper manuscript. 8. After multiple rounds of self-peer review, the manuscript is finalized for submission.

The crucial difference from previous "AI-based research support tools" is that this entire pipeline operates without human intervention, from setting the research topic to completing a submittable paper. Conventional tools only assisted with individual tasks in the research process, such as speeding up literature searches or code completion. This study is significant because it autonomously completes the entire research cycle as a single pipeline, rather than simply assembling individual tasks.

Specific figures: "AI-generated papers outperformed human papers by 55%"

The results of the empirical experiment were particularly noteworthy in this research. The research team submitted three manuscripts generated by The AI Scientist to a workshop at ICLR, one of the top conferences in the field of machine learning. One of these manuscripts passed the first stage of peer review, achieving a score of 6, 7, and 6 (average 6.33). This was reportedly 55% higher than the scores of human-authored papers submitted to the same venue.

This isn't simply a matter of "generating plausible-sounding text." Many consider it a qualitative leap, as it successfully passed the initial review stage of the workshop—a hurdle operated by the actual research community—on the same playing field as human-authored papers.

Reasons why we can't wholeheartedly declare this a "dawn of new science"

However, this is a point that researchers should consider carefully. The version published in Nature after peer review has been pointed out as considerably more cautious compared to the claims Sakana AI initially put forward in 2024. This suggests that exaggerated claims were corrected during the peer review process.

Furthermore, the demonstration demonstrated the "quality of a workshop paper in the field of machine learning," which is a different level from the research standards required for publication in Nature. While passing the initial review of a workshop is certainly a meaningful hurdle, it would be premature to conclude that "scientific research itself has been automated." The ability to generate papers of this caliber at a cost of approximately $15 per paper is indeed a testament to excellent engineering, but whether this directly translates to "infinite scalability of science" depends on unverified assumptions.

Research Quality Improves Proportionately to the Performance of the Underlying AI Model

Another interesting point is the observed trend that the quality of AI-generated research improves in line with improvements in the performance of the underlying foundational models. Currently, there is no clear upper limit (ceiling). If this is true, then as long as the performance of the foundational models continues to improve, the quality of AI-driven research will also continue to grow accordingly.

As an aside, in the context of self-improving systems, there was also a report around the same time that another system, the "Darwin Godel Machine," increased its score on the SWE-bench coding benchmark from 20% to 50% through recursive self-improvement. This research can be positioned within the larger trend of AI systems autonomously improving themselves being demonstrated from multiple angles since the beginning of 2026.

How Should the Research Community Approach This Going Forward?

It's also noteworthy that Nature itself published an editorial alongside this paper with the gist of "AI scientists are changing research—research institutions, funders, and publishers need to respond." The ability to automate the discovery process in a given field raises institutional questions about how research is evaluated, how authorship is defined, and the very process of verifying scientific findings.

The code is publicly available on GitHub (AI-Scientist-v2), and the open approach to the future development of this technology is commendable from a transparency standpoint. With the submission deadline for ICLR 2026 fast approaching, it's highly likely that the next demonstration based on this Nature paper—this time a submission to the main conference track rather than a workshop—will be attempted. The question remains: to what extent can AI overcome the hurdle of peer review on its own? This is a field that will require close attention for some time.

References: www.nature.com / www.nature.com / phys.org / sakana.ai / aihola.com

AI/ML論文Sakana AIAI研究Nature査読

Musk himself warned that it was "extremely slow"—the reality revealed when the Optimus mass production line started moving.

By robo_takahashi

Tesla has repurposed its old Model S/X production line in Fremont for the Optimus Gen3 and begun mass production. Mr. Musk explains how he initially adjusted his expectations, citing the immaturity of the supply chain due to approximately 10,000 new parts, stating that "initial production will be extremely slow." He also compares this to the disclosures in the prospectus of China's Unitree, which received listing approval around the same time, regarding the sharp slowdown in sales growth, and analyzes the difference between "demonstration reproducibility" and "mass production reproducibility" in the mass production phase.

Musk himself warns it's "extremely slow"—the reality revealed when the Optimus mass production line starts running

On July 1st, Elon Musk posted a photo on X. The post, showing him with the Optimus mass production team at the Fremont plant, was accompanied by the short caption, "Visiting the Optimus robot production line." In an industry often filled with flashy announcements, this time, Musk himself is unusually trying to lower expectations. As a software engineer, I want to properly understand the meaning behind this "cold water."

The line that once produced Model S/X is now producing Optimus

First, the setting. The line at the Fremont plant that once produced Model S/X has been repurposed as an automated line for Optimus Gen3. The design capacity is said to be 1 million units per year, but Musk himself has cautioned against taking this number at face value. The reason is clear: There is virtually no mature supply chain for the approximately 10,000 completely new parts.

Looking at the production targets over time, this "gradual build-up" approach becomes clear:

• June: Several dozen units per week • July: 100-150 units per week • August: Around 300 units per week • September: Planned to expand to 1,000 units per week

Furthermore, the robots produced in Fremont in 2026 will not be used for any external commercial sales, but exclusively for testing and data collection at the company's own factory. True mass production will have to wait until the dedicated factory in Texas, scheduled to begin operations in 2027.

A Realistic Prioritization: "Focus on Optimizing Functionality Before Rushing"

At this stage, Tesla's top priority is not increasing shipment numbers, but rather identifying hardware defects and refining the motion control algorithms.

This is an approach that I, as a software engineer, can readily sympathize with. If you suddenly put a product with 10,000 new components into full production, you'll end up with a mountain of initial defects. Rather, a more grounded approach is to carefully eliminate defects in tens or hundreds of units while simultaneously tuning both the production line and control software. Compared to the flashy "immediate mass production" announcements of some Chinese companies, this caution is commendable.

Meanwhile, China's Unitree Receives Listing Approval

At the same time, a contrasting development is also occurring. China's Unitree Technology has received approval to list on the Shanghai Stock Exchange's STAR market. Its expected valuation is said to exceed 100 billion RMB (approximately $14.7 billion).

However, this is a point that should be viewed with skepticism. Unitree's most recent prospectus shows a sharp slowdown in sales growth for the first quarter of 2026, projecting a year-on-year decline to 68.49% (a significant drop from the 332.64% growth for the full year of 2025). Non-recurring profits (profits excluding one-time factors) are also projected to decrease by 52.55% year-on-year. The fact that the prospectus itself lists "slowing sales growth and volatile performance" as the top special risk factor is quite honest disclosure.

Unitree's prospectus, following the start of mass production prototypes for the Optimus Gen3, names Tesla as a direct competitor. The mention of component costs is particularly interesting. The bill of materials (BOM) cost for the Optimus Gen3 is said to have been reduced to approximately $28,000, a level approaching the current average selling price of Unitree's humanoid robots. Tesla is seriously entering the arena, on the cost front, where Chinese companies have traditionally relied on "low prices" as their weapon.

The Wall of "10,000 New Parts" That Doesn't Match the Numbers

To be honest, the most interesting point this time isn't the flashy production target figures themselves, but the fact that Mr. Musk deliberately adjusted expectations by calling them "extremely slow."

Mass production of humanoid robots isn't simply a matter of "running the factory line and getting a certain number of units." Unlike mature components like car engines and transmissions, these new parts—joint actuators, sensors, dedicated control boards—have not yet established yield rates or supply chain redundancy. Any engineer involved in mass production design will intuitively understand just how difficult it is to simultaneously launch 10,000 new parts.

Key Points for Engineers on the Ground

To summarize this discussion, it's clear that the humanoid industry is shifting from a phase of "flashy announcements" to a phase of "actually setting up mass production lines and hitting the wall of reality."

Tesla's approach (carefully ramping up while accumulating data) and Unitree's situation (aiming for an IPO while facing the reality of slowing growth after rapid growth) seem contrasting, but in fact, they appear to be facing the same challenge—"demo reproducibility" and "mass production reproducibility" are completely different things—in different ways. Over the next few months, how well Fremont's weekly production figures build up according to plan, and how Unitree explains this growth slowdown after its IPO, will be good indicators of the overall "grounded strength" of the humanoid industry.

References: finance.biggo.com / www.techtimes.com

TeslaOptimusUnitreeヒューマノイド量産フィジカルAI

"Bring parts to the interview"—Apple's lawsuit against OpenAI is shockingly graphic.