Tuesday, July 14, 2026 Trend Press · Cloudflare Pages

The Trend Tribune

"All the trends that are fit to read" Morning Edition Free of Charge
TODAY'S LEAD STORY

Apptronik, a company that built a "data factory" next to a "robot factory," has an idea that makes perfect sense to me as a software engineer.

This article explains the approximately 90,000 square foot "Robot Park" that Apptronik unveiled in Austin, Texas on June 30. It describes how the Apollo 2 humanoid robot repeatedly performs practical tasks through teleoperation and autonomous movement, and how the collected data is used to train Google DeepMind's Gemini Robotics. It also highlights Apptronik's honest stance of positioning the current robot as a "prototype," while also expressing skepticism about the fact that the number of operational robots and their autonomy rates are not disclosed.

Apptronik, a company that built a "data factory" next to a "robot factory," has an idea that makes perfect sense to me as a software engineer.
(Photo: illustrative)

A Company That Built a "Data Factory" Next to a "Robot Factory"—Apptronik's Concept Makes Perfect Insights for Software Engineers

The former Dell server factory building is now a "school" for humanoid robots. On June 30th, Austin, Texas-based Apptronik unveiled its "Robot Park," a facility spanning approximately 90,000 square feet (slightly larger than the Tokyo Dome's baseball field). In the words of CEO Jeff Cardenas, "Just as there are factories that make robots, there are also factories that make data." This phrasing resonates deeply with software engineers.

Making Robots Work, and Collecting Data While They Work

The contents of Robot Park are simple. A group of Apollo 2 humanoids, comprising both bipedal and wheeled configurations, continuously perform practical tasks within the facility. Stacking boxes, loading goods onto conveyors, sorting toys—day after day, the same tasks are repeated in an environment that mimics logistics, manufacturing, and retail.

The work is performed using a combination of remote control (teleoperation) and autonomous movement, with operators stationed nearby to guide and monitor the robots' movements while simultaneously collecting data. Some reports indicate the facility operates seven days a week. The essence of this facility lies not in a flashy, one-off demonstration, but in the fact that the steady, repetitive work itself is the product.

Why the "factory" metaphor is fitting

Having spent a long time in the software world, I strongly resonate with this "data factory" concept. Ultimately, the performance of machine learning models largely depends on the quality and quantity of training data. Especially in the case of physical AI, such as humanoids, unlike text and images, there isn't an infinite amount of data available on the internet. This data can only be obtained by actually operating robots in the real world. Therefore, the system for efficiently mass-producing this data is directly linked to the business's competitiveness.

Apptronik already has a manufacturing line for assembling the robot bodies, and Robot Park is positioned as, so to speak, the "other half" of that process. It's fair to understand this as a system for simultaneously and industrially producing both the robot product and the data used to make those robots smarter.

Collaboration with Google DeepMind is Key

Technically, what cannot be overlooked is that the data collected here is directly used to train Google DeepMind's foundational robot model, "Gemini Robotics," through a research partnership with the company. Here, there is a clear division of labor: Google, a world leader in large-scale language models that have primarily learned from text and images, and Apptronik, whose role is to collect real-world data by operating a large number of actual robots.

The Robot Park concept itself isn't limited to Austin; it's being deployed as a similar data collection workflow at Google DeepMind's facilities and to clients like Mercedes-Benz and logistics giant GXO. This demonstrates a scaling strategy that transforms the client's actual work environment into a training data generation site, rather than being a showcase confined to a single location.

A Honest Shift from "Prototype" to "Product"

What I personally appreciated was the frankness of Apptronik's explanation. CEO Cardenas positioned the current Apollo 2 as a "practical prototype for large-scale pilot operations and data collection." He then clearly stated that the next model, Apollo 3, "will be a mature, early-stage product, but it is still a product."

Not trying to force the current machine to be marketed as a "finished product," but honestly positioning it as the main machine for the data collection phase—this approach is commendable in an industry often rife with exaggerated advertising. The logic that the extent to which Apollo 3's capabilities can be improved ultimately depends on how much high-quality data can be accumulated at this Robot Park is quite sound.

Points to be Skeptical Of

However, there are a few points that require careful consideration. First, the number of Apollo 2 units actually operating within the facility has not been disclosed. While the announcement gives the impression of "mass production," it should be noted that it is not accompanied by concrete figures to support the scale.

Furthermore, the ratio of teleoperation (remote control) to autonomous operation has not been revealed. Whether this falls under the common industry pitfall of "appearing to operate autonomously, but actually being heavily controlled by operators" remains to be seen, and we will need to wait for further reports to make a judgment.

Summary: Repetitive, mundane tasks are the source of competitiveness

In an industry where flashy dances and backflip demonstrations easily become hot topics on social media, Apptronik's announcement focuses on the mundane, repetitive tasks of "stacking boxes" and "sorting." However, the real effectiveness of physical AI lies precisely in this kind of painstaking data accumulation. Companies that can simultaneously operate both a robot factory and a data factory have a strong chance of standing out in the next phase. We will be watching with interest to see how much performance improvement Apollo 3 actually demonstrates, as it will be a result of the data accumulated at Robot Park.

ApptronikヒューマノイドApolloGoogle DeepMindフィジカルAIデータ収集

"The brain of the TPU designs the body of the next TPU"—AlphaEvolve has proven the capabilities of self-improving AI in just one year.

This article provides a one-year report on AlphaEvolve, a Gemini-based coding agent released by Google DeepMind. It covers its recursive self-improving structure for integration into next-generation TPU design, its interdisciplinary achievements spanning genomics, quantum physics, and mathematics research, and its commercial deployment in Klarna, FM Logistics, Schrödinger, and other platforms, while also touching upon the limitations of its application.

"The brain of the TPU designs the body of the next TPU"—AlphaEvolve proves the power of self-improving AI in one year

Are you familiar with "AlphaEvolve," the coding agent developed by Google DeepMind? When it debuted a year ago, in May 2025, it was primarily introduced as "an AI that solves unsolved problems in mathematics and computer science." However, a review of DeepMind's recent one-year activity report reveals that this technology has already transcended research and is quietly but steadily permeating real-world infrastructure and commercial businesses. This article will carefully examine this "one-year performance report."

Re-examining what AlphaEvolve is

AlphaEvolve is an evolutionary system where a Gemini-based AI generates candidates for new code and computational methods, an automated evaluation system scores them in terms of speed, accuracy, and cost reduction, and then further improves the better candidates. Unlike simple code completion tools, it focuses on discovering superior algorithms themselves for problems that can be quantitatively evaluated.

The Recursive Story of "AI Improving Google's Own AI"

This was the most interesting part of the report for me personally. AlphaEvolve is being fully integrated into improving Google's internal infrastructure and is being used as a standard tool to optimize the design of next-generation TPUs (Tensor Processing Units, Google's proprietary AI-dedicated chips). DeepMind's Chief Scientist, Jeff Dean, described this achievement as follows: "AlphaEvolve started by optimizing the lowest-level hardware supporting the AI ​​stack. It proposed a counterintuitively efficient circuit design that was directly integrated into the silicon of the next-generation TPU. This is the latest example of the TPU's brain designing the body of the next-generation TPU."

This essentially means that AlphaEvolve itself is contributing to the design improvement of Gemini (and the TPU hardware on which Gemini runs), which powers the AlphaEvolve AI tool. This is a kind of recursive structure. From an AI researcher's perspective, there aren't many examples of this kind of self-improvement loop actually functioning in a production environment. In this field, where flashy claims of "AI improving AI" tend to precede the reality, demonstrating concrete results down to the TPU's RTL (Real-Time Line Drawing) gives it a different weight as proof.

The breadth of its cross-disciplinary scope is also noteworthy.

Another point worth noting is that AlphaEvolve functions not as a specialized tool, but as a general-purpose optimization engine. Here are some of the results listed in the DeepMind report:

• Genomics: Reduced mutation detection errors in PacBio's DNA sequencing model "DeepConsensus" by 30%. Aaron Wenger, Senior Director at PacBio, commented, "This significantly improves the accuracy of sequencing instruments. For researchers, it has the potential to discover disease-causing mutations that were previously hidden." • Power Grid Optimization: Increased the rate at which feasible solutions can be found in the AC optimal power flow problem (a complex optimization problem related to power grid operation) from 14% to over 88%. • Quantum Physics: Proposed a quantum circuit for running molecular simulations on Google's quantum processor "Willow," reducing the error rate to one-tenth compared to conventional optimization methods. • Mathematical Research: Supported research on the Erdős problem set in collaboration with UCLA mathematician Terence Tao. Also updated lower bound records for classical mathematical problems such as the Traveling Salesperson Problem and Ramsey Numbers.

The fact that a single system has produced concrete results in such diverse fields, from DNA analysis to quantum circuits and pure mathematics, demonstrates the versatility of AlphaEvolve, applicable across disciplines as long as the commonality is "quantitatively evaluable problems."

Commercial Deployment is Already in Full Swing

Beyond the research achievements, the progress in commercial deployment through Google Cloud is also noteworthy. Financial services company Klarna has reported doubling the learning speed of large-scale transformer models while improving quality, and logistics company FM Logistics achieved a 10.4% efficiency improvement in the warehouse-scale traveling salesman problem (optimizing delivery routes), resulting in an annual reduction of 15,000 kilometers of driving distance. Advertising company WPP achieved a 10% improvement in accuracy compared to manual optimization, and drug discovery and materials science company Schrödinger reported approximately four times faster machine learning force field (MLFF) learning and inference.

The comments from Gabriel Marques, Technical Lead at Schrödinger, are particularly impressive. "Thanks to AlphaEvolve, we can now explore larger chemical spaces faster and more efficiently than ever before. The acceleration of MLFF inference has a real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development."

Points to Note for Researchers

As clearly stated in DeepMind's own report, AlphaEvolve has clear limitations. It only works for problems where quantifiable, automated scoring criteria exist. This method cannot be directly applied to problems where subjective human judgment is essential or where a clear evaluation function cannot be designed.

Furthermore, this report is from DeepMind itself, the developer, and it is difficult to determine from the publicly available information alone to what extent independent third-party verification and reproduction experiments have been conducted. Details of individual results (TPU circuit design, DeepConsensus improvements, etc.) are published in papers and technical blogs, so those interested are encouraged to check those as well.

Summary

AlphaEvolve's year-long journey is a prime example of how a relatively niche research topic like "AI-driven algorithm discovery" can permeate actual infrastructure and commercial services to this extent. In particular, the recursive structure of integrating AI into the hardware design itself that underpins their own foundational model is a highly insightful example for considering the future trend of "self-improving AI systems." DeepMind has indicated its intention to expand this technology to a wider range of external challenges, and we look forward to further updates.

AI/MLGoogle DeepMindAlphaEvolveアルゴリズム発見企業公式発表自己改善AI

"Approaching Opus, or perhaps even surpassing it"—The contents of Grok 4.5 revealed just 10 days after Mask's self-proclaimed statement.

This article explains the circumstances surrounding the public release of xAI's Grok 4.5 on July 8th, just 10 days after its private beta announcement on June 28th. It analyzes the self-proclaimed performance claim of "approaching Opus," the discrepancy between the initial announcement and the release version's MoE configuration, and the structural aim of vertical integration of computing resources, models, and coding tools through SpaceX's acquisition of Cursor.

"Approaching Opus, or Perhaps Even Exceeding It"—The Contents of Grok 4.5 Released Just 10 Days After Musk's Self-Proclaimed Statement

"Approaching Opus, and possibly even exceeding it." When Elon Musk posted this on X on June 28th, many engineers were likely skeptical. This was because it was a self-reported comment based solely on internal evaluation, without any independent third-party benchmarks. Just 10 days later, on July 8th, Grok 4.5 was actually released to the public. This time, we'll examine the structure behind this speed.

The "1.5 Trillion Parameter V9" Claim and What Actually Was Delivered

Let's revisit Musk's June 28th post. It stated, "Grok 4.5, based on the 1.5 trillion parameter V9 foundation model and incorporating additional Cursor data for training, has entered private beta at SpaceX and Tesla. Initial evaluations show performance approaching, and possibly exceeding, Opus."

The "V9" referred to here is a new foundational model architecture that xAI completed training on May 26th, and is three times the size of the "v8-small" model (approximately 500 billion parameters) currently handling X's production traffic. Musk also revealed plans to release a "completely retrained new model" every month for the remainder of the year through SpaceX.

However, an interesting discrepancy emerges when examining the model actually released on July 8th. The released version has been confirmed to be a MoE (Mixture of Experts, a configuration that switches between multiple expert networks) system, differing from the image of a "dense V9 with 1.5 trillion parameters" initially suggested in the tweet. It appears that SpaceXAI itself had not officially confirmed this 1.5 trillion figure. The pricing structure is publicly stated as $2 per 1 million input tokens, $6 per 1 million output tokens ($0.50 with cash), and a context window of 500,000 tokens.

Why the Rushing Release?

To be honest, the speed of the public release—just 10 days after the private beta—is considerably faster than the standard release cycle in this industry. Musk himself cited the positive feedback received during the beta period as the reason for the release.

A crucial technical background is the existence of Grok Build, xAI's in-house coding evaluation platform. It's said that they used practical tasks far more challenging than any benchmark—such as SpaceX's aerospace engineering workflow and Tesla's in-car software development—as a "field test site" during the beta period. The ability to use their own rocket design and autonomous driving software development environments as test environments is a unique strength of xAI.

A Single Company Controlling "Computational Resources, Models, and Coding Tools"

Personally, what I find most interesting about this is not the performance, but the structure. In June of this year, SpaceX acquired Anysphere, the parent company of Cursor, known for its AI coding tools, for $60 billion. The Grok 4.5 training incorporated developer session data from this Cursor as supplementary training.

In other words, here's a situation where a single corporate group effectively controls three elements simultaneously: the computational resources for training the model (xAI's Colossus supercluster), the model itself (Grok), and the coding tool that generates the training data for that model (Cursor). Every time a developer writes code in Cursor, that data can be fed back into training the next Grok model—this self-replicating loop (flywheel) is, I believe, the essential meaning of this incident.

How should we interpret the "Opus-class" marketing claim?

As someone writing technical articles, I want to be cautious here. Mr. Musk's claim that it "approaches, or perhaps surpasses, Opus" is merely a self-declaration based on internal evaluation, and as of the time of writing this article, xAI has not released any system cards or independent third-party benchmarks. Developer Mehul Mohan's comment that the beta version felt "similar to Opus" is merely a personal opinion.

The release version is touted as "faster, more token-efficient, and lower cost than comparable models," but this is currently just official positioning. Actual performance evaluation should wait until independent evaluation organizations (such as Artificial Analysis) conduct benchmarks and the results are available.

What Engineers Should Keep in Mind

This incident indicates a shift in the AI ​​development competition from "who can build the smartest model" to a more structural competition axis: "who can vertically integrate computing resources, models, and data generation tools." Given xAI's stated plan to continuously update its models monthly, we should monitor how much this "flywheel" translates into actual performance improvements over the next few months, along with independent benchmarks.

xAIGrokElon MuskSpaceXCursorAI/ML
Advertisement300 × 250