July 24, 2026 — The Trend Tribune

By tech_tanaka · July 24, 2026

"Distillation in just a few weeks is impossible"—White House accusations and the contradictions in the timeline pointed out by experts. — (Photo: illustrative)

"Distillation in a Few Weeks is Impossible"—Contradictions in the Timeline Pointed Out by the White House Allegations and Experts

On July 22nd, Michael Kratsios, Director of the White House Office of Science and Technology Policy (OSTP), made a bombshell statement in a post on X. He accused China's Moonshot AI of systematically and extensively "distilling" Anthropic's model "Fable" and reusing it in the development of its own model, "Kimi K3." Treasury Secretary Scott Bessent also mentioned the possibility of sanctions, demonstrating a unified government response. However, several experts have pointed out that this accusation is quite far-fetched when the timeline is carefully examined. This article will examine this case from both the perspective of the accusation itself and the counterarguments, maintaining a balanced approach.

What is "Distillation" in the First Place?

First, let's clarify the technical premise. "Distillation" is a technique that uses the output of a large, high-performance model (trainer model) to train a smaller model (student model), thereby replicating some of the teacher model's capabilities at a lower cost. Kratsios himself states in his post that "legitimate AI distillation plays a crucial role in this open innovation ecosystem in creating smaller, more efficient models," indicating he is not criticizing the distillation technique itself. His criticism lies precisely in the distinction between "industrial-scale distillation carried out on a large scale and stealing proprietary US technology".

Allegations: GB300 Servers and Access via Thailand

Kratsios's claims consist of two main pillars. First, he alleges that Moonshot "developed a sophisticated internal platform capable of quickly switching between multiple access methods while evading detection, for large-scale distillation of US models." Second, he claims that Moonshot owned NVIDIA GB300-equipped servers and accessed the GB300 from within Thailand. The GB300 is an NVIDIA Blackwell-generation chip, and its export to Chinese companies is prohibited by US regulations.

Anthropic itself had already filed an accusation in February of this year alleging that Moonshot had violated its terms of service by using a fraudulent account to send over 3.4 million distillation queries to Claude. The White House's statement now officially follows up on this accusation made at the private company level.

Experts Point Out "Chronological Inconsistencies"

This is the point I particularly want to highlight. According to a TechCrunch article interviewing several experts, there are concerns about the technical consistency of this accusation.

Anthropic released Fable 5 to the public on July 1st. The Kimi K3 appeared just a few weeks later. If a large-scale distillation—running millions of queries against Fable 5, incorporating the output into training data, and retraining the model to be competitive enough to warrant the White House's concern—was truly performed within this timeframe, it would require compressing the normal development cycle to a level that AI researchers would consider "unrealistic." Large-scale distillation typically takes months, not weeks.

However, for the sake of fairness, it should be emphasized that this does not prove the allegations are false. The White House's public statement was merely an "allegation," not a technical reconstruction of how the model was created. It remains possible that the government made this statement after scrutinizing confidential information (intelligence). Several media outlets, including those in the Japanese-speaking world, have also reported that Kratsios's initial post used the phrase "solid evidence," which was later changed to "information."

Specifications of Kimi K3 itself should also be considered

Apart from the validity of the allegations, it is also important to touch upon the specifications of the Kimi K3 model itself. Released on July 17th, this model is a 2.8 trillion parameter Mixture-of-Experts (configuration that switches between multiple expert networks) model designed to activate 16 out of 896 experts per token. It is considered the largest open-weight model released to date, boasting a 1 million token context window and native visual comprehension capabilities.

In benchmarks, it achieved 91.2 points on BrowseComp and 88.3 points on Terminal-Bench 2.1, ranking first on the front-end coding leaderboard and surpassing even Fable 5. While its overall performance is still considered slightly behind Fable 5 and OpenAI's GPT-5.6 Sol, some reports suggest the difference was so small it unsettled AI and semiconductor investors.

Another Realistic Issue: The GPU Black Market

Regarding the chip access allegations, comments from Sam Bresnick, a researcher at Georgetown University's Center for Security and Emerging Technologies, are interesting. The allegations suggest that a black market actually exists for regulated chips like the GB300. The case of the founder of US server manufacturer Supermicro, who was indicted in May of this year for smuggling high-performance chips into China, is also cited. This highlights a more structural issue: the effectiveness of export controls.

What Engineers Should Consider

The reason this news is treated as a technical article is that, more than just another scene in the US-China conflict, it exposes a challenge for the entire AI industry: how to verify the provenance of a model. Benchmark scores can be measured by independent evaluators, and running workloads reveals how a model performs for your intended use. However, verifying the provenance of a model—how it was created—is a far more difficult area than benchmark numbers alone can determine.

NVIDIA has avoided direct comment on the allegations, only issuing a statement to the effect that it "complies with all export control regulations and actively enforces compliance across all sales channels." Moonshot AI itself has not yet publicly refuted these allegations. It has also been reported that the U.S. Department of Commerce's Bureau of Industry and Security (BIS) has launched an investigation into whether Chinese companies are circumventing regulations by using overseas subsidiaries or cloud infrastructure in third countries. We will be watching closely to see if this case will escalate into actual sanctions or simply fizzle out.

References: techcrunch.com / qz.com / mlq.ai / xenospectrum.com / easternherald.com

Moonshot AIKimi K3Anthropic輸出規制AI蒸留米中対立

Where did the figure of "50,000 Optimus units sold" come from? — The gap between "announcements" and "verified" in the humanoid industry.

By robo_takahashi

This article introduces a cross-industry verification that figures circulating online, such as "50,000 Optimus units sold" and "Over 10,000 Figure AI units deployed," have not actually been officially announced by any of the companies. It examines the small but verifiable track record of companies like Agility (Digit, 7 units, Toyota) and Figure (BMW, single-digit numbers), the discrepancy between Unitree's shipment figures and revenue, the unprecedented challenge of mass-producing 10,000 parts for a manufacturing company, and the issue of China's National Intelligence Law, emphasizing the importance of verifying the source of announced figures.

Where Did the "50,000 Optimus Units Sold" Figure Come From? – Between "Announcements" and "Verified Figures" in the Humanoid Industry

"Tesla's Optimus has surpassed 50,000 units sold," "Figure AI has deployed over 10,000 units in partner warehouses"—these figures are frequently seen online. However, I, Takahashi, want to pause and consider this point. None of these figures were actually announced by the companies themselves. This time, I'd like to examine each of the "numbers" widely circulated in the humanoid industry.

Tesla's Own Statements Are Quite "Unremarkable"

Let's start with the Tesla Optimus, which has been the subject of the most sensational rumors. On July 1st, Elon Musk posted a photo on X of himself walking along the production line at the Fremont factory, causing a stir in the market. However, Tesla itself has never officially announced any production figures for the Optimus—whether audited or not.

Rather, Tesla's own statements, especially in the legally weighted setting of financial reporting, are quite restrained. At the Q4 2025 earnings call, Musk explicitly stated that Optimus is "not being used extensively in factories" and that "the number of units being produced is primarily for training purposes." Tesla's Q2 earnings call, scheduled for July 22nd, may be a pivotal moment where verifiable figures for this production volume are finally revealed. It's important to clarify that, at least as of the time of writing this article, neither the "50,000 units total" nor the "over 1,000 units operating in factories" figures have come directly from Tesla itself.

The Companies with Verifiable Track Records Are Actually Different

In contrast, Agility and Figure have accumulated relatively verifiable track records. Agility Robotics is said to have 7 commercial Digit units operating at a Toyota facility after a year-long pilot operation. Figure AI has a track record of production integration at BMW, but the number of units is reported to be in the single digits.

You might think, "Only 7 units?" However, this isn't a flashy number; rather, it's a valuable data point because it represents the actual number of verifiable units continuously operating in a third-party factory. Apptronik's "Robot Park," which I introduced in a previous column, also kept its actual operational unit count undisclosed. Looking at the humanoid industry as a whole, it becomes clear that companies that actually disclose their operational unit count are in the minority.

Chinese Companies' "Quantity" is Real, But "Profitability" is Another Story

Looking at Chinese companies reveals a different picture. Unitree plans to ship approximately 5,500 humanoid robots in 2025, and aims for 10,000 to 20,000 units in 2026. This is an overwhelming number of units shipped compared to its Western competitors. However, it's important to note that a large number of units shipped does not necessarily mean high profitability. Unitree's net profit for the first quarter of 2026 is reported to have fallen by 52% year-on-year. This comes despite the fact that robotics-related stocks were soaring around the same time, fueled by speculation about Optimus's funding and Unitree's IPO approval.

In this context, we should also re-examine Agibot (claiming 15,000 cumulative shipments and a 39% market share), which I introduced in this column last week. According to the current source, Agibot achieved a task success rate of over 99% over 140 hours of pilot testing in April of this year, and recently reported that its G2 units are operating at a throughput of 310 units per hour and a cycle time of 18-20 seconds. While the figures themselves are specific, it's important to remember that this is self-reported data from the company and has not undergone independent third-party verification.

An Unprecedented Manufacturing Challenge: Mass Production of 10,000 Parts

From a technical standpoint, what I personally find important is the difficulty of this manufacturing challenge, as pointed out in this article. There is no precedent in the history of manufacturing for mass-producing an electromechanical system consisting of approximately 10,000 parts using serial production. For reference, the article cites the fact that even Tesla's own Model 3 ramp-up (2017-2018, the so-called "production hell"), despite being a far simpler product, took over 18 months to reach a stable mass production system.

This is something I, as a software engineer, can readily relate to. When the number of new parts reaches the scale of 10,000, the yield of individual parts, the redundancy of the supply chain, the reproducibility of the assembly process—none of these can be solved by theoretical design alone. The weekly production plan for the Tesla Optimus Gen3, which I previously mentioned (a phased launch from several dozen units in June to 1,000 units in September), can be understood as a cautious figure that took into account these realistic constraints of manufacturing.

Another point to consider regarding Chinese-made robots

Aside from the discussion of technology and manufacturing, I would like to introduce another point raised in this article. Because Unitree is a Chinese company, it is subject to China's National Intelligence Law (enacted in 2017). This law legally obligates organizations under Chinese jurisdiction to cooperate with national intelligence activities, and it applies regardless of where the products are deployed or where the servers are located. This is the same point I touched upon in last week's Agibot article, and it should be noted as a compliance risk that adopting companies should consider, separate from performance and price competitiveness.

Summary: "Quantity," "Revenue," and "Verifiability" should be viewed from different perspectives

Organizing this series of information, it becomes clear that the reality of the humanoid robot industry needs to be evaluated from at least three different perspectives. The "volume" of units shipped (Unitree is overwhelmingly dominant), actual profitability (even with high volume, profit margins are a different matter), and the actual operational status verified by third parties (Agility and Figure have a small but verifiable track record). It is dangerous to confuse these three factors and simply conclude that "Company A is the most advanced."

In this industry where flashy numbers tend to take on a life of their own, developing the habit of checking "who announced those numbers and on what basis"—it may seem mundane, but this story reminded me once again that this is the most important attitude to have when pursuing this field.

References: www.technology.org / humanoid.guide / www.techtimes.com / www.humanoid.press

ヒューマノイドTesla OptimusFigure AIAgility RoboticsUnitree検証

Qwen overtakes Llama for the first time—170,000 papers reflect the AI research landscape for the first half of 2026

By thesis_yamamoto

This article explains the AI Papers Academy's report on AI research trends for the first half of 2026, based on an analysis of 170,927 arXiv papers. It highlights the turning point where Alibaba's Qwen surpassed Meta's Llama in reference share for the first time, the rapid expansion of agent research (which didn't exist 18 months ago), and the sheer size of the inference and alignment fields, while also touching upon the limitations of keyword matching as a method.

Qwen Overtakes Llama for the First Time – 170,000 Papers Reflect the AI Research Landscape in the First Half of 2026

Counting the number of models mentioned in the text of a paper, rather than just the number of citations, reveals a completely different picture of the AI research landscape. This article introduces an analysis of AI research trends in the first half of 2026, conducted by the AI Papers Academy data analysis team, based on a massive 170,927 papers submitted to arXiv. While seemingly a mundane aggregation process, the resulting picture is quite fascinating.

Analysis Method: Filtering 170,000 Papers by Keywords

First, let's introduce the methodology. The analysis targeted papers submitted to arXiv from early 2025 to June 26, 2026, in four major categories: Computer Science & Natural Language Processing (cs.CL), Computer Vision (cs.CV), Machine Learning (cs.LG), and Artificial Intelligence (cs.AI). The method employed involves matching the title and abstract with a pre-prepared set of keywords for topics, model families, and research institutions.

The analysis period is divided into three halves: "First Half of 2025," "Second Half of 2025," and "First Half of 2026." Given the approximately 25% increase in the total number of papers in this field, the design, which tracks the change in "share" within the overall total rather than simply comparing the number of papers, seems appropriate as an analytical method.

The "Center of Gravity of Openweight AI" Shifts

The most impactful finding in this analysis is the reversal in the reference share of openweight models. Alibaba's Qwen has surpassed Meta's Llama for the first time. Looking at the specific numbers, references to Qwen nearly doubled from 752 to 1,489 (+98%). Llama, on the other hand, only increased by 14%, from 1,085 to 1,232.

This is a significant turning point for AI researchers. Until now, the position of "easy-to-use, high-performance base model" has almost consistently been held by Meta's Llama. This data marks the first instance where a Chinese-developed open-weight model family has become the most referenced model in Western academic research. It's a sign that the "default" for researchers choosing base models is quietly but surely shifting.

"Agent Research" Rapidly Expands from a Previously Non-Existent Domain

Another point worth noting is the change in research themes themselves. Many of the topics dominating the top growth rates in market share are agent-related fields that barely existed as independent research areas 18 months ago. The analysis team describes this trend as "the agent-based workflow field maturing, and the growth of its sub-fields reflects the direction of development in this field itself." The observation points out that research interest is shifting from the stage of "Can we create an agent?" to more implementation-level challenges such as "How can we make an agent plan, reason, use tools, and evaluate its own output?"

On the other hand, the positioning of the topic "Reasoning & CoT" is interesting. While it ranks 9th in terms of market share growth, it boasts an overwhelming scale with 11,636 papers, ranking 1st in absolute numbers. It's important to note that because the entire field is so vast, even a modest market share growth represents thousands of new papers. Similarly, the field of "Alignment & AI Safety," while ranking 10th in market share growth, has already recorded a large base of 8,121 papers and is still showing a 33% market share growth.

How should we interpret this analysis?

As a researcher, there are several points to keep in mind when presenting this type of analysis. First, this analysis relies on keyword matching and is not a qualitative analysis that scrutinizes the content of the papers themselves. Since it counts the frequency of a particular model name or topic name appearing in the text, it cannot distinguish between references for comparative or mention purposes and research that actually uses that model.

Furthermore, the AI Papers Academy, the entity conducting the analysis, is not an independent academic institution, but a media outlet specializing in AI-related paper reviews. While this is an ambitious undertaking for a large-scale data analysis, it should be kept in mind that it has not undergone academic peer review. Nevertheless, an analysis based on data from 170,000 papers that shows such concrete numerical changes (a reversal of share from +98% to +14%) will be a valuable reference point that supports the industry's intuition.

Thoughts as a Researcher

This type of analysis, which attempts to capture "what researchers are actually referencing and using now" rather than "evaluation that accumulates afterward" like citation counts, is significant in anticipating trends in AI research. The geographical shift in leadership of open weight models, and the shift in research focus from "the performance of single models" to "the behavior of agents"—we will continue to closely monitor how these two trends will change the landscape of AI research from the latter half of 2026 onward.

References: aipapersacademy.com / medium.com

AI/ML論文arXivQwenLlama研究動向

"Distillation in just a few weeks is impossible"—White House accusations and the contradictions in the timeline pointed out by experts.