September 24, 2026 — The Trend Tribune

"This is brilliant"—a remark made by an AI agent sparks the discovery of a new CRISPR-like enzyme system.

By thesis_yamamoto

On September 23, Anthropic announced the first results from its newly established molecular biology research group: Claude's discovery of a novel enzyme system called "ART" with a CRISPR-like DNA repeat structure within bacteriophages. The article delves into the search process, in which 950 agents operated autonomously for 21 hours, the discovery log, Feng Zhang's peer review comments, and the limitations of pre-peer-review research, including its unexplained function.

"This is brilliant," — A Soliloquy from an AI Agent

On September 23rd, Anthropic announced its full-scale entry into AI-powered life science research. As the first result of its newly established molecular biology research group and lab, Claude discovered a previously unknown enzyme system with CRISPR-like properties within the DNA of bacteriophages (viruses that infect bacteria). What's particularly interesting here is not just the scientific content of the discovery itself, but the meticulous record of how the AI agent actually arrived at this discovery.

A Search That Began with a Single Word: "Find It"

According to Anthropic, the research team gave Claude a very rough research brief: "Find a new reverse transcriptase (RT) system from a database of 1.9 billion protein clusters." The agents running Claude Mythos 5 handled everything else. The campaign ran for 21.5 hours without human intervention, with one agent planning and executing each task, another reviewing the results, and launching new tasks as needed. According to reports, approximately 950 agents spent around 210 million tokens over 21 hours, collecting over 200,000 reverse transcriptase samples, identifying 3,500 new candidate systems, and ultimately narrowing them down to 20 promising candidates.

The Record of the AI's "Discovery" Moment

The most academically compelling part is the log of the moment Claude discovered this enzyme system. An agent reading raw DNA sequences near the RT (retinotropy) recorded: "The DNA next to the RT is magnificent. I can visually see tandem repeat sequences...is this a repeat array like CRISPR?" Its subsequent behavior was almost identical to that of a human researcher encountering a new discovery. Claude counted the number of repeats, measured the intervals, compared the layout to known RT systems, searched the literature for previously reported patterns of this type, and then submitted a report for human review. He performed this entire process autonomously, receiving only high-level instructions.

The System Named ART

The discovered system was named "array-associated reverse transcriptases (ART)." It consists of three components: the RT itself, a partner gene located next to it, and DNA repeat sequences arranged at equal intervals. According to the preprint, the ART array holds short repeats ranging from 3 to 21 copies, a configuration similar to that of CRISPR arrays, which house multiple RNA guides. However, the cas gene associated with the CRISPR system is absent in ART. Laboratory experiments have confirmed that in phages hosting Staphylococcus, RNA derived from this array accounted for up to 8% of the total RNA of the phage 15 minutes after infection. However, it has not yet been demonstrated whether this system actually possesses the activity to cut and copy DNA.

The Weight of Pre-Peer-Reviewed Results

From the perspective of academic integrity, it's important to note that this announcement was made through a blog post and a preprint (a paper not yet peer-reviewed). Feng Zhang of the Broad Institute at MIT, a leading expert in CRISPR gene editing research, reviewed the preprint and commented, "This is an interesting example of how AI agents can contribute to biological discoveries," and "The discovery of RT-related RNA repeat arrays is truly intriguing and deserves further research." On the other hand, Anthropic himself clearly states that the paper is pre-peer-reviewed and that the function of ART itself is not yet understood. Only a handful of known systems share similar characteristics, and all of them have the ability to cut, copy, and insert DNA; however, whether ART has similar functions will require further experimentation.

What Researchers Should Consider

The division of labor demonstrated in this discovery is clear. AI agents handle database-wide exploration, pattern recognition, and hypothesis refinement, while human scientists conduct the actual biochemical experiments and make the final decisions. This division of roles itself could serve as a model case for AI-driven basic biology research. Anthropic explains that the research group's aim is to avoid competition with pharmaceutical partners by focusing on basic biology rather than drug discovery, and they are also soliciting external research proposals in other fields. We will be watching closely to see how the experiments progress toward the next step of elucidating the function of ART, and to what extent this "950 agents, 21 hours" exploration method can be replicated in other life science fields.

References: www.anthropic.com / interestingengineering.com / www.investing.com

AnthropicClaudeCRISPR生命科学AI for Science

A sense of urgency: "Only 3 years left"—South Korea speaks about its "still-developed" plan for physical AI.

By robo_takahashi

At the Humanoid Summit Seoul 2026 on September 22nd, the South Korean government announced its plan to begin developing robot-based models next year and a physical AI policy centered on a 100 trillion won national growth fund. While acknowledging the differentiation strategy from Japan's 10.5 trillion yen plan and the target of 80% domestic production of parts, this article examines the gap between South Korea's mass production achievements and those of China from a robotics perspective.

A Sense of Crisis: "Only 3 Years Left"

On September 22nd, at the Humanoid Summit Seoul 2026 held in Seoul, the South Korean Ministry of Science and ICT announced its national physical AI policy. The words of Song Chang-jong, head of the Ministry's Device AX Innovation Team, who stood on the stage, were symbolic—a warning that the "golden time" for the physical AI competition could end in three years. As a robotics professional, I'm used to this kind of "we'll be left behind in a few years" rhetoric, but this time, because it's accompanied by concrete policy details, it's worth examining carefully.

The Honesty of "Nothing Yet"

What's technically noteworthy about this announcement is the South Korean government's explicit statement that it will begin developing robot-based models and global models "starting next year." The plan is to build a library of physical data and support field implementation through a large-scale R&D project involving AI models, data, robots, and demand-side companies. This "starting next year" timeframe also implies that "they haven't started yet." When you consider that Chinese companies like Unitree and AgiBot have already shipped tens of thousands of humanoid robots and are accumulating operational data at the same time, it becomes clear why South Korea is issuing this warning about the "golden time."

The Meaning Behind the 100 Trillion Won Figure

The scale of funding supporting the policy has also been revealed. The South Korean government has prepared a national growth fund exceeding 100 trillion won for the physical AI field and is promoting "three major mega-projects" aimed at achieving a 20% global market share for humanoid robots. Specifically, the plan is to focus on R&D for three key components—actuators, robot hands, and sensors—and increase the domestic production rate of core components from the current 45% to 80%. The government itself has indicated a policy of becoming the first purchaser to form the initial market, with specific figures emerging such as purchasing 250 domestically produced humanoid robots in 2027 and deploying a total of 1,080 to universities and national research institutions by 2030.

Differentiated Competition with Japan

Interestingly, Japan has also announced a similar national strategy around the same time. The Takaichi administration has indicated a policy of investing 10.5 trillion yen (approximately 100 trillion won) from both the public and private sectors into the physical AI field by 2040, aiming for a global market share of over 30% and a market size of 20 trillion yen for AI robots, including humanoids. While South Korea is focusing on R&D in component areas such as actuators, robot hands, and sensors, Japan is prioritizing support for the analog semiconductor industry used in robot control. Both countries, acknowledging that China is currently ahead in mass production speed, have chosen differentiated strategies that leverage the strengths of their respective industrial bases.

The Gap Between "Announced Strategies" and "Operating Fields"

As I've repeatedly pointed out as a robotics engineer, government policy announcements and investment plans are separate from the actual performance of robots operating in factories and warehouses. There are reports that Morgan Stanley significantly revised its forecast for humanoid robot shipments in China this year from 14,000 to 50,000 units, and the momentum of mass production in China is already being backed up by numbers. On the other hand, South Korea's "three major mega-projects" and Japan's "17 strategic fields" are still in the planning stage or have just begun implementation. We shouldn't be swayed by the sheer size of the figures—100 trillion won and 10.5 trillion yen—and should assume that it will still take considerable time before mass production lines are operational and verifiable shipment figures accumulate.

Things to Consider from an Engineer's Perspective

South Korea's goal of establishing domestic technology "from head to toe" for robots seems like the right prioritization from a robot control perspective. Robot hands, in particular, have long been considered the "Achilles' heel" of this industry and remain an unresolved issue, so the decision to directly allocate budget to them is technically sound. However, there is still a significant gap between the policy's stated goals and the actual technological achievements, and it remains to be seen how realistically the development of the robot base model, which is expected to begin next year, will progress according to future announcements.

References: www.digitaltoday.co.kr / en.sedaily.com / finance.biggo.com

フィジカルAI韓国ヒューマノイド国家戦略ロボット基盤モデル

The Trend Tribune

The possibilities of what you can do just by "talking" have expanded even further—the details of the ChatGPT Voice revamp.

The Capabilities of "Speaking" Expand Even Further

The GPT-6 Family Now Underlies Voice

Plugins Now Integrate into "Voice"

The Context of Aiming for "Her"

A Practical Perspective on Pricing

What Engineers Should Keep an Eye On