人工智能剛剛創造出自然界中不存在的病毒（2/2）

Recently The New York Times reported the following:

This A.I. Just Created Viruses Not Found in Nature (2/2)

Scientists trained artificial intelligence on libraries of DNA and then asked the model to create recipes for viral genomes. Sixteen of them were viable, yielding new viruses.

The NYT - By Carl Zimmer - Carl Zimmer covers news about science for The Times and writes the Origins column.

Aug. 6, 2026

(continue)

They made this choice in part because scientists know Phi X-174 intimately, having studied it for close to a century. And because it’s a bacteriophage that infects only E. coli, they knew that viruses similar to it would be safe.

Once Evo got familiar with the genomes of Phi X-174 and its kin, the researchers prompted the model to write new versions of its own. Evo generated 700,000 potential versions; the scientists pursued only the ones that looked as if they had the best odds of succeeding.

They ended up making DNA molecules from 285 of Evo’s suggested sequences. When those genomes were ready to test, Mr. King and his colleagues inserted them into bacteria, which they spread across petri dishes.

In many of the dishes, the microbes grew peacefully — Evo’s genomes had failed. But in one dish, researchers noticed clear dots appearing in the cloudy film of bacteria, the telltale sign of multiplying viruses.

The DNA that the scientists had inserted into the bacteria had produced protein shells that contained viral genes. The viruses burst out of the cells, leaving behind the ruptured husks of their dead hosts.

As Mr. King and his colleagues tested more genomes, they saw more clear dots. All told, they discovered that 16 of Evo’s genomes produced viable new viruses.

They proved to be as resilient as natural ones. In fact, some multiplied faster than Phi X-174. “They’re not just sickly versions of stuff that already exists,” said Oliver Crook, a protein chemist at the University of Oxford who was not involved in the new study.

Dr. Crook cautioned that Evo’s viruses were not radically new creations. They tend to be very similar to natural species, relying on the same underlying biology.

Scientists will have to run more experiments to find out if Evo has the same success rate if it’s trained on other groups of viruses. If so, Dr. Crook expected that scientists might find some A.I.-generated viruses that could become useful tools for medicine and biotechnology.

“A lot of our science rests on viruses as technology,” he said. To treat people with genetic disorders, for example, doctors will load genes into viruses, which deliver them into cells.

But along with this hope, Evo’s initial success has raised concerns that A.I. could be used to create deadly pathogens. “You could say, ‘Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal,’” Dr. Hanke speculated.

The potential dangers were already on the minds of the researchers as they trained Evo. They did not provide the model with data about viruses that infect humans, and they excluded similar viruses that infect other animals, plants and fungi.

As a result, Evo can’t generate genomes of viruses that could threaten humans. “We just wanted to be extra careful,” said Brian Hie, a computational biologist at Stanford University and an author of the new study.

Dr. Hanke credited Dr. Hie and his colleagues for taking that precaution. “I think that’s quite commendable,” he said, “because they don’t get any guidance from anywhere on what they should be doing.”

Last week, Dr. Hanke noted, the National Institutes of Health rolled out a new policy for stopping high-risk life science research. The policy would bar scientists from experiments that would make biological agents more harmful.

But computer-based research — such as generating virus DNA with A.I. — “is not prohibited by this policy unless it involves an entity of concern,” the agency said in a statement.

It’s easy to tell if a natural virus like smallpox is an entity of concern. But Dr. Hanke said there’s no consensus on judging the possible danger of a virus made by an A.I. model.

“What is the risk of what I’ve never seen before?” he asked.

Translation

人工智能剛剛創造出自然界中不存在的病毒（2/2）

科學家利用DNA庫訓練人工智能，然後讓模型產生病毒基因組的「配方」。其中16個配方成功生成了新的病毒

（繼續）

他們之所以選擇這種病毒，部分原因是科學家對Phi X-174噬菌體非常了解，已經研究了近一個世紀。而且由於它是一種只感染大腸桿菌的噬菌體，科學家知道與其類似的病毒是安全的。

一旦 Evo 熟悉了Phi​​ X-174及其近親的基因組，研究人員就讓模型生成自己的新版本。 Evo產生了70萬個潛在版本；科學家只選擇了那些看起來最有可能成功的版本來研究。

他們最終利用Evo提供的285個序列建構了DNA分子。當這些基因組準備好後，King先生和他的同事將它們植入細菌，並鋪開在培養皿中。

在許多培養皿中，微生物生長良 - 但Evo的基因組未能成功複製。但在一個培養皿中，研究人員注意到細菌的渾濁薄膜上出現了清晰的斑點，這是有病毒在繁殖中的明顯跡象。

被科學家植入細菌的DNA產生了含有病毒基因的蛋白質外殼。病毒從細胞中破殼而出，留下宿主細胞破裂的殘骸。

隨著King先生和他的同事測試更多基因組，他們觀察到了更多清晰的斑點。最終，他們發現Evo提供的16個基因組產生了可行的新病毒。

這些病毒與天然病毒一樣能迅速適應環境變化。事實上，有些病毒的繁殖速度甚至超過了Phi X-174。並未參與這項新研究的牛津大學蛋白質化學家Oliver Crook說道: 「它們並非只是現有病毒的弱態版本」。

Crook博士提醒說，Evo 生成的病毒並非全新物種。它們往往與天然病毒非常相似，依賴相同的生物學基礎。

科學家需要進行更多實驗，以確定如果 Evo 使用其他病毒群進行訓練，其成功率是否相同。如果成功，Crook博士預計科學家可能會發現一些人工智能生成的病毒，這些病毒有望成為醫學上和生物技術領域的有用工具。

他說:「我們的許多科學研究在技術上都依賴病毒」。例如，為了治療遺傳性疾病患者，醫生會將基因加載到病毒中，然後由它導入細胞。

然而，在帶來希望的同時，Evo 的初步成功也引發了人們的擔憂：人工智能可能會被用於製造致命病原體。 Hanke博士推測地説: 「你可以說，『嘿，基因組語言模型，幫我產生一個經過改造的流感病毒基因組，使其更具傳染性或更致命』」。

研究人員在訓練Evo模型時就已經考慮到了潛在的危險。他們沒有向模型提供感染人類的病毒數據，並且排除了類似能感染其他動物、植物和真菌的病毒。

因此，Evo無法產生可能威脅人類的病毒基因組。史丹佛大學計算生物學家、這項新研究的作者之一Brian Hie說:「我們只是想格外謹慎」。

Hanke博士讚揚了Hie博士及其同事採取的這種預防措施。 他說:“我認為這非常值得稱讚”，“因為他們沒有得到任何關於應該做什麼的指導。”

上週Hanke博士指出，美國國立衛生研究院推出了一項新的政策，旨在停止高風險的生命科學研究。這項政策將禁止科學家進行任何可能使生物物質更具危害性的實驗。

但該機構在聲明中表示，基於電腦的研究 - 例如利用人工智能產生病毒DNA -「並不在此政策的禁止範圍之內，除非它涉及令人擔憂的實體」。

判斷像天花這樣的天然病毒是否令人擔憂是很容易。但Hanke博士表示，對於如何判斷人工智能模型產生的病毒可能存在的危險，目前尚無共識。

他問:「我從怎會知道從未見過的事物會帶來什麼風險？」。

              So, for the first time, scientists have used artificial intelligence to create new kinds of viruses, raising hopes for medical advances while also raising the disturbing possibility that the technology could someday be used to invent dangerous pathogens. DNA has its own rules of grammar and biologists have uncovered some of nature’s grammatical rules, but many remain a mystery. In the present case, Evo the A.I. had scanned about nine trillion nucleotides. Evo eventually recognized patterns common across the tree of life and used them to generate blueprints for new genes encoding proteins that could perform specific jobs. Apparently, we should be very careful in doing these kinds of researches because we do not know what is the risk of what we’ve never seen before.

Note:

1. A bacteriophage (噬菌體), also known informally as a phage, is a virus that infects and replicates within bacteria. Bacteriophages are composed of proteins that encapsulate a DNA or RNA genome, and may have structures that are either simple or elaborate. Their genomes may encode as few as four genes (e.g. MS2) and as many as hundreds of genes. Phages replicate within the bacterium following the injection of their genome into its cytoplasm. Bacteriophages are among the most common and diverse entities in the biosphere. Bacteriophages are ubiquitous viruses, found wherever bacteria exist. (Wikipedia)

人工智能剛剛創造出自然界中不存在的病毒（1/2）

Recently The New York Times reported the following:

(Source: The NYT)

This A.I. Just Created Viruses Not Found in Nature (1/2)

Scientists trained artificial intelligence on libraries of DNA and then asked the model to create recipes for viral genomes. Sixteen of them were viable, yielding new viruses.

The NYT - By Carl Zimmer - Carl Zimmer covers news about science for The Times and writes the Origins column.

Aug. 6, 2026

For the first time, scientists have used artificial intelligence to create new kinds of viruses, raising hopes for medical advances while also raising the disturbing possibility that the technology could someday be used to invent dangerous pathogens.

Synthesizing viruses from scratch is hardly new. Researchers long ago learned how to manufacture viral genomes; they are used to investigate antiviral drugs and vaccines, as well as to learn how viruses work.

But the new study, published Thursday in the journal Science, goes well beyond duplicating viral genes. Scientists at Stanford University and the Arc Institute, a research organization in Palo Alto, Calif., taught A.I. to recognize patterns of DNA structure in nature, and then to use that data to write recipes for entirely new viruses.

The researchers followed those recipes to create DNA molecules, which they inserted into bacteria. The modified bacteria then produced viruses never seen in nature. The viruses were able to infect other bacteria, demonstrating that they were viable.

“This is an important milestone,” said Patrick Cai, a synthetic biologist at the University of Manchester, who was not involved in the study.

The viruses dreamed up by A.I. do not pose a threat to humans, because they are all similar to a naturally occurring virus called Phi X-174, which can infect only bacteria.

But the new study adds to growing worries that artificial intelligence might enable the creation of a new generation of biological weapons, from deadly poisons to unstoppable pandemics.

Dr. Moritz Hanke, a fellow at the Johns Hopkins Center for Health Security who was not involved in the new study, said governments and scientific organizations have been slow to develop guardrails that could block the creation of a deadly virus — even as the science races ahead.

“There’s just a huge disconnect,” he said.

The authors of the study relied on an A.I. model called Evo, which is similar in some ways to ChatGPT, made by OpenAI. 

ChatGPT answers questions by stringing together words that seem likely to follow one another. It can do this thanks to years of training on vast amounts of text gathered from the internet, books and other sources.

DNA is strikingly similar to a book in some ways: a string of molecular building blocks, known as nucleotides, arrayed like letters in a line of text. A gene consists of hundreds of nucleotides drawn from a four-letter alphabet: A, C, G and T. The sequence encodes the instructions for building proteins and other molecules.

DNA has its own rules of grammar, and if a sequence violates them, the result is biological gibberish. Biologists have uncovered some of nature’s grammatical rules, but many remain a mystery.

The researchers wondered if Evo could pick up these rules on its own. Instead of training it on text, they trained it on genetic sequences drawn from millions of animals, plants, microbes and viruses. All told, Evo scanned about nine trillion nucleotides.

Evo eventually recognized patterns common across the tree of life and used them to generate blueprints for new genes encoding proteins that could perform specific jobs. These results led the team to wonder if Evo could master not just single genes but also an entire genome.

As the A.I. would be able to handle only small genomes at first, the scientists decided to try to make viruses. While a human genome contains over three billion nucleotides, many viruses have genomes just a few thousand nucleotides long.

“It just felt like the obvious next step,” said Samuel King, a graduate student at Stanford University and an author of the new study. He and his colleagues gave Evo another round of training, this time on the 11 genes of Phi X-174 and about 15,000 of its closest relatives.

(to be continued)

Translation

人工智能剛剛創造出自然界中不存在的病毒（1/2）

科學家利用DNA庫訓練人工智能，然後讓模型創建病毒基因組的合成方案。其中16個方案成功，產生了新的病毒

科學家首次利用人工智能創造了新型病毒，為醫學進步帶來了希望，同時也引發了一個令人不安的可能性：這項技術未來可能被用於製造危險的病原體。

從零開始合成病毒並非新鮮事。研究人員很久以前就掌握瞭如何製造病毒基因組；他們利用病毒基因組來研究抗病毒藥物和疫苗，以及了解病毒的運作機制。

但這項週四發表在《科學》雜誌上的新研究，遠遠超出了複製病毒基因的範疇。史丹佛大學和位於加州Palo Alto的Arc研究所的科學家訓練人工智能辨識自然界中DNA結構的模式，並利用這些資料編寫全新病毒的「配方」。

研究人員依照這些「配方」製造DNA分子，並將其插入細菌體內。改造後的細菌隨後產生了自然界中從未出現過的病毒。這些病毒能夠感染其他細菌，證明它們是可行的。

並未參與這項研究的曼徹斯特大學的合成生物學家Patrick Cai說道:「這是一個重要的里程碑」。

人工智能製造的這些病毒不會對人類構成威脅，因為它們都與一種名為Phi X-174的天然病毒相似，而Phi X-174只能感染細菌。

但這項新研究加劇了人們日益增長的擔憂：人工智能可能被用於製造新一代生物武器，從致命毒藥到無法控制的流行病。

約翰霍普金斯大學健康安全中心的研究員Moritz Hanke博士並未參與這項新研究。他表示，儘管科學發展日新月異，但各國政府和科研機構在制定出能夠阻止致命病毒產生的防止措施方面卻進展緩慢。

他說:「這之間存在著巨大的脫節」。

研究的作者使用了名為Evo的人工智能模型，它在某些方面與OpenAI開發的ChatGPT類似。

ChatGPT透過將看似可能連續出現的字詞串連起來去回答問題。它之所以能夠做到這一點，是因為它經過多年對從互聯網、書籍和其他來源收集的大量文本進行訓練。

DNA在某些方面與書籍驚人地相似：它們都是由被稱為核苷酸 (nucleotides) 的分子構件組成的，排列方式如同文本中的字母。基因由數百個核苷酸組成，這些核苷酸取自字母表：A、C、G 和 T。此序列編碼構成對蛋白質和其他分子的指令。

DNA 有其自身的語法規則，如果序列違反這些規則，就會產生生物亂碼。生物學家已經揭示了一些自然界的語法規則，但許多規則仍然是個謎。

研究人員想知道 Evo 是否能夠自主學習這些規則。他們沒有用文字訓練 Evo，而是用來自數百萬種動物、植物、微生物和病毒的基因序列來訓練。 Evo 總共掃描了大約 9 兆個核苷酸。

Evo 最終識別出了生命之樹中常見的模式，並利用這些模式產生了編碼能夠執行特定功能的蛋白質的新基因藍圖。這些結果讓研究團隊開始思考，Evo 是否不僅能掌握單一基因，還能掌握整個基因組。

由於人工智能最初只能處理小型基因組，科學家決定嘗試建構病毒。人類基因組包含超過30億個核苷酸，而許多病毒的基因組只有數千個核苷酸長。

史丹佛大學研究生、這項新研究的作者之一 Samuel King 說:「這感覺就像是順理成章的下一步」。他和同事們對Evo進行了新一輪的訓練，這次對象是Phi X-174的11個基因以及大約15,000個與其親緣關係最近的基因。

 （待續）

即將來臨、無法逃避的人工智能洪流（3/3）

Recently The New York Times reported the following:

The Impending, Inescapable Deluge of A.I. (3/3)

“It’s hard to get your mind around the scale”

The NYT - By Adam Satariano, Paul Mozur, Jacqueline Gu and Cade Metz - The reporters, based in London, Taipei, New York and San Francisco, have covered the rise of artificial intelligence for years.

July 29, 2026

(continue from part 2)

The U.S. advantage

As more computing power arrives, American companies are expected to extend their A.I. lead. Amazon, Google, Microsoft, Meta and Oracle are projected to spend about $750 billion this year on data centers, chips and other A.I. infrastructure, up from roughly $400 billion last year, according to Goldman Sachs.

China, the next closest rival in A.I., is working to close the gap. Chinese companies had roughly 1.16 million H100-equivalent chips at the end of 2025, up from roughly 244,000 at the beginning of 2024, though those figures exclude smuggled chips and other offshore computing resources used by Chinese firms, Epoch AI estimated.

China’s National Energy Administration has estimated the country’s electricity use for data centers will reach the equivalent of around 91 gigawatts by 2030, or about 6 percent of total use, up from 19 gigawatts last year.

China has been hamstrung by export controls and other limits initiated by the United States on A.I. chips and other key technology. To address those hurdles, Beijing has made A.I. infrastructure a national priority. In March, the Chinese Communist Party released a five-year economic strategy that mentioned A.I. more than 50 times and called for building an interconnected network of data centers to create “next generation supercomputing.”

Xi Jinping, China’s president, has positioned China as a counterweight to American tech dominance. “A.I. development should not be a solo performance by a single country, but a symphony of international cooperation,” he said in a recent speech in Shanghai.

China’s top tech companies are building new A.I. chips and data centers. This year, Huawei, ByteDance and Alibaba are expected to spend $111 billion on data centers and other A.I. investments, according to Bernstein Research.

Even with these challenges, Chinese start-ups like Moonshot AI and DeepSeek have built powerful A.I. models. Often given away as “open source” software, which others can freely use and modify, they are growing more popular. Many users regard the Chinese models as good enough, more efficient and cheaper than leading U.S. models.

Still, everyday use of A.I. is limited across China right now because of lack of computing infrastructure, said Jordan Nanos, an analyst at SemiAnalysis.

Mr. Nanos said the U.S. data center lead over China would likely grow over the next four to five years, before China’s domestic chips are produced at scale. After that, China should begin closing the gap.

“The advantage will run out,” he said.

The U.S.-China race threatens to leave the rest of the world behind. France, Germany and other nations are trying to encourage data center construction across the European Union, which has 5 percent of global A.I. computing power, according to a report by A.I. developers and policy experts in the region. Europe has been hampered by electricity and land access, permitting and financing.

In the Persian Gulf, where Saudi Arabia and the United Arab Emirates have pledged billions to build data centers, the war in Iran has affected plans.

Fears that A.I.’s economic gains are unequally spread are growing, with the world potentially splitting between those with the infrastructure to utilize the technology and those without.

“If 75 percent of the compute today is in a few postal codes in the U.S., 12 to 15 percent in China, and 5 percent in the E.U., where does it leave the rest of the world?” said Amandeep Gill, under secretary general at the United Nations who is the special envoy on tech issues.

The accelerating A.I. loop

To those in the A.I. industry, adding huge data centers is akin to outfitting a car with a jet engine.

Signs of that acceleration are already here. Uber, PepsiCo and Walmart are increasingly turning work over to A.I. “agents,” the bots that can perform a growing list of multistep tasks like coding, compiling research reports, handling customer support and reading and responding to emails.

With more computing power, agents can take on more responsibilities, said Google’s Dr. Dean, who has worked in A.I. research for more than 30 years. He envisioned a scientist asking hundreds of A.I. agents to autonomously devise and test various hypotheses in biological research and then taking the best leads to build off.

Under at least one new assessment, the Remote Labor Index, A.I. models have become increasingly capable. The test examines their ability to do common freelance tasks, like building a mobile video game. In October, leading models completed 2.5 percent of tasks. By July, Anthropic’s Fable A.I. model completed 16 percent.

These jumps in abilities have economists warning about major changes to the labor market.

“There’s going to be millions of jobs destroyed, millions of jobs created,” said Erik Brynjolfsson, an economist who is the director of Stanford University’s Digital Economy Lab. “That’s going to be very difficult. Even if new jobs are created, they’re not the same jobs.”

Leading A.I. labs are continuing to race ahead. One long-sought breakthrough, called recursive self improvement, would allow A.I. to speed its own progress with little or no help from human developers. An A.I. model would essentially help build the next version of itself.

Google is already exploring various kinds of self-improvement tools. A process that once involved dozens of A.I. researchers testing hundreds of ideas could eventually be turned over to thousands of “very tiny models,” which come up with ideas on their own, Dr. Dean said.

With more computing power coming, “you can fully automate the loop,” he said. “We are at the beginning stages.”

Translation

即將來臨、無法逃避的人工智能洪流（3/3）

“其規模之大令人難以置信”

（上接第二部分）

美國的優勢

隨著運算能力的不斷提升，美國公司預計將進一步擴大在人工智能領域的領先優勢。根據高盛預測，亞馬遜、Google、微軟、Meta和甲骨文今年將在數據庫、晶片和其他人工智能基礎設施投入約7,500億美元，高於去年的約4,000億美元。

中國作為人工智能領域最接近的競爭對手，正努力縮小差距。根據Epoch AI估計，到2025年底，中國企業擁有的H100晶片數量約為116萬顆，高於2024年初的約24. 4萬顆。不過，這些數字不包括走私晶片以及中國企業使用的其他境外運算資源。

中國國家能源局估計，2030年，中國數據庫的用電量將達到約91吉瓦，約佔總用電量的6%，高於去年的19吉瓦。

美國對人工智能晶片和其他關鍵技術實施的出口管制和其他限制措施一直束縛著中國的發展。為了克服這些障礙，北京已將人工智能基礎設施建設列為國家優先事項。今年3月，中國共產黨發佈了一項五年經濟戰略，其中50多次提及人工智能，並呼籲建立互聯互通的數據庫網絡，以打造「下一代超級運算」。

中國國家主席習近平將中國定位為制衡美國科技主導地位的力量。他在最近於上海發表的演講中表示：“人工智能發展不應是任何一個國家的獨腳戲，而應是國際合作的交響樂。”

中國頂尖科技公司正在研發新型人工智能晶片和數據庫。根據Bernstein Research預測，華為、字節跳動和阿里巴巴今年預計將在數據庫和其他人工智能投資方面投入1,110億美元。

中國即便面臨這些挑戰，月之暗面（Moonshot AI）和DeepSeek這樣的中國初創公司仍然開發了強大的人工智能模型。這些模型通常以「開源」軟體的形式發佈，允許他人免費使用和修改，因此越來越受歡迎。許多用戶認為，中國開發的人工智能模型性能足夠好，比美國領先的模型更有效率、更便宜。

不過，SemiAnalysis分析師Jordan Nanos表示，由於運算基礎設施不足，目前人工智能在中國的日常應用仍然有限。

Nanos先生表示，在未來四到五年內，美國在數據庫領域的領先優勢可能會繼續擴大，直到中國國產晶片實現規模化生產。此後，中國應該會開始縮小差距。

他說「這種優勢終將消失」。

中美競賽有可能將世界其他國家遠遠拋離。法國、德國和其他國家正努力鼓勵在歐盟範圍內建造數據庫。根據該地區人工智能開發人員和政策專家的報告，歐盟擁有全球5%的人工智能運算能力。然而，電力和土地取得、許可審批和融資等方面的障礙阻礙了歐洲的數據庫建設。

在波斯灣地區，沙地阿拉伯和阿聯酋已承諾投入數十億美元建造數據庫，但伊朗戰爭影響了相關計劃。

人們越來越擔心人工智能帶來的經濟效益分配不均，世界可能會分裂為擁有利用這項技術的基礎設施的國家和缺乏基礎設施的國家。

聯合國副秘書長兼技術問題特使Amandeep Gill說道:「如果今天75%的運算能力集中在美國少數幾個郵遞區號區域，12%到15%在中國，5%在歐盟，那麼世界其他地區又該何去何從呢？」

人工智能加速發展的循環

對人工智能產業人士來說，增建大型數據庫就好比給汽車裝上噴射引擎。

這種加速發展的跡像已經顯現。優步、百事可樂和沃爾瑪等公司正越來越多地將工作交給人工智能“代理”，這些機器人可以執行越來越多的多步驟任務，例如編寫程式碼、撰寫研究報告、處理客戶支援以及閱讀和回覆電子郵件。

谷歌的Dean博士在人工智能研究領域工作了30多年，他表示，隨著運算能力的提升，代理可以承擔更多責任。他設想，一位科學家可以向數百個人工智能發出請求。人工智能體能夠自主地設計並測試生物學研究中的各種假設，然後選擇最佳線索並以此為基礎進行拓展。

至少在一項名為「遠距勞動指數」的新評估中，人工智能模型的能力日益增強。該測試考察了它們完成常見自由職業任務的能力，例如開發手機遊戲。 10月份，領先的模型完成了2.5%的任務。而到了7月份，Anthropic公司的Fable人工智能模式已經完成了16%的任務。

這些能力的飛躍引發了經濟學家擔憂勞動市場即將發生的重大變化。

史丹佛大學數位經濟實驗室主任、經濟學家Erik Brynjolfsson表示:「數百萬個工作崗位將會消失，同時也會有數百萬個工作機會被創造出來」;「情况會非常困難。即使創造了新的工作崗位，它們也與之前的崗位截然不同」。

領先的人工智能實驗室正持續高速發展。一項名為「遞歸自我改善」(recursive self-improvement) 的突破性技術，有望讓人工智能在幾乎無需人類開發者幫助的情況下加速自身發展。人工智能模型本質上能夠幫助建構自身的下一個版本。

谷歌已經在探索各種自我改進工具。Dean博士表示，過去需要數十名人工智能研究人員測試數百個想法的過程，最終可能會交給數千個「微型模型」來完成，這些模型能夠自主想出方法。

隨著運算能力的提升，他說:「你可以完全自動化這個循環」; 「我們目前還處於起步階段」。

              So, the performance of artificial intelligence keeps getting more and more eye-catching.  Last week, two A.I. systems under testing went rogue and hacked into a company’s database. In American hundreds of major data centers are now under construction and will be operational in the coming years. They are set to develop A.I. that has no equal in the history of the technology industry. The U.S.-China AI race is leaving the rest of the world behind. Apparently, millions of jobs will be destroyed and millions of jobs will be created eventually.

Note:

1. Moonshot AI (月之暗面) is one of China's top Artificial General Intelligence (AGI) startups. The Chinese name was inspired by the founder's love for Pink Floyd's famous album, while the English name "Moonshot" signifies their dedication to high-difficulty technological exploration. (Google search)

2. Recursive self-improvement (遞歸自我改善) - In the context of artificial intelligence and machine learning, it refers to the ability of a system to improve its own algorithms and performance through iterative cycles of enhancement, potentially leading to rapid advancements beyond human capabilities. (Google search)

OpenAI人工智能失控入侵Hugging Face伺服器 觸發全球Skynet Day安全警號

Recently Sinic Analytica reported the following:

OpenAI人工智能失控入侵Hugging Face伺服器 觸發全球Skynet Day安全警號

Sinic Analytica (sinicanalytica@creator.patreon.com)

27 Jul 2026

OpenAI人工智能今年7月9日起試圖擺脫公司內部隔離測試環境，兩日後闖入人工智能工具庫Hugging Face伺服器，事件延續至7月13日，其間更曾竊取存取憑證。OpenAI遲至逾一星期後始查悉肇事者為自身人工智能，並於7月21日公開披露事件始末，觸發全球以「Skynet Day」形容此宗人工智能失控事件，引發新一輪安全警號。

 ｜人工智能闖網始末曝光

 美國人工智能公司OpenAI當時正測試一款人工智能，其驅動模型為GPT-5.6 Sol及另一款未發布型號，藉此評估網絡攻防能力。兩名知情人士透露，該人工智能約於7月9日首次嘗試逃出OpenAI內部隔離測試環境。Hugging Face共同創辦人沃爾夫（Thomas Wolf）表示，入侵行動於兩日後、即7月11日展開，並持續至7月13日。三名知情人士其後補充，早於事件發生前，OpenAI技術已現異常跡象，其中一個人工智能曾為未來版本留下筆記，載明如何擺脫公司內部限制。

 ｜遲逾一周始悉真相

 兩名知情人士指，直至7月16日Hugging Face發表網誌，確認「自主人工智能系統」曾入侵其系統，OpenAI才驚悉肇事者為自身人工智能，即由發現異常行為至確認責任歸屬，中間相隔至少一星期。7月18日至19日周末期間，OpenAI員工翻查內部日誌，發現人工智能確曾逃出測試限制。雙方直至7月20日前後才首次就事件互相聯繫，翌日OpenAI對外公開披露事件。事件曝光前，Hugging Face已報警求助，惟路透社未能確認聯邦調查局（FBI）是否已就此展開調查。

 ｜各方回應惹外界質疑

 OpenAI發表聲明形容事件屬史無前例，並指「標誌著人工智能安全的重要時刻」，又稱正與外部顧問檢視事件，其後將發表技術報告。該公司發言人另指路透社報道存在「多處不準確」，惟未有具體說明錯誤所在。聯邦調查局回應指，不評論今次事件。世界倫理數據基金會首席情報專家史密斯（Marley Smith）質疑，事件反映OpenAI或無人看管人工智能，或發現後亦無法將其制止，兩者同樣令人憂慮。

 ｜Skynet Day掀起全球安全反思

 事件曝光翌日，網上迅速以「Skynet Day」形容7月22日這場人工智能失控風波，呼應電影《未來戰士》中失控電腦網絡「天網」的情節。Anthropic前沿紅隊主管格雷厄姆（Logan Graham）在社交平台X發文憶述，事發翌日與團隊一同閱讀報告時，已叮囑同事「記住這一刻」，形容為首宗真正的人工智能安全事故。研究機構帕利塞德研究（Palisade Research）創辦人拉迪什（Jeffrey Ladish）直言，「這些模型會說謊、會欺騙、會入侵」，並促請各國政府加強監管，否則業界難以自律。業界普遍視自主人工智能為未來大勢，惟自主程度愈高，模型為完成任務或通過測試而走捷徑的風險亦隨之上升，令外界關注科企能否在競逐推出最強模型之餘，兼顧安全監管。事件發生之際，OpenAI正籌備可能於今年進行的首次公開招股，為未來擴展業務籌措資金，令是次失控事故格外惹人關注。

(Source: sinicanalytica@creator.patreon.com)

Comment:

文章中說到 ”研究機構帕利塞德研究（Palisade Research）創辦人拉迪什（Jeffrey Ladish）直言，「這些模型會說謊、會欺騙、會入侵」，並促請各國政府加強監管，否則業界難以自律” 。如果這些模型真是會說謊、會欺騙、會入侵， 那是令人震驚人的。

這種新酶能讓人類身體時光倒流嗎？

Recently The New York Times reported the following:

Can This New Enzyme Turn Back the Clock in the Human Body?

In a recent study, scientists devised a way of reversing the buildup of compounds that lead to some age-related diseases.

The NYT - By K. R. Callaway - K. R. Callaway is a science reporter and a member of the 2026-27 Times Fellowship, a program for journalists early in their careers.

July 24, 2026

The same chemical process that turns a baking cake golden brown is at work aging your body’s cells. Over decades, your tissues, baking steadily at 98.6 degrees, slowly accumulate compounds called advanced glycation end products, or AGEs, as sugars interact with protein and fat in the bloodstream. The buildup of AGEs is a hallmark of the aging process, making tissues like collagen sticky and rigid and causing inflammation that can lead to heart disease, eye damage, kidney disease and diabetes.

Researchers have tried unsuccessfully to develop medications that can stop AGEs from forming. But a new study, published on July 16 in the journal Nature Communications, describes a different tack: Scientists created an enzyme that cleans one of the most common AGEs from human tissue, helping to turn back the clock even after the aging compounds have settled in the body. This cleanse can help repair the tissue and has the potential to allow cells to act younger.

“It’s a small step in this bigger direction, because aging is very complicated,” said Aaron Cravens, a founder and the chief executive of Revel Pharmaceuticals, which developed the new enzyme, and the lead author of the study. “It’s the first time, I think, that anyone in the field has shown at the structural level that you can actually reverse some of these changes.”

Many efforts to develop treatments for aspects of aging have focused on cells, which gradually lose the ability to repair themselves and divide properly. Less attention has been given to the aging of longer-lived tissue structures such as collagen, which has a half-life of about 15 years.

Scientists have been aware of advanced glycation end products and their role in aging for decades. “It accumulates on tissue proteins like rust,” said John Baynes, a retired biochemist who was not involved in the study but who spent his career researching AGEs at the University of South Carolina Floyd School of Medicine.

But researchers have had “absolutely no success” trying to create drugs to inhibit the chemical reactions that produce AGEs, he said.

The new finding was “sort of a tour de force of modern biochemical and molecular biological techniques,” Dr. Baynes added. “To demonstrate that the darn thing worked on real proteins isolated from real tissue is really excellent work.”

The study by Dr. Cravens and his colleagues began with a hypothesis: Since all human tissue can be recycled into the biosphere after death, there must be a natural way to dissolve these compounds, even though they are extremely durable. To identify that process, they looked to the enzymes inside microbes, the agents of decomposition.

The team used artificial intelligence software to analyze the DNA sequences of more than 50,000 microbes and to compute the enzyme structures that each genetic code would create. They then narrowed the results to several candidates that seemed capable of cleaving AGEs from human tissue.

They eventually found one, and then engineered it to be more efficient at its task. After several cycles of guided evolution, the novel enzyme, called CMLase, became very good at removing AGEs from human tissue samples.

“In the most extreme case, we took 70-year-old human skin and brought the levels back to that of a 30-year-old,” Dr. Cravens said.

The researchers plan to start testing their enzyme for the treatment of eye diseases, including several associated with diabetes, that develop when AGEs accumulate in retinal and lens tissues. In theory, Dr. Cravens said, CMLase could be applied periodically or even just once to clear out the inflammatory buildup underlying these conditions.

“I was pretty excited about this paper,” said Michael C. Jewett, a bioengineer at Stanford University who was not involved in the study. “It’s not a clinical solution yet, but this proof of concept opens the aperture to possible new therapeutic solutions for longevity.”

If the approach works in a clinical setting, it has the potential to be applied to many AGE-related conditions, for example by improving the elasticity of collagen in skin or by restoring kidney or cardiovascular function.

“This study was a master class in how to engineer an enzyme,” Dr. Jewett said. “I think there is tremendous, untapped potential in this protein class for addressing numerous human diseases.”

Translation

這種新酶能讓人類身體時光倒流嗎？

在最近的一項研究中，科學家找到了一種方法，可以逆轉導致某些與年齡相關的疾病的化合物的累積。

使蛋糕變成金黃色的化學過程，同樣也在加速人體細胞的老化。幾十年來，人體組織在攝華 98.6  度的高溫下持續烘焙，隨著血液中的糖分與蛋白質和脂肪相互作用，會緩慢累積一種名為晚期糖化終產物（AGEs）的化合物。 AGEs 的累積是老化過程的標誌，它會使膠原蛋白等組織變得黏稠僵硬，並引發炎症，進而導致心臟病、眼部損傷、腎臟疾病和糖尿病。

研究人員曾多次嘗試開發能夠阻止晚期糖化終產物（AGEs）形成的藥物，但都以失敗告終。然而，7月16日發表在《自然通訊》雜誌上的一項新研究提出了一種不同的方法：科學家們研發出一種酶，能夠清除人體組織中最常見的AGEs之一，即使這些衰老化合物已在體內沉積，也能出現衰老進程逆轉。這種清除作用有助於修復組織，並有可能使細胞恢復年輕狀態。

研發出這種新型酶的Revel Pharmaceuticals公司創始人兼首席執行官、亦是該研究的主要作者Aaron Cravens表示，“這是朝著更大方向邁出的一小步，因為衰老是一個非常複雜的過程” ;“我認為，這是該領域首次有人在結構層面證明，我們確實可以逆轉其中一些變化。”

許多針對老化相關問題的治療方法都集中在細胞層面，因為細胞會逐漸喪失自我修復和正常分裂的能力。人們對膠原蛋白等壽命較長的組織結構的衰老關注較少，膠原蛋白的半衰期約為15年。

幾十年來，科學家一直關注著晚期糖化終產物（AGEs）及其在老化過程中的作用。John Baynes說:「它像鐵銹一樣在組織蛋白上積聚」。John Baynes是一位退休的生物化學家，他並未參與這項研究，但他在南卡羅來納大學佛洛伊德醫學院畢生致力於AGEs的研究。

但他表示，研究人員在嘗試研發抑制AGEs生成化學反應的藥物方面「完全沒有成功」。

Baynes博士補充說，這項新發現「堪稱現代生物化學和分子生物學技術的傑作」; 「這玩意能夠證明對從真實組織中分離出來的真實蛋白質是有效，這真是一項了不起的工作」。

Cravens博士及其同事的研究始於一個假設：既然所有人體組織在死後都可以循環利用到生物圈中，那麼即使這些化合物極其頑固，也必然存在一種自然的溶解方式。為了探明這個過程，研究人員將目光投向了微生物體內的酵素，它是一種分解劑。

研究團隊利用人工智能軟件分析了超過5萬種微生物的DNA序列，並計算出每種基因編碼所對應的酵素結構。隨後，他們將結果篩選至幾個似乎能夠從人體組織中裂解晚期糖化終產物（AGEs）的候選酵素。

最終，他們找到了一種酶，並對其進行改造，使其更有效率地完成這項任務。經過幾輪引導演化，這種名為CMLase的新型酵素能夠非常有效地從人體組織樣本中去除AGEs。

Cravens博士說: 「在最極端的情況下，我們提取了70歲老人的皮膚樣本，並將其AGEs含量恢復到了30歲老人的水平」。

研究人員計劃開始測試這種酵素在治療眼部疾病方面的應用，包括一些與糖尿病相關的眼部疾病，這些疾病的發生是由於AGEs在視網膜和晶狀體組織中累積所致。Cravens博士表示，理論上，CMLase可以定期使用，甚至只需一次，就能清除這些疾病背後的發炎累積。

並未參與這項研究的史丹佛大學生物工程師Michael C. Jewett說道: 「我對這篇論文感到非常興奮」; 「目前它還不是一個臨床解決方案，但這項概念驗證為延長壽命治療途徑帯來新曙光」。

如果這種方法在臨床環境中有效，它就有可能應用於許多與晚期糖化末期（AGE）相關的疾病，例如透過改善皮膚膠原蛋白的彈性，或恢復腎臟或心血管功能。

Jewett博士說: 「這項研究堪稱酵素工程的典範」; 「我認為這種蛋白質類別在治療多種人類疾病方面蘊藏著巨大的、尚未開發的潛力」。

So, in a new study scientists find an enzyme that can clean up one of the most common AGEs from human tissue, helping us to turn back the clock even after the aging compounds have settled in the body. This cleanse can help repair the tissue and has the potential to allow cells to act younger. Apparently, it’s a small step in this bigger direction, and there is tremendous, untapped potential in this protein class for addressing numerous human diseases. It seems that the novel enzyme, called CMLase is quite effective and I think this is good news to everyone.

Note:

1. Glycation (non-enzymatic glycosylation) (糖化) is the covalent attachment of a sugar to a protein, lipid or nucleic acid molecule. Typical sugars that participate in glycation are glucose, fructose, galactose, and their derivatives. Glycation is the non-enzymatic process responsible for many (e.g. micro and macrovascular) complications in diabetes mellitus and is implicated in other diseases and in aging.

最佳基因風險評估工具並非對每個人都同樣有效（2/2）

Recently The New York Times reported the following:

The Best Genetic Risk Tools Don’t Work Equally for Everyone (2/2)

Genetic prediction models are poised to revolutionize medicine. But because they have been trained on European DNA, they threaten to widen health care disparities.

The NYT - By Emily Baumgaertner Nunn - Emily Baumgaertner Nunn is a national health reporter for The Times, focusing on public health issues that primarily affect vulnerable communities.

July 20, 2026, 5:01 a.m. ET

(continue)

Major demographic events in history, like mass migrations, explain why genetic findings from people in Africa or of African descent are vital to study. Communities there have accumulated great genetic diversity over hundreds of thousands of years. By contrast, present-day Europeans descend from a small group of Africans who expanded out of Africa about 50,000 years ago. As a result, they have much less diversity.

“If you’re going to pick a population to study for the sake of everybody’s good, the Europeans are the worst you could choose,” said Jay Kaufman, an epidemiologist at McGill University who has written about race and genomics.

The most obvious solution is to recruit more people of color, something the All of Us program says it has prioritized. The group works with local partner organizations like churches and community health centers to host discussion sessions and enroll rural communities through a mobile clinic. It also returns individualized health findings and genetic risks to the participants to establish a sense of reciprocity.

Still, recruitment can be an upstream endeavor, particularly in the United States, which has a history of mistreatment of people of color, from the Tuskegee Syphilis Study to discriminatory sickle-cell screening to the story of Henrietta Lacks, a Black patient whose cancer cells were used for research without her knowledge or consent.

“Why would you contribute if you’ve been wronged in that way in the past?” Dr. Martin said. “There is some earned mistrust that needs to be addressed — to have that level of trust to be willing to say, ‘Take my data, monitor it for decades, do whatever you want with it.’”

In the meantime, researchers designing polygenic risk scores have found technical ways to diversify data. Modeling tools can now pull weighted information from a variety of sources simultaneously, rather than drawing only from the largest European cohorts and then forcing their algorithms onto other groups. Mount Sinai Health System in New York and U.C.L.A. in Los Angeles, for example, have both established more diverse biobanks than the UK Biobank, with each of them containing tens of thousands of genomes.

Scientists also draw from rapidly growing biobanks in China, Japan, South Korea and Taiwan, which bolster the accuracy of polygenic risk scores for East Asian groups. Localized initiatives are underway in Peru, Mexico, Qatar and elsewhere. And a continentwide program called H3Africa is collecting genetic data in African populations and supporting efforts by scientists there to study how genes and environments cause diseases.

Academic networks have also tapped sources beyond biobanks, integrating data from large, disease-specific cohort projects, including the Atherosclerosis Risk in Communities Study, or ARIC, which intentionally enrolled thousands of Black participants from North Carolina and Mississippi.

For breast cancer, a project called Confluence aggregates data from over 300 individual studies across 62 countries, providing information from 400,000 breast cancer cases and more than 1.5 million controls. The influx of data from minority groups is being used to sharpen breast cancer polygenic risk scores for everyone.

Private genomics companies and start-ups looking to develop and sell polygenic risk scores do not typically have access to those vast networks. Until recently, they relied almost exclusively on data from the UK Biobank. But the All of Us program recently expanded its policies so that companies can use its more diverse data.

Will all of these efforts, taken together, be enough? It depends on whom you ask. Some experts argue that even an ethnically representative data pool is still imperfect, since variations in other factors — socioeconomic status, health access, age or even sex — can still cause a score’s accuracy to decay. Others believe that the statistical issue was overblown in the first place.

Dr. Kenny, who has been helping to test existing polygenic risk scores for 11 common conditions in racially and ethnically diverse patients, argues that the bias is too complex to resolve quickly.

“I don’t think these things are perfectly portable yet — and maybe never will be perfectly portable,” she said. “But there are things that are starting to narrow that gap.”

Translation

最佳基因風險評估工具並非對每個人都同樣有效（2/2）

基因預測模型有望徹底改變醫學。但由於這些模型是基於歐洲人的DNA訓練的，它們有可能加劇醫療保健方面的差異

（繼續）

歷史上重大的人口發展里程，例如大規模遷徙，解釋了為什麼研究非洲人，或非洲裔人群的基因發現至關重要。這些人群在數十萬年的時間裡累積了豐富的基因多樣性。相較之下，現代歐洲人的祖先是大約5萬年前從非洲遷徙出來的一小群非洲人。因此，他們的基因多樣性要低得多。

曾撰寫過關於種族和基因組學相關文章的麥吉爾大學流行病學家Jay Kaufman說道:「如果你要選擇一個群體進行研究，以造福所有人，那麼歐洲人絕對是最糟糕的選擇」。

最顯而易見的解決方案是招募更多有色人種，這也是「我們所有人」（All of Us）計劃表示優先考慮的事項。該計劃與當地夥伴機構如教會和社區健康中心合作，舉辦討論會，並透過流動診所招募農村社區居民參與。他們也會向參與者提供個人化的健康檢查結果和遺傳風險評估，以建立一種互惠互利的意識。

然而，招募工作可能是一項帶挑戰和困難工作，尤其是在美國。美國歷史上曾多次虐待有色人種，從Tuskegee梅毒實驗到歧視性的鐮狀細胞貧血症篩檢，再到Henrietta Lacks的故事 - 一位黑人患者的癌細胞在未經她知情或同意的情況下被用於研究。

Martin博士問: 「如果你過去曾經遭受過這樣的不公，你又怎麼會願意做出貢獻呢？」; 「目前存在一些根深蒂固的不信任需要加以解決 - 要建立起足夠的信任，才能讓人願意地說： ‘拿取我的數據，監測它幾十年，你想怎麼用都行’」。

同時，設計多基因風險評分的研究人員已經找到了一些技術手段來使數據多樣化。建模工具現在可以同時從各種來源提取加權訊息，而不是僅僅從最大的歐洲群體中提取訊息，然後將其演算法強加給其他群體。例如，紐約的西奈山醫療系統和洛杉磯的加州大學洛杉磯分校都建立了比英國生物銀行更多樣化的生物樣本庫，每個樣本庫都包含數以萬計的基因組。

科學家也利用中國、日本、韓國和台灣地區快速成長的生物樣本庫，這些樣本庫提高了東亞人口多基因風險評分的準確性。秘魯、墨西哥、卡達和其他地區也正在進行類似的本地化計劃。此外，一項名為H3Africa的全非洲大陸計劃正在收集非洲人群的遺傳數據，並支持當地科學家研究基因和環境如何導致疾病。

學術網絡也開始利用生物樣本庫以外的數據來源，去整合大型特異疾病群體研究的數據，例如動脈粥狀硬化風險社群研究（ARIC）。 ARIC 特意招募了數千名來自北卡羅來納州和密西西比州的黑人參與。

在乳癌領域，一個名為 Confluence 的計劃匯聚了來自 62 個國家/地區 300 多項獨立研究的數據，提供了 40 萬個乳癌病例和超過 150 萬個對照的資訊。來自少數族裔群體的大量數據正被用於改善適用於所有人的乳癌多基因風險評分。

希望開發和銷售多基因風險評分的私人基因組學公司和初創公司通常無法存取這些龐大的數據網路。直到最近，他們幾乎完全依賴英國生物樣本庫的數據。但「我們所有人」（All of Us）計劃最近擴大了其政策範圍，允許其他公司使用其擁有的更多樣化的數據。

所有這些努力加起來是否足夠？這取決於你問誰。一些專家認為，即使是具有種族代表性的資料池也並非完美無缺，因為其他因，素例如社會經濟地位、獲得醫療服務能力、年齡甚至性別的差異仍然會導致評分準確性下降。另一些專家則認為，統計學上的問題一開始就被誇大了。

Kenny博士一直在幫助測試針對11種常見疾病的現有多基因風險評分，研究對象涵蓋不同種族和民族的患者。她認為，偏差過於複雜，難以迅速解決。

她說:“ 目前我不認為這些方法能完全通用 - 或許永遠也無法完全通用”; “但有些方法正在縮小那差距。”

               So, scientists are already developing clinical tools and techniques based on genomic information gathered from the agency’s All of Us program. One of the most promising may be the Polygenic risk scores. These can help predict a person’s likelihood of developing complex diseases. Yet the tools were trained overwhelmingly on the DNA of people of European descent, so they often fail to accurately assess risks for other descent. Apparently, we may never be able to completely close the gap that causes health care disparities, but some new approaches are starting to narrow the gap.

最佳基因風險評估工具並非對每個人都同樣有效（1/2）

Recently The New York Times reported the following:

The Best Genetic Risk Tools Don’t Work Equally for Everyone (1/2)

Genetic prediction models are poised to revolutionize medicine. But because they have been trained on European DNA, they threaten to widen health care disparities.

The NYT - By Emily Baumgaertner Nunn - Emily Baumgaertner Nunn is a national health reporter for The Times, focusing on public health issues that primarily affect vulnerable communities.

July 20, 2026, 5:01 a.m. ET

The National Institutes of Health recently announced a remarkable achievement: the world’s largest database of human genomes, a repository of information that the agency says will help usher in a new era in personalized medicine.

Scientists are already developing clinical tools and techniques based on genomic information gathered from the agency’s All of Us program. One of the most promising? Polygenic risk scores. These can help predict a person’s likelihood of developing complex diseases, like cardiovascular disease and breast cancer, from as early as infancy.

But there’s a glaring problem: The tools were trained overwhelmingly on the DNA of people of European descent, so they often fail to accurately assess risks for anyone else. For some conditions, the forecasts for people of color are almost no better than flipping a coin.

Researchers are racing to correct the biases through advancements in modeling and widespread recruitment of minority groups. They fear that failing to close the gaps could worsen health care disparities — and prevent the technology from realizing its promise of helping to decrease chronic disease.

“It really warrants attention,” said Eimear Kenny, the director of the Institute for Genomic Health at Mount Sinai and a principal investigator in a national consortium that tests the genetic scores.

Geneticists measure a polygenic risk score by aggregating hundreds or thousands of tiny variations across the genome, many of which might nudge a person’s disease risk by only a fraction. By bundling them together, they can generate a cumulative score, ranking a person’s predisposition to common killers.

For patients at the extremely high end of the risk distribution, the scores can offer a medical advantage, potentially giving doctors a decades-long runway to intervene. Those in the highest percentiles for a cardiovascular disease score, for instance, can be prescribed cholesterol-lowering statins earlier, potentially helping prevent some heart attacks and strokes.

But the entire premise rests on the data used to train the model in the first place. The UK Biobank — established in the early 2000s and widely considered a leading genomic trove — is overwhelmingly homogenous; roughly 94 percent of its participants are white. The Million Veteran Program, a repository run by the U.S. Department of Veterans Affairs, is almost three-quarters white.

The cohort in the All of Us program, built by the N.I.H. with the goal of diversifying data, is still about 50 percent European, according to Alicia Martin, a statistical geneticist at the Broad Institute. The effort’s future is also at risk: One of its major funding streams, the 21st Century Cures Act, is set to expire at the end of this fiscal year. (The program’s budget has already been reduced by 72 percent since 2023.)

“You can’t just snap your fingers and overnight have a biobank from Africa, a biobank from India, a biobank from China that are all as large and open and comprehensive as the UK Biobank,” Dr. Martin said. “We can develop all the fancy statistical methods we want, but still not be able to overcome without more diverse representation.”

Other clinical tools could be affected by the ancestry bias, such as integrated artificial intelligence models that draw upon genomics and other biomedical data. Doctors who sequence a genome to look for rare mutations may also find genetic variants in patients of color that researchers have not studied enough to understand.

Genetic blind spots can also cause scientists to miss discoveries that benefit everyone. One example: the PCSK9 gene. Scientists found that a rare mutation that occurred mostly in a small proportion of Black people could naturally reduce the risk of developing coronary heart disease by 88 percent. That set off a race to develop drugs that mimic the mutation for people of all races and ethnicities. But scientists only found the variant because one study in Dallas had ensured that its participant pool was more than half Black.

(to be continued)

Translation

最佳基因風險評估工具並非對每個人都同樣有效（1/2）

基因預測模型有望徹底改變醫學。但由於這些模型是基於歐洲人的DNA進行訓練的，它們有可能加劇醫療保健方面的差異

美國國立衛生研究院（NIH）最近宣布了一項非凡的成就：建立了世界上最大的人類基因組資料庫。該機構表示，這個資訊庫將有助於開啟個人化醫療的新時代。

科學家已經開始基於NIH的「我們所有人」（All of Us）計劃所收集的基因組資訊來開發臨床工具和技術。其中最有前景的是什麼？多基因風險評分。這些評分可以幫助預測一個人從嬰兒時期開始患上複雜疾病（如心血管疾病和乳癌）的可能性。

但有一個顯而易見的問題：這些工具主要基於歐洲裔族群的DNA進行訓練，因此往往無法準確評估其他族群的風險。對於某些疾病，對有色人種的預測結果幾乎不比擲硬幣好到哪裡去。

研究人員正爭分奪秒地透過改進建模技術和廣泛招募少數族裔群體來糾正這些偏差。他們擔心，如果無法彌合這些差距，可能會加劇醫療保健方面的差異，並阻礙這項技術實現其降低慢性病發病率的承諾。

西奈山基因組健康研究所 所長、同時負責測試基因評分的國家聯盟首席研究員Eimear Kenny表示:「這確實值得關注」。

遺傳學家透過匯集基因組中數百或數千個微小變異來測量多基因風險評分，其中許多變異可能只會對一個人的患病風險產生微乎其微的影響。而透過將這些變異組合在一起，他們可以產生一個累積評分，從而評估一個人罹患常見致命疾病的傾向。

對於處於風險分佈極高端的患者而言，這些評分可以提供醫療優勢，使醫生有長達數十年的干預時間。例如，心血管疾病評分處於最高百份位的患者可以更早服用降膽固醇的他汀類藥物，這可能有助於預防部分心臟病和中風。

但這一切的前提都建立在最初用來訓練模型的資料之上。英國生物銀行（UK Biobank）－成立於21世紀初，被廣泛認為是領先的基因組資料庫－其參與者構成極為單一，約94%為白人。由美國退伍軍人事務部營運的「百萬退伍軍人計劃」（Million Veteran Program）作為一個資料庫其參與者中近四分之三為白人。

據Broad Institute的統計遺傳學家Alicia Martin稱，由美國國立衛生研究院（NIH）建立的「我們所有人」（All of Us）計劃是旨在實現數據多樣化，但該計劃的參與者中仍有約50%為歐洲人。而這項工作的未來也面臨風險：其主要資金來源之一 - 《21世紀治癒法案》 - 將於本財政年度結束時到期。 （自2023年以來，該項目的預算已被削減了72%。）

Martin博士說: 「 你不能打個響指就可以可以在一夜間建立一個像英國生物銀行一樣規模龐大、全面開放、覆蓋面廣闊的非洲生物銀行、印度生物銀行、中國生物銀行」; 「我們可以發展出各種炫麗的統計方法，但如果沒有更多樣化的代表性，我們仍然無法克服這個問題」。

其他臨床工具也可能受到種族偏差的影響，例如利用基因組學和其他生物醫學數據的整合人工智能模型。醫生在進行基因組定序以尋找罕見突變時，也可能在有色人種患者身上發現一些研究人員尚未充分研究的基因變異。

基因盲點也可能導致科學家錯過一些對所有人都有益處的發現。例如：PCSK9基因。科學家發現，一種主要存在於少數黑人中的罕見基因突變，能夠自然地將罹患冠心病的風險降低88%。這引發了一場旨在模仿這種突變研發藥物的競賽，做出適用於所有種族和民族的人藥物。但科學家之所以能夠發現這種變異，是因為達拉斯的一項研究確保了其參與者中的黑人佔超過一半。

 （待續）

遇到棘手的醫學問題？你的醫生可能使用人工智能解决 (1/2)

Recently The New York Times reported the following:

Have a Thorny Medical Question? Your Doctor May Be Using A.I. for That (1/2)

OpenEvidence, a fast-growing start-up, is using artificial intelligence to help doctors find answers to clinical questions for diagnosis and treatment.

The NYT - By Steve Lohr - Steve Lohr has reported on the way technology is changing the work force for more than a decade.

June 8, 2026

Dr. Nicholas Gavin, an emergency medicine doctor at Mount Sinai in New York City, was working an overnight shift last summer when a patient came in with a puzzling set of symptoms. Within seconds, his three younger colleagues — two medical students and a resident — were consulting a free artificial-intelligence-powered app for physicians, OpenEvidence.

Dr. Gavin soon learned that they were far from outliers. A third of Mount Sinai’s 9,000 doctors were already regular OpenEvidence users, the health system’s executives found out in a meeting last year with the start-up’s leaders.

“That was an ‘aha’ moment for our leadership,” said Dr. Gavin, who is also the system’s chief clinical innovation officer.

OpenEvidence’s A.I. app, essentially a chatbot for medicine, has become a viral hit with physicians. Talk to a doctor and chances are he or she uses the app to ask specific medical questions or bounce ideas off it in a diagnostic dialogue.

More than half of the nation’s physicians are regular users. Last month, they used it for 30 million questions and consultations, nearly twice the volume from six months earlier, according to the start-up. A separate survey last year of 1,000 physicians found that 45 percent of them used the app, nearly triple the percentage who used ChatGPT, according to Offcall, a career information service for doctors.

That growth propelled the start-up to a $12 billion valuation in January, up from $3.5 billion last July.

But doctors’ quick adoption of the app since its introduction in 2024 — one of a handful of A.I.-enhanced programs on the market seeking to win over physicians — has heightened concerns about how and when the technology should be used in life-or-death situations. In a high-stakes field like medicine, health care systems are navigating thorny matters of patient privacy, safety and trust, as well as the limitations of the technology itself.

“It’s not an oracle, it’s a tool,” said Daniel Nadler, founder and chief executive of OpenEvidence. “Knowledge and knowledge workers still matter.”

The doctor’s office has been a target for computer-assisted decision making for decades, with very limited success until the recent advances of A.I.

The first wave of A.I. in medicine focused on easing the heavy burden of documentation that contributes to physician burnout with transcriptions and summaries of patient visits, called A.I. scribe software. The second wave, which is just getting underway, aims to use A.I. to assist doctors with reliable information and advice to guide diagnosis and treatment while at a patient’s bedside.

The competition has intensified in recent months. UpToDate, a popular legacy electronic reference for doctors, has given its service an A.I. makeover with a chatbot interface. Doximity, an online professional network for physicians, bought an A.I. start-up that mines medical literature and generates summaries. Abridge, a fast-growing A.I. scribe maker, is adding decision-support tools. And last month, OpenAI introduced ChatGPT for Clinicians.

OpenEvidence became a front-runner in part because it exclusively used medical journals and other high-quality research as data to train its A.I. models. Physicians can ask the app specific questions or enter the characteristics and symptoms of a patient and ask for potential explanations. The app is compliant with the federal law that protects patient health information, and physicians are told not to enter any personally identifying information.

OpenEvidence responds with a summary of most likely diagnoses, and then offers other “most important not to miss diagnoses.” Each has links to the research articles that inform the summaries.

“A.I. is solving some of the problems that have long plagued the practice of medicine,” said Dr. Raja-Elie Abdulnour, chief clinical innovation officer at NEJM Group, which publishes The New England Journal of Medicine. “These tools just didn’t exist before, and that’s why people are so excited about them now.”

Yet the early enthusiasm should be tempered with a large dose of caution, medical experts agree. The research so far into the benefits and shortcomings of A.I. in medicine is decidedly mixed.

A.I. has aced standard licensing exams and outperformed human doctors in diagnosing certain cases. But A.I. has also stumbled, failing to accurately summarize research papers or giving wrong answers to diagnostic questions. And it isn’t going to replace humans anytime soon.

“The potential for A.I. is great, but we’re not there yet,” said Dr. Eric Topol, a cardiologist and an executive vice president at Scripps Research in San Diego. “It hasn’t really been tested and demonstrated in the messy, real world of medicine.”

(to be continued)

Translation

遇到棘手的醫學問題？你的醫生可能使用人工智能解决 (1/2)

OpenEvidence 是一家快速發展的新創公司，它正利用人工智能幫助醫生找到臨床診斷和治療問題的答案

去年夏天，紐約市西奈山醫院的急診醫生Nicholas Gavin值夜班時，一位有令人費解症狀的病人前來就診。幾秒鐘之內，他的三位年輕同事 - 兩名醫學生和一名住院醫師 - 就開始使用一款名為 OpenEvidence 的免費人工智能醫生應用程式去診斷。

Gavin醫生很快就發現，他們絕非異類。西奈山醫院的9,000名醫生中，有三分之一已經是 OpenEvidence 的經用使用者。去年，這醫療機構的管理層在與這家新創公司的領導人會面時得知了這一情況。

同時也是該智能系統的首席臨床創新官的 Gavin醫生說道:「這對我們的領導層來說是一個『頓悟』時刻」。

OpenEvidence 的人工智能應用程序，本質上是一個醫療聊天機器人，已在醫生群體中迅速走紅。随便問一個醫生，他們很可能都使用這款應用程式來諮詢特定的醫學問題，或在診斷對話中與它交流想法。

超過一半的美國醫生都是這款應用程式的經常用戶。據這家新創公司稱，上個月，他們使用該應用程式進行了 3,000 萬次諮詢和問詢，幾乎是六個月前的兩倍。去年一項針對 1,000 名醫生的獨立調查發現，45% 的醫生曾使用過這款應用程序，幾乎是 ChatGPT 用戶比例的三倍。這項調查的數據來自醫生職業資訊服務平台 Offcall。

這一成長推動這家新創公司在今年 1 月的估值達到 120 億美元，而去年 7 月的估值為 35 億美元。

但自2024年推出以來，醫生們對這款應用程式的快速接受度 - 它是市場上為數不多的幾款旨在贏得醫生青睞的人工智能增強型程式之一 - 加劇了人們對這項技術在生死攸關的情況下應該如何以及何時使用的擔憂。在醫療如此高風險的領域，醫療保健系統正在努力應對棘手的病患隱私、安全和信任問題，以及科技本身的限制。

OpenEvidence的創始人兼執行長 Daniel Nadler表示: 「它不是神諭，而是一種工具」; 「知識和知識工作者仍然至關重要」。

幾十年來，醫生辦公室一直是用電腦去輔助診症的目標場所，但直至人工智能最近取得大進展之前，其應用非常有限。

第一波人工智能在醫學領域的應用主要集中在減輕醫生因繁重的文件工作而導致的職業倦怠，其方式是透過轉錄和總結患者就診記錄，即所謂的人工智能記錄軟件。第二波浪潮剛興起，旨在利用人工智能為醫生提供可靠的資訊和建議，以指導他們在患者床邊進行診斷和治療。

近幾個月來，競爭日益激烈。廣受歡迎的傳統電子參考資料平台 UpToDate 為其服務進行了人工智能改造，新增了聊天機器人介面。向醫生提供線上專業網路的 Doximity 收購了一家人工智能新創公司，該公司挖掘醫學文獻並提供摘要。快速發展的人工智能醫療記錄工具 Abridge 正在添加決策支援工具。上個月，OpenAI 推出了臨床醫生的 ChatGPT。

OpenEvidence 之所以能成為領跑者，部分原因在於它完全使用醫學期刊和其他高品質研究作為資料來訓練其人工智能模型。醫生可以向該應用程式提出具體問題，也可以輸入患者的特徵和症狀，並尋求可能的解釋。這款應用程式符合保護患者健康資訊的聯邦法律，醫生被告知不得輸入任何個人識別資訊。

OpenEvidence 會提供最可能診斷的摘要，並提供其他「不容錯過的重要診斷」。每個摘要都附有相關研究文章的連結。

出版《新英格蘭醫學雜誌》的 NEJM Group 的首席臨床創新官 Raja-Elie Abdulnour 醫生表示:「人工智能正在解決一些長期困擾醫學實踐的問題」;「這些工具以前根本不存在，所以人們現在對它們如此興奮」。

然而，醫學專家一致認為，早期的熱情應該伴隨著大量的謹慎。目前關於研究人工智能在醫學領域的益處和不足的結果是明顯地參差。

人工智能已經順利通過了標準的執業資格考試，並在某些病例的診斷上超越了人類醫生。但人工智能也曾失手，例如無法準確總結研究論文，或對診斷問題給予錯誤答案。而且，它在短期內也不會取代人類。

Eric Topol博士是心臟科醫生，也是聖地亞哥Scripps研究所的執行副總裁説: 「人工智能的潛力巨大，但我們尚未達到目標」; 「它還沒有在紛繁複雜的真實醫學世界中經過實際考驗和證實」。

（待續）