Recently The New Times reported the following:
Some Scientists Have ‘Magic Hands’ in the Lab. This A.I.
Is Learning Why. (1/2)
Even the most adept researchers may not know exactly what
they do to get successful results. An A.I. model is trying to figure it out by
watching every move.
The NYT - By Carl Zimmer
Aug. 27, 2026
On a recent afternoon, scientists buzzed around a lab in
Cambridge, Mass., performing experiments. Their equipment was standard: a lab
hood for working with dangerous chemicals, an incubator for growing cells.
Their procedures were identical to those in biology labs everywhere.
A closer look revealed some oddities, though. Along with lab coats and gloves, most of the scientists wore miniature cameras on headbands. Three additional cameras peered down at each work station from a shelf.
Lab notebooks were strangely absent. Instead, the scientists quietly narrated their work, murmuring into microphones. A team huddled at one end of the lab, inspecting videos of the experiments.
This was the real research taking place in this lab. With a system of sensors and software intended to capture science as it happens, down to the millisecond, scientists were training artificial intelligence models to recognize every object the scientists used and every action they performed.
The A.I. even composed its own narrative, describing every few seconds of video with sentences like, “The operator resuspends the pellet by pipetting it up and down 10 times.”
The technology is the creation of a start-up called Transfyr, which came out of stealth mode this week with $25 million in seed funding. The company is tackling an age-old challenge in science: the hidden factors that make some experiments succeed and others fail.
Failure can take many forms. Researchers may spend months preparing a line of engineered cells, only to have them mysteriously die along the way. A government Covid test seems to work in one lab, but fails to detect the virus when others use it.
A biotech company creates a promising new drug; when it hands off the protocol for large-scale production, suddenly it just doesn’t work.
“This is just an incredibly painful problem,” said Anna Marie Wagner, a co-founder of Transfyr. “There’s finger-pointing back and forth. Was your protocol wrong? Or did you screw something up? These are very, very expensive mistakes in terms of time, money, and lives.”
Success can be just as mysterious as failure. Some researchers consistently get experiments to work, earning befuddled admiration from colleagues. Scientists even have a special term for this gift: magic hands.
The idea may come as a surprise to people who don’t spend their lives in labs. Science is not supposed to be magic.
When scientists carry out experiments, they keep careful records, both in lab notebooks and later in published scientific papers. Other researchers use that information to repeat the experiment.
But every scientist discovers sooner or later that essential knowledge is not necessarily written down.
“A protocol is a recipe,” said Jonathan Livny, a senior research scientist at the Broad Institute who has collaborated with Transfyr. “You can give somebody a recipe, and it’s not going to make them a great chef.”
Like apprentice chefs, scientists spend years in training, shadowing experts, asking questions and trying out procedures for themselves.
To make an experiment work, they may have to make thousands of minor decisions over the course of a day. A tube needs to be shaken — let it whir on a vibrating platform, or just flick it back and forth by hand?
“The really good people sometimes don’t know why they’re that good,” Dr. Livny said. “They don’t remember the 20 times they did something another way and it failed. They’ve blocked all that sadness out.”
In the 1960s, the Hungarian chemist and philosopher Michael Polyani gave this mysterious expertise a name: tacit knowledge. “We know more than we can tell,” he liked to say.
Scientists have generally come to agree with Polanyi that tacit knowledge is an essential part of research. But it can also slow down progress.
In theory, science moves forward as researchers build on previous work. When a new study comes out, other scientists will attempt to replicate the work. If the original result was correct, a replication ought to produce the same results.
Surprisingly often, it doesn’t. And tacit knowledge is part of the problem. Scientists can’t know exactly how to replicate a study simply by reading about it.
(to be continued)
Translation
有些科學家在實驗室裡擁有「魔術手」。人工智能正在學習其中的奧秘。 (1/2)
即使是最熟練的研究人員，也可能不完全清楚自己是如何取得成功的。一款人工智能模型正試圖透過觀察他們的每一個動作來找出答案
最近一個下午，在麻薩諸塞州劍橋市的一間實驗室裡，科學家們忙著進行各種實驗。他們的設備都很標準：用於處理危險化學品的通風櫃，以及用於培養細胞的培養箱。他們的實驗流程也與世界各地生物實驗室的流程完全相同。
然而，仔細觀察後卻發現了一些奇怪之處。除了實驗服和手套，大多數科學家頭上都戴著微型攝影機。另外三個攝影機則從架上俯視著每個工作台。
奇怪的是，實驗紀錄簿不見了。取而代之的是，科學家們對著麥克風低聲講述他們的工作。一個小組聚集在實驗室的一端，正在查看實驗的錄影。
這才是這個實驗室真正進行的研究。科學家利用一套旨在精確到毫秒地捕捉科學過程的感測器和軟件系統，訓練人工智能模型來識別他們使用的每個物體和執行的每一個動作。
人工智能甚至能夠產生自己的敘述，用諸如每隔幾秒鐘用「操作員用移液器上下吸取10次，使沉澱物重新懸浮」之類的句子來描述影片。
這項技術出自一家名為 Transfyr 的新創公司，該公司本週結束了隱密地運營，並獲得了2,500萬美元的種子輪融資。該公司正在攻克科學界一個長久以來的難題：導致某些實驗成功而另一些實驗失敗的隱藏因素。
失敗的形式多種多樣。研究人員可能花費數月時間建構出一組細胞，卻發現這些細胞在建構過程中神祕死亡。政府研發的一種新冠病毒檢測方法在一個實驗室似乎有效，但在其他實驗室卻無法檢測到病毒。
一家生技公司研發出一種很有前景的新藥；當它把生產方案交給大規模生產團隊時，突然發現這種藥根本不起作用。
Transfyr 的聯合創始人 Anna Marie Wagner 說: “這真是個令人無比頭疼的問題”; “大家互相指責。是不是你的流程錯了？還是你哪裡搞砸了？這些錯誤在時間、金錢和生命方面都代價極其高昂。”
成功和失敗一樣神秘莫測。有些研究人員總是能成功完成實驗，贏得同事們既困惑又欽佩的讚嘆。科學家甚至給這種天賦取了個專門的詞：魔術之手。
對於那些不常待在實驗室的人來說，這種說法或許會讓他們感到驚訝。科學不應該是魔術。
科學家在進行實驗時，會仔細記錄實驗過程，既會記在實驗筆記裡，也會在之後發表的科學論文中加以闡述。其他研究人員會利用這些資訊來重複實驗。
但每位科學家遲早都會發現，關鍵的知識並非總是會被記錄下來的。
曾與 Transfyr 公司合作的博德研究所 高級研究科學家 Jonathan Livny 說道:「實驗方案就像食譜」; 「你可以給別人一份食譜，但這並不能讓他們成為一位偉大的廚師」。
就像學徒廚師一樣，科學家需要花費數年時間進行培訓，跟隨專家學習，不斷提問，並親自嘗試各種實驗步驟。
為了讓實驗成功，他們可能一天之內要做出成千上萬個細微的決定。例如，當需要搖晃試管 - 是讓它在振動平台上旋轉，還是用手輕輕地搖動它？
Livny 博士說: 「真正優秀的人有時甚至不知道自己為什麼如此優秀」; 「他們不記得自己曾經嘗試過其他方法卻失敗了20次。他們已經把那些失敗感都隔掉了」。
在1960年代，匈牙利化學家兼哲學家 Michael Polyani 將這種神秘的專業知識命名為「融會知識」。他常說: 「我們知道的比我們能表達出來的要多」。
科學家普遍認同 Polanyi 的觀點，即融會知識是研究的重要組成部分。但它也可能阻礙研究進展。
理論上，科學的進步源自於研究者在前人工作的基礎上不斷拓展。當一項新研究發表時，其他科學家會嘗試重複這項研究。如果最初的結果是正確的，那麼重複實驗應該會得出相同的結果。
然而，令人驚訝的是，能得出相同的結果不是經常發生。而融會知識正是問題所在。科學家無法僅憑閱讀就準確地知道如何重複一項研究。
（待續）
