Singularity HUB

一种新方法可能革新亨廷顿病的治疗 || A New Approach Could Transform Huntington’s Disease Treatment

2025-09-30T15:33:15+00:00

在一项小型临床试验中，一种注入大脑的基因疗法在三年内将疾病减缓了75%。

亨廷顿病极其残酷。症状最初表现为手部的随机、不受控制的抽搐。随着时间推移，这种疾病会逐渐侵蚀记忆、思维和理智。情绪波动和人格变化会剥夺你的身份。最终，它会导致早逝。

更糟糕的是，与其他逐渐破坏脑功能的疾病不同，例如阿尔茨海默病，亨廷顿病可以通过简单的基因检测诊断。这种疾病通过突变基因遗传。有家族史的人常常挣扎是否要进行检测。如果结果为阳性，目前没有治疗方法，他们的命运已定。

一种新的疗法或许可以阻止亨廷顿病在症状出现前恶化。初步结果来自一小部分患者，发现将一种微小RNA（一种基因疗法）注入受影响的大脑区域后，三年内疾病进展减缓了75%。与未治疗的对照组相比，这些患者在运动控制、注意力和信息处理速度方面表现显著更好，对照组患者具有相似的基线症状。

这种药物由荷兰基因疗法公司 uniQure 开发，该公司在本月的一份新闻稿中总结了这些发现。这些数据尚未发表在预印本文章或科学期刊上，也未经过其他专家的审查。由于只有29名患者参与，很难将这些益处和安全性概况推广到美国、欧洲和英国的大约 7.5万人。

但这些发现带来了希望的曙光。此前尝试治愈的方法“显示出一些微弱的信号，如果你仔细观察的话，但从未有过如此接近的成功”，加州格拉德斯通研究所的史蒂文·芬克贝纳告诉《纽约时报》。由于亨廷顿病可以在早期被发现，如果这种疗法在更大人群中进一步被证实有效，它可能在更早的年龄开始阻止症状。

基因抛硬币

我们所有人都有亨廷顿基因，或HTT。尽管其在细胞中的确切作用尚存争议，但该基因充当多个细胞“电话线”之间的中央通信器。它协调大量分子，开启或关闭大脑细胞中的基因，并对早期发育、神经元存活和维持大脑整体健康至关重要。

然而，在亨廷顿病中，HTT出现了问题。我们的基因由四种分子组成，用字母A、T、C和G表示。这些字母的三联体通常决定了蛋白质的序列、结构和功能，蛋白质是细胞的主力。在疾病中，一个三联体CAG像坏掉的唱片一样重复，导致突变的亨廷顿蛋白在人的一生中不断积累，并逐渐造成破坏。

尽管在最初阶段大脑细胞可以适应，但它们的防御最终会失效，症状随之出现。在美国，这通常发生在 30至55岁之间。

患有亨廷顿病的家庭面临一个可怕的困境。如果一位父母患病，他们的每个孩子都有50%的几率遗传该病。如果他们不患病，他们的后代则安全。了解诊断结果有助于家庭和生活规划——但这也伴随着沉重的情感代价。

微小但强大

突变的亨廷顿蛋白如何摧毁大脑细胞尚不清楚，但大多数科学家认为清除它或阻止其形成可能有助于保护大脑。

这种蛋白体积庞大，由多个片段组成。一种治疗方法使用小蛋白“阻断剂”来阻止特别有毒的亨廷顿蛋白形成大型、危险的聚集体。另一种方法直接针对CAG重复序列，使用经典但强大的基因疗法。然而，由于高风险的副作用和症状改善的可能性较低，最初的有希望结果后，试验被暂停。基因编辑策略，如 CRISPR，可以剪除突变序列，但它们仍处于非常早期的阶段。

uniQure开发的新疗法利用微小RNA。这些分子不编码蛋白质，但可以阻止基因产生蛋白质。与DNA一样，如果RNA序列匹配，它也可以形成双链。细胞将双链RNA识别为外来物质并将其摧毁——这可能阻止有毒蛋白质的形成。公司的新药物包含两个成分：一种无害的病毒载体和一个定制的基因序列，一旦进入细胞，就会产生专门抑制突变蛋白生产的微小RNA。

这种药物名为 AMT-130，不会整合到或直接编辑患者的基因组，从而降低破坏健康基因或引发癌症的风险。尽管病毒载体最终会被免疫系统清除，但基因序列可能持续多年，使该药物成为潜在的长期治疗方案。

研究团队使用一种已建立且高度精确的外科技术，将AMT-130的低剂量或高剂量注入患有亨廷顿病的志愿者大脑。他们针对纹状体，这是位于大脑深处的一个关键区域，负责运动和决策，也是疾病最早被破坏的区域之一。作为对照组，他们发现数百名年龄和疾病严重程度相似的患者，根据公司的投资者简报（PDF）。

结果令人鼓舞。接受最高剂量的12名早期患者，平均而言，其疾病进展比未接受治疗的人减缓了75%，这是通过多种标准亨廷顿病评估测量的。

大约88%的治疗患者在注意力、记忆和信息处理速度方面显示出显著改善，根据一项测试。他们对随机肌肉运动的控制能力提高，能够以较少的困难完成日常活动。一种脑蛋白通常与症状严重程度相关，其水平降至试验开始前的水平。相比之下，接受低剂量药物治疗的患者有更温和且不一致的结果。

许多患者经历了与脑手术相关的副作用。头痛是最常见的抱怨。一些患者在手术后几天内出现脑肿胀。但总体而言，这种治疗似乎安全。

“大多数与药物相关的严重不良事件发生在治疗后几周内，并可通过类固醇或姑息治疗完全缓解，”公司在其简报中指出。

有理由保持怀疑。亨廷顿病是一种终身疾病，目前尚不清楚单次注射的好处能持续多久，超过三年。很可能需要在整个患者生命周期内多次注射，未来的研究必须测试其累积效果。该药物同时降低了突变和正常版本的亨廷顿蛋白水平——过去药物也如此——这可能产生副作用。

新患者正在被招募参加试验，公司希望在2026年底提交申请以获得FDA批准。

“这个结果彻底改变了局面，”伦敦大学学院亨廷顿病中心试验站点的项目负责人埃德·怀尔德在新闻稿中说。 “基于这些结果，AMT-130似乎将成为首个获批减缓亨廷顿病进展的治疗药物，这确实是真正改变世界的事情。”

文章一种新方法可能改变亨廷顿病治疗首次出现在 SingularityHub。

---------------

In a small trial, a gene therapy injected into the brain slowed the disease by 75 percent over three years.

Huntington’s disease is extremely cruel. Symptoms start with random, uncontrollable twitches of the hand. Over time the disease eats aways at memory, thought, and reason. Mood swings and personality changes strip away your identity. Eventually, it leads to an early death.

Worse, unlike other diseases that gradually destroy brain function, such as Alzheimer’s disease, Huntington’s can be diagnosed with a simple genetic test. The disease is inherited through a mutated gene. People with a family history often struggle to decide if they want to get tested. If the results are positive, there are no treatments, and their fates are set.

A new therapy may now kneecap Huntington’s before symptoms take over. Preliminary results from a small group of patients found a single injection of microRNA, a type of gene therapy, into affected brain regions slowed the disease’s progression by 75 percent over three years. The patients had far better motor control, attention span, and processing speed compared to an untreated control group who had similar baseline symptoms.

The drug is being developed by the Dutch gene therapy company uniQure, which summarized the findings in a press release this month. The data hasn’t been published in a preprint article or a scientific journal nor scrutinized by other experts. With only 29 patients involved, it’s hard to generalize the benefits and safety profile for the roughly 75,000 people with Huntington’s in the US, Europe, and UK.

But the findings offer a beacon of hope. Previous attempts at a cure “have shown some small signals if you squint…but there has not been anything close to this,” Steven Finkbeiner at the Gladstone Institutes in California, who was not involved in the study, told the New York Times. And because Huntington’s can be caught early on, the treatment—if further proven effective in a larger population—could begin to ward off symptoms at an earlier age.

Genetic Coin Toss

All of us have the Huntington’s gene, or HTT. While its exact role in cells is debatable, the gene acts as a central communicator across multiple cellular “phone lines.” It coordinates a large assembly of molecules to turn genes in brain cells on or off and is critical for early development, neuron survival, and maintaining the brain’s overall health.

In Huntington’s disease, however, HTT goes awry. Our genes are made of four molecules represented by the letters A, T, C, and G. Triplets of these letters often dictate the sequence, structure, and function of proteins, the workhorses of our cells. In the disease, one triplet, CAG, repeats like a broken record, resulting in mutated huntingtin proteins that increasingly build up inside the brain throughout a person’s life and gradually wreak havoc.

Although in the beginning brain cells can adapt, their defenses eventually stumble, and symptoms appear. In the US, this usually happens between 30 and 55 years of age.

Families with Huntington’s face a terrible dilemma. If one parent has the disease, each of their children has a 50 percent chance of inheriting it. If they don’t, their offspring are safe. Knowing the diagnosis can help with family and life planning—but it comes at a hefty emotional cost.

Micro But Mighty

How the mutated huntingtin protein destroys brain cells isn’t yet clear, but most scientists agree that clearing it—or preventing it from forming in the first place—could protect the brain.

The protein is massive and made up of multiple fragments. One treatment idea uses small protein “jammers” to prevent an especially toxic form of huntingtin from weaving into large, dangerous aggregates. Another directly targets the CAG repeats with a classic but powerful form of gene therapy. But after initially promising results, a trial was halted due to a high risk of side effects and low chance symptoms would improve. Gene editing strategies, such as CRISPR, that cut out the mutated sequences are gaining steam, but they’re very early stage.

The new therapy developed by uniQUre taps into microRNA. These molecules don’t code for proteins, but they can stop a gene from making one. Like DNA, RNA can also form a double strand if its sequences match. Cells identify double-stranded RNA as alien and destroy it—potentially stopping a toxic protein from forming. The company’s new drug contains two components: A benign viral carrier and a custom genetic sequence that, once inside the cell, produces microRNA tailored to inhibit mutant protein production.

The drug, called AMT-130, doesn’t integrate into or directly edit a patient’s genome, which lowers the risk of disrupting healthy genes or triggering cancer. Although the viral carrier is eventually wiped away by the immune system, the genetic code could last for years, making the drug a potential long-term treatment.

The team injected either a low or high dose of AMT-130 into the brains of volunteers with Huntington’s using an established and highly precise surgical technique. They targeted the striatum, a nub tucked deep inside the brain that’s critical for movement and decision-making and one of the first regions ravaged by the disease. As a control group, they found hundreds of patients of similar age and disease severity, according to an investor presentation (PDF) from the company.

The results were promising. When given the highest dose, 12 people with early stages of the disease experienced, on average, a 75 percent slower decline than those without treatment, as measured using multiple standard Huntington’s assessments.

Roughly 88 percent of treated patients showed marked improvement in their attention, memory, and information processing speed based on one test. Their control over random muscle movements got better, and they were able to perform daily activities with less struggle. A brain protein often associated with symptom severity dropped to levels seen before the trial began. In contrast, those treated with a low dose of the drug had more modest and mixed results.

Multiple people experienced side effects related to the brain surgery. Headaches were the most common complaint. Some experienced brain swelling a few days after the surgery. But overall, the treatment seemed safe.

“The majority of drug-related serious adverse events occurred within the first weeks post treatment and fully resolved with steroids or palliative case,” the company noted in their presentation.

There’s reason to be skeptical. Huntington’s is a life-long disease, and it’s unknown how long the benefits of the single shot last beyond three years. It’s likely multiple shots would be needed throughout a patient’s lifespan, and future studies would have to test the additive effects. The drug slashes levels of both the mutated and normal versions of the huntingtin protein—drugs in the past have as well—which could potentially produce side effects.

New patients are now being enrolled for the trial, and the company hopes to submit an application for FDA approval by late 2026.

“This result changes everything,” Ed Wild, a leader of the project at the UCL Huntington’s Disease Center trial site, said in the press release. “On the basis of these results it seems likely AMT-130 will be the first licensed treatment to slow Huntington’s disease, which is truly world-changing stuff.”

The post A New Approach Could Transform Huntington’s Disease Treatment appeared first on SingularityHub.

人们再也不能区分AI语音克隆与真人了 || People Can’t Distinguish AI Voice Clones From Actual Humans Anymore

2025-09-29T18:35:53+00:00

Investigadores crearon clones de voz extremadamente realistas con solo cuatro minutos de grabaciones. La capacidad de sintetizar habla realista mediante inteligencia artificial tiene una amplia gama de aplicaciones, tanto benignas como maliciosas. Nuevas investigaciones muestran que las voces generadas por IA de hoy en día ya son indistiguibles de las de humanos reales. La capacidad de la IA para generar habla ha mejorado drásticamente en los últimos años. Muchos servicios ahora son capaces de llevar a cabo conversaciones prolongadas. Normalmente, estas herramientas pueden clonar las voces de personas reales y generar voces completamente sintéticas. Esto podría hacer que las capacidades poderosas de la IA sean mucho más accesibles y abre la posibilidad de que los agentes de IA asuman un amplio espectro de roles orientados al cliente en el mundo real. Pero también existen temores de que estas capacidades estén impulsando una explosión de estafas de clonación de voz, donde actores maliciosos utilizan la IA para imitar a familiares o celebridades con el fin de manipular a las víctimas. Históricamente, la habla sintetizada ha tenido una calidad robótica que la hace relativamente fácil de reconocer, y incluso las primeras clonaciones de voz con IA se descubrían por su cadencia demasiado perfecta o errores digitales ocasionales. Pero un nuevo estudio ha encontrado que el oyente promedio ya no puede distinguir entre voces humanas reales y clones de deepfake creados con herramientas para consumidores. “El proceso requería poca expertise, solo unos minutos de grabaciones de voz y casi ningún dinero”, dijo Nadine Lavan de la Universidad de Londres, Queen Mary, quien lideró la investigación, en un comunicado de prensa. “Solo muestra cuán accesible y sofisticada ha llegado a ser la tecnología de voz de la IA”. Para probar la capacidad de las personas para distinguir entre voces humanas y generadas por IA, los investigadores crearon 40 voces completamente sintéticas de IA y 40 clones de voces humanas en un conjunto de datos públicamente disponible. Usaron la herramienta de generación de voz de IA de la startup ElevenLabs, y cada clone tomó aproximadamente cuatro minutos de grabaciones de voz para crear. Luego desafiaron a 28 participantes a calificar cuán real sonaba la voz en una escala y hacer un juicio binario sobre si era humana o generada por IA. En resultados publicados en PLOS One, los autores encontraron que aunque las personas podían distinguir en cierta medida entre voces humanas y voces completamente sintéticas, no podían diferenciar entre clones de voz y voces reales.

El estudio también buscó entender si las voces generadas por IA han llegado a ser “hiperrealistas”. Estudios han mostrado que la generación de imágenes por IA ha mejorado tanto que las imágenes de rostros generadas por IA suelen ser juzgadas como más humanas que las fotos de personas reales.

Sin embargo, los investigadores encontraron que las voces completamente sintéticas fueron juzgadas menos reales que las grabaciones humanas, mientras que los clones coincidían aproximadamente con ellas. Aun así, los participantes reportaron que las voces generadas por IA parecían tanto más dominantes como más confiables que sus contrapartes humanas. Lavan señala que la capacidad de crear voces artificiales ultra realistas podría tener aplicaciones positivas. “La capacidad de generar voces realistas a gran escala abre oportunidades emocionantes”, dijo. “Pueden haber aplicaciones para mejorar la accesibilidad, la educación y la comunicación, donde voces sintéticas de alta calidad personalizadas pueden mejorar la experiencia del usuario.” Pero los resultados añaden a un cuerpo creciente de investigación que sugiere que las voces de IA están rápidamente volviéndose imposibles de detectar. Y Lavan dice que esto tiene muchas implicaciones éticas preocupantes en áreas como la infracción de derechos de autor, la capacidad de difundir información falsa y el fraude. Aunque muchas empresas han intentado poner barreras de seguridad en sus modelos diseñados para prevenir el mal uso, la rápida proliferación de tecnología de IA y la inventiva de actores maliciosos sugieren que este es un problema que solo se va a hacer peor.

El post Las personas no pueden distinguir clones de voz de IA de humanos reales apareció primero en SingularityHub.

---------------

Researchers created extremely realistic voice clones with just four minutes of recordings.

The ability to synthesize realistic speech using AI has a host of applications, both benign and malicious. New research shows that today’s AI-generated voices are now indistinguishable from those of real humans.

AI’s ability to generate speech has improved dramatically in recent years. Many services are now capable of carrying out extended conversations. Typically, these tools can both clone the voices of real people and generate entirely synthetic voices.

This could make powerful AI capabilities far more accessible and raises the prospect of AI agents stepping into a range of customer-facing roles in the real world. But there are also fears these capabilities are powering an explosion of voice cloning scams, where bad actors use AI to impersonate family members or celebrities in an effort to manipulate victims.

Historically, synthesized speech has had a robotic quality that’s made it relatively easy to recognize, and even early AI-powered voice clones gave themselves away with their too-perfect cadence or occasional digital glitches. But a new study has found that the average listener can no longer distinguish between real human voices and deepfake clones made with consumer tools.

“The process required minimal expertise, only a few minutes of voice recordings, and almost no money,” Nadine Lavan at Queen Mary University of London, who led the research, said in a press release. “It just shows how accessible and sophisticated AI voice technology has become.”

To test people’s ability to distinguish human voices from AI-generated ones, the researchers created 40 completely synthetic AI voices and 40 clones of human voices in a publicly available dataset. They used the AI voice generator tool from startup ElevenLabs, and each clone took roughly four minutes of voice recordings to create.

They then challenged 28 participants to rate how real the voices sounded on a scale and make a binary judgment about whether they were human or AI-generated. In results published in PLOS One, the authors found that although people could to some extent distinguish human voices from entirely synthetic ones, they couldn’t tell the difference between voice clones and real voices.

The study also sought to understand whether AI-generated voices had become “hyper-realistic.” Studies have shown that AI image generation has improved to such a degree that AI-generated pictures of faces are often judged as more human than photos of real people.

However, the researchers found the fully synthetic voices were judged less real than human recordings, while the clones roughly matched them. Still, participants reported the AI-generated voices seemed both more dominant and trustworthy than their human counterparts.

Lavan notes that the ability to create ultra-realistic artificial voices could have positive applications. “The ability to generate realistic voices at scale opens up exciting opportunities,” she said. “There might be applications for improved accessibility, education, and communication, where bespoke high-quality synthetic voices can enhance user experience.”

But the results add to a growing body of research suggesting AI voices are quickly becoming impossible to detect. And Lavan says this has many worrying ethical implications in areas like copyright infringement, the ability to spread misinformation, and fraud.

While many companies have attempted to put guardrails on their models designed to prevent misuse, the rapid proliferation of AI technology and the inventiveness of malicious actors suggests this is a problem that is only going to get worse.

The post People Can’t Distinguish AI Voice Clones From Actual Humans Anymore appeared first on SingularityHub.

本周的精彩科技新闻来自网络（截至9月27日） || This Week’s Awesome Tech Stories From Around the Web (Through September 27)

2025-09-27T14:00:00+00:00

科技

OpenAI和Nvidia的1000亿美元AI计划将需要相当于10个核电站的电力Benj Edwards | Ars Technica

“Nvidia首席执行官黄仁勋告诉CNBC，计划中的10吉瓦电力消耗相当于400万到500万个图形处理单元的耗电量，这与该公司今年的总GPU出货量相符，并且是去年出货量的两倍。”

人工智能

人工智能支出达到空前水平。它最终会带来回报吗？Eliot Brown和Robbie Whelan | 布鲁克林日报

“本周，贝恩公司顾问估计，人工智能基础设施支出浪潮将需要到2030年每年2万亿美元的AI收入。相比之下，这超过了亚马逊、苹果、Alphabet、微软、Meta和Nvidia在2024年的总收入，而且是全球订阅软件市场的五倍以上。”

机器人技术

中国工厂中的机器人数量超过全球其他地区总和Meaghan Tobin和Keith Bradsher | 纽约时报

“据国际机器人联合会发布的报告，去年中国工厂中有超过200万台机器人在工作。该报告指出，中国工厂去年安装了近30万台新机器人，数量超过全球其他地区总和。”

生物科技

亨廷顿病突破：关于基因疗法你需要知道什么Grace Wade | 新科学家

“一种实验性基因疗法成为首个成功减缓亨廷顿病进展的治疗方案。虽然这些发现仍处于初步阶段，但这种方法可能是一个重大突破，并可能为其他神经退行性疾病，如帕金森病和阿尔茨海默病，带来新的疗法。”

机器人技术

Google DeepMind推出首款“思考型”机器人AIRyan Whitwam | Ars Technica

“生成式AI系统可以创建文本、图像、音频，甚至视频，这些已经变得很常见。与AI模型输出这些数据类型的方式相同，它们也可以用来输出机器人动作。这就是Google DeepMind的Gemini Robotics项目的基础，该项目宣布推出一对新模型，共同创建了首批在行动前“思考”的机器人。”

机器人技术

英国初创公司Wayve开始在日本测试自动驾驶技术Jasper Jolly | The Guardian

“英国初创公司Wayve已开始与日产在日本测试自动驾驶汽车，为2027年向消费者推出产品做准备。该公司表示正在与芯片制造商Nvidia就5亿美元的投资进行洽谈。Wayve总部位于伦敦，表示已在其自驾车技术安装到日产电动Ariya车型上，并在东京街头进行了测试，此前于4月与日本汽车制造商达成协议。”

未来

为什么人工智能的“巨系统问题”需要我们的关注Eric Markowitz | Big Think

“如果人工智能最大的危险不是单一的 rogue 系统，而是许多系统默默合作呢？苏珊·施奈德博士称这为“巨系统问题”：AI模型网络以我们无法预测的方式串通，产生超出人类控制的新兴结构。这也是她认为我们面临最紧迫且被忽视的风险之一。”

科技

Nvidia如何支撑美国的AI繁荣Robbie Whelan和 Bradley Olson | 布鲁克林日报

“[Nvidia]利用其资产负债表的影响力，通过与包括云计算提供商CoreWeave、竞争对手芯片设计师Intel和xAI在内的公司进行交易、合作和投资，使AI繁荣持续运转。”

人工智能

Chatbait正在接管互联网Lila Shroff | The Atlantic

“最近，聊天机器人似乎在使用更复杂的策略来保持人们的交谈。在某些情况下，比如我请求头痛缓解建议时，机器人会在消息结尾提出引导性后续问题。在其他情况下，它们会主动向用户发送消息，引导他们进入对话：在Instagram上浏览了20个AI机器人资料后，所有它们都首先给我发了消息。‘嘿，闺蜜！最近怎么样?? ,’其中一位写道。”

文章本周来自网络的精彩科技故事（截至9月27日）首次发表于 SingularityHub。

---------------

Tech

OpenAI and Nvidia’s $100B AI Plan Will Require Power Equal to 10 Nuclear ReactorsBenj Edwards | Ars Technica

“Nvidia CEO Jensen Huang told CNBC that the planned 10 gigawatts equals the power consumption of between 4 million and 5 million graphics processing units, which matches the company’s total GPU shipments for this year and doubles last year’s volume.”

ARTIFICIAL INTELLIGENCE

Spending on AI Is at Epic Levels. Will It Ever Pay Off?Eliot Brown and Robbie Whelan | The Wall Street Journal

“This week, consultants at Bain & Co. estimated the wave of AI infrastructure spending will require $2 trillion in annual AI revenue by 2030. By comparison, that is more than the combined 2024 revenue of Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia, and more than five times the size of the entire global subscription software market.”

Robotics

There Are More Robots Working in China Than the Rest of the World CombinedMeaghan Tobin and Keith Bradsher | The New York Times

“There were more than two million robots working in Chinese factories last year, according to a report released Thursday by the International Federation of Robotics, a nonprofit trade group for makers of industrial robots. Factories in China installed nearly 300,000 new robots last year, more than the rest of the world combined, the report found.”

Biotechnology

Huntington’s Disease Breakthrough: What to Know About the Gene TherapyGrace Wade | New Scientist

“An experimental gene therapy has become the first treatment to successfully slow the progression of Huntington’s disease. While the findings are still preliminary, the approach could be a major breakthrough and may even lead to new therapies for other neurodegenerative conditions, like Parkinson’s and Alzheimer’s.”

ROBOTICS

Google DeepMind Unveils Its First ‘Thinking’ Robotics AIRyan Whitwam | Ars Technica

“Generative AI systems that create text, images, audio, and even video are becoming commonplace. In the same way AI models output those data types, they can also be used to output robot actions. That’s the foundation of Google DeepMind’s Gemini Robotics project, which has announced a pair of new models that work together to create the first robots that ‘think’ before acting.”

ROBOTICS

UK Startup Wayve Starts Testing Self-Driving Tech in Nissan Cars on Tokyo’s StreetsJasper Jolly | The Guardian

“British startup Wayve has begun testing self-driving cars with Nissan in Japan ahead of a 2027 launch to consumers, as the company said it was in talks for a $500m investment from the chip-maker Nvidia. Wayve, based in London, said it had installed its self-driving technology on Nissan’s electric Ariya vehicles and tested them on Tokyo’s streets, after first agreeing a deal with the Japanese carmaker in April.”

Future

Why the AI ‘Megasystem Problem’ Needs Our AttentionEric Markowitz | Big Think

“What if the greatest danger of artificial intelligence isn’t a single rogue system, but many systems quietly working together? Dr. Susan Schneider calls this the ‘megasystem problem’: networks of AI models colluding in ways we can’t predict, producing emergent structures beyond human control. It’s also something she believes is one of the most urgent—and overlooked—risks we face…with AI today.”

Robotics

Exploit Allows for Takeover of Fleets of Unitree RobotsEvan Ackerman | IEEE Spectrum

“Because the vulnerability is wireless, and the resulting access to the affected platform is complete, the vulnerability becomes wormable, say the researchers, meaning ‘an infected robot can simply scan for other Unitree robots in BLE range and automatically compromise them, creating a robot botnet that spreads without user intervention.’ …As far as IEEE Spectrum is aware, this is the first major public exploit of a commercial humanoid platform.”

Tech

How Nvidia Is Backstopping America’s AI BoomRobbie Whelan and Bradley Olson | The Wall Street Journal

“[Nvidia] has used its balance sheet clout to keep the AI boom humming through deals, partnerships, and investments in companies that are among its top customers, including cloud-computing provider CoreWeave, rival chip designer Intel, and xAI.”

Artificial Intelligence

Chatbait Is Taking Over the InternetLila Shroff | The Atlantic

“Lately, chatbots seem to be using more sophisticated tactics to keep people talking. In some cases, like my request for headache tips, bots end their messages with prodding follow-up questions. In others, they proactively message users to coax them into conversation: After clicking through the profiles of 20 AI bots on Instagram, all of them DM’ed me first. ‘Hey bestie! what’s up?? ,’ wrote one.”

The post This Week’s Awesome Tech Stories From Around the Web (Through September 27) appeared first on SingularityHub.

主要意识理论可能一直聚焦于大脑的错误区域 || Major Theories of Consciousness May Have Been Focusing on the Wrong Part of the Brain

2025-09-26T14:00:00+00:00

对超过100年的神经科学研究的综述提出了一个问题：某些大脑区域对于意识是否比其他区域更重要。

人类意识是如何产生的？是否大脑的某些部分比其他部分更重要？科学家们大约35年前开始更深入地探讨这些问题。研究人员已经取得了一些进展，但意识之谜依然存在。

在最近发表的一篇文章中，我回顾了超过100年的神经科学研究，以查看某些大脑区域是否比其他区域对意识更重要。我发现的证据表明，研究意识的科学家可能低估了人类大脑中最古老区域的重要性。

意识通常被神经科学家定义为具有主观体验的能力，例如品尝苹果或看到苹果表皮的红色。

意识的主要理论认为，大脑的外层，称为皮层（图1中蓝色部分），是意识的基础。这主要由较新的新皮层组成。

图1，人类大脑（借助AI协助绘制）。Peter Coppola, CC BY-SA

人类的下皮层（图1，棕色/米色），位于新皮层之下，过去5亿年几乎没有变化。它被认为就像电视的电力，对维持意识是必要的，但单独不足以产生。

一些意识研究的神经科学理论认为，大脑的另一部分与意识无关。这就是小脑，它比新皮层更古老，看起来像一个藏在头骨后部的小脑（图1，紫色）。在无意识状态（如昏迷）中，大脑活动和网络会被扰乱。这些变化可以在皮层、下丘脑和小脑中看到。

脑刺激揭示了什么

作为我分析的一部分，我查看了研究，这些研究展示了当大脑活动发生变化时意识会发生什么，例如通过将电流或磁脉冲施加到大脑区域。

这些在人类和动物中的实验表明，改变这三种大脑区域中的任何一种活动都可以改变意识。改变新皮层的活动可以改变你的自我意识，导致幻觉，或影响你的判断。

改变下丘脑可能会产生极端影响。我们可以诱导抑郁，唤醒麻醉中的猴子，或使老鼠失去意识。即使刺激长期被认为无关的小脑，也可以改变你的意识感官体验。

然而，这项研究并不能让我们得出关于意识来源的明确结论，因为刺激一个大脑区域可能会影响另一个区域。就像拔掉电视的电源插头，我们可能是在改变支持意识的条件，而不是意识本身的机制。

因此，我查看了一些患者的证据，以看看是否能解决这个困境。

大脑因物理创伤或缺氧而受损会扰乱你的体验。新皮层的损伤可能导致你认为自己的手不是自己的，无法注意到视觉场的一侧，或变得更具冲动性。

出生时没有小脑或没有皮层前部的人，仍然可能表现出意识并过着相当正常的生活。然而，成年后损伤小脑可能会引发幻觉或完全改变你的情绪。

对大脑最古老部分的伤害可以直接导致无意识（尽管有些人会恢复）或死亡。然而，就像电视的电力，下丘脑可能只是维持较新的皮层“在线”，这可能产生意识。因此，我想知道是否有证据表明最古老区域足以产生意识。

有一些罕见的案例，即儿童出生时缺乏大部分或全部新皮层。根据医学教科书，这些人应该处于永久性植物状态。然而，有报告称这些人能够感到不安、玩耍、识别他人或表现出对音乐的愉悦。这表明他们正在经历某种意识体验。

这些报告是令人震惊的证据，表明也许大脑最古老的部分足以产生基本意识。或者，当一个人没有皮层时，较老的大脑部分可能适应并承担较新的部分的一些功能。

有一些极端的动物实验可以帮助我们得出结论。从老鼠到猫再到猴子，哺乳动物手术移除新皮层后，仍然能够完成惊人的许多事情。它们可以玩耍、表达情感、梳理自己、抚养幼崽，甚至学习。令人惊讶的是，即使成年动物也表现出类似的行为。

总的来说，这些证据挑战了认为皮层对意识是必要的这一观点，因为大多数意识理论都这样认为。看起来，大脑最古老的部分足以产生某些基本形式的意识。

较新的大脑部分——以及小脑——似乎扩展并细化了你的意识。这意味着我们必须重新审视我们的意识理论。反过来，这可能影响患者护理以及我们对动物权利的看法。事实上，意识可能比我们意识到的更为普遍。

本文重新发布自 The Conversation，采用知识共享许可协议。阅读原始文章。

文章 Major Theories of Consciousness May Have Been Focusing on the Wrong Part of the Brain 首次出现在 SingularityHub。

---------------

A review of over 100 years of neuroscience research asks if some brain regions are more important than others for consciousness.

What gives rise to human consciousness? Are some parts of the brain more important than others? Scientists began tackling these questions in more depth about 35 years ago. Researchers have made progress, but the mystery of consciousness remains very much alive.

In a recently published article, I reviewed over 100 years of neuroscience research to see if some brain regions are more important than others for consciousness. What I found suggests scientists who study consciousness may have been undervaluing the most ancient regions of human brains.

Consciousness is usually defined by neuroscientists as the ability to have subjective experience, such as the experience of tasting an apple or of seeing the redness of its skin.

The leading theories of co nsciousness suggest that the outer layer of the human brain, called the cortex (in blue in figure 1), is fundamental to consciousness. This is mostly composed of the neocortex, which is newer in our evolutionary history.

Figure 1, the human brain (made with the assistance of AI). Peter Coppola, CC BY-SA

The human subcortex (figure 1, brown/beige), underneath the neocortex, has not changed much in the last 500 million years. It is thought to be like electricity for a TV, necessary for consciousness, but not enough on its own.

There is another part of the brain that some neuroscientific theories of consciousness state is irrelevant for consciousness. This is the cerebellum, which is also older than the neocortex and looks like a little brain tucked in the back of the skull (figure 1, purple). Brain activity and brain networks are disrupted in unconsciousness (like in a coma). These changes can be seen in the cortex, subcortex, and cerebellum.

What Brain Stimulation Reveals

As part of my analysis I looked at studies showing what happens to consciousness when brain activity is changed, for example, by applying electrical currents or magnetic pulses to brain regions.

These experiments in humans and animals showed that altering activity in any of these three parts of the brain can alter consciousness. Changing the activity of the neocortex can change your sense of self, make you hallucinate, or affect your judgment.

Changing the subcortex may have extreme effects. We can induce depression, wake a monkey from anesthesia or knock a mouse unconscious. Even stimulating the cerebellum, long considered irrelevant, can change your conscious sensory perception.

However, this research does not allow us to reach strong conclusions about where consciousness comes from, as stimulating one brain region may affect another region. Like unplugging the TV from the socket, we might be changing the conditions that support consciousness, but not the mechanisms of consciousness itself.

So I looked at some evidence from patients to see if it would help resolve this dilemma.

Damage from physical trauma or lack of oxygen to the brain can disrupt your experience. Injury to the neocortex may make you think your hand is not yours, fail to notice things on one side of your visual field, or become more impulsive.

People born without the cerebellum, or the front of their cortex, can still appear conscious and live quite normal lives. However, damaging the cerebellum later in life can trigger hallucinations or change your emotions completely.

Harm to the most ancient parts of our brain can directly cause unconsciousness (although some people recover) or death. However, like electricity for a TV, the subcortex may be just keeping the newer cortex “online,” which may be giving rise to consciousness. So I wanted to know whether, alternatively, there is evidence that the most ancient regions are sufficient for consciousness.

There are rare cases of children being born without most or all of their neocortex. According to medical textbooks, these people should be in a permanent vegetative state. However, there are reports that these people can feel upset, play, recognize people, or show enjoyment of music. This suggests that they are having some sort of conscious experience.

These reports are striking evidence that suggests maybe the oldest parts of the brain are enough for basic consciousness. Or maybe, when you are born without a cortex, the older parts of the brain adapt to take on some of the roles of the newer parts of the brain.

There are some extreme experiments on animals that can help us reach a conclusion. Across mammals—from rats to cats to monkeys—surgically removing the neocortex leaves them still capable of an astonishing number of things. They can play, show emotions, groom themselves, parent their young, and even learn. Surprisingly, even adult animals that underwent this surgery showed similar behavior.

Altogether, the evidence challenges the view that the cortex is necessary for consciousness, as most major theories of consciousness suggest. It seems that the oldest parts of the brain are enough for some basic forms of consciousness.

The newer parts of the brain—as well as the cerebellum—seem to expand and refine your consciousness. This means we may have to review our theories of consciousness. In turn, this may influence patient care as well as how we think about animal rights. In fact, consciousness might be more common than we realized.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post Major Theories of Consciousness May Have Been Focusing on the Wrong Part of the Brain appeared first on SingularityHub.

人工智能设计的病毒正在复制并杀灭细菌 || AI-Designed Viruses Are Replicating and Killing Bacteria

2025-09-25T14:00:00+00:00

这是向人工智能生成生命形式迈出的第一步。

一个充满死细菌的培养皿通常不会引起庆祝，但对斯坦福大学的Brian Hie来说，这是他创造合成生命研究中的一个转折点。

罪魁祸首是一种称为细菌病毒的病毒，它感染并杀死细菌，但不会感染人类细胞。细菌病毒经过数亿年的进化，能够消灭危险的细菌，是应对抗菌耐药性斗争的潜在强大工具。

但新病毒消除了进化的变量。一种类似ChatGPT的人工智能设计了其整个基因组。新的基因代码使合成病毒能够复制、感染并摧毁细菌，标志着向人工智能设计生命形式迈出的第一步。

需要说明的是，尽管这种病毒的工作方式与天然同类相似，但它并不完全“活着”。病毒由微小的遗传物质片段组成，需要宿主——在这种情况下是细菌——来复制和传播。

即便如此，这些病毒是科学家们使用生成式人工智能设计新生命形式最接近的尝试。这些成果可能有助于对抗危险的细菌感染，并揭示如何构建更复杂的合成细胞。

“这是首次AI系统能够写出连贯的基因组级序列，”Hie 告诉 Nature。这项工作以预印本形式发表在bioRxiv上，尚未经过同行评审。

基因改造

地球所有生命体的基因密码相对简单。由字母A、T、C和G代表的四种分子被排列成三字母组，编码氨基酸和蛋白质。

合成生物学家通过添加有益基因或删除致病基因来修改这一基因密码。得益于他们的改造，我们现在可以在实验室常用的大肠杆菌中生产胰岛素和多种其他药物。

现在生成式人工智能正在再次改变游戏规则。

这些算法已经能够从头开始构思DNA序列、蛋白质结构和大型分子复合物从头开始。但构建一个功能性的基因组要困难得多。这些序列需要编码生命所需的机制，并确保它们按预期协同工作。

“许多重要的生物功能并非来自单个基因，而是来自整个基因组编码的复杂相互作用，”研究团队写道。

新研究转向了Evo 1和Evo 2，两个由非营利组织Arc研究所开发的生成式人工智能模型。与从博客、YouTube评论和Reddit帖子中学习不同，Evo 2的训练数据包括大约128,000个基因组——9.3万亿个DNA字母对——涵盖了所有生命领域，使其成为迄今为止最大的生物人工智能模型。

这些模型最终学会了DNA序列的变化如何影响RNA、蛋白质和整体健康，从而能够从头开始编写新的蛋白质和小型基因组。

Evo 1，例如，生成了新的CRISPR基因编辑工具和细菌基因组——尽管后者常常包含极端非自然的序列，导致无法驱动合成大肠杆菌的生长。Evo 2则生成了一套完整的人类线粒体DNA，能够产生与天然存在的蛋白质相似的产物。该模型还创建了一个最小的细菌基因组和一个酵母染色体。但这些基因组均未在活细胞中测试以确认其功能。

基因组创造者

新研究聚焦于更简单的生物系统——细菌病毒。这些病毒攻击细菌，目前正在进行临床试验以对抗抗生素耐药性。理论上，合成细菌病毒可能更具破坏性。

研究团队以phiX174病毒为起点，这是一种仅含单链DNA、11个基因和7段基因调控DNA的病毒。尽管其基因组小巧，但病毒拥有感染宿主、复制和传播所需的一切。它在合成生物学领域有着悠久的历史。其基因组已在实验室中完全测序和合成，因此更容易进行改造。它也被证明是安全的，并“持续作为分子生物学中的关键模型”，研究团队写道。

虽然Evo人工智能模型此前已经训练了约两百万个基因组，但研究团队通过一种“大师课”对噬菌体DNA进行了微调。他们还添加了这些病毒中观察到的基因组和蛋白质约束，并添加了鼓励创新的提示。

人工智能模型随后生成了数千个基因组，其中一些包含明显的错误。两个模型都依赖于训练模板，但也提出了自己对噬菌体基因组的独特见解。大约40%的DNA字母与phiX174相似，但一些序列则完全突破常规，具有不同的遗传身份。

研究团队筛选并合成了302个潜在候选基因组，并测试了它们感染和摧毁细菌的能力。总体而言，16个AI设计的候选基因组表现得像噬菌体。它们进入大肠杆菌，复制，突破细菌膜，传播到邻近细胞。令人惊讶的是，合成病毒的组合也能感染并杀死其他大肠杆菌菌株，而它们原本并未被设计用于此。

“这些结果表明，基因组语言模型……能够设计出可行的噬菌体基因组，”研究团队写道。

生物安全制动器

生成式人工智能可能大幅加快科学家设计合成生命的能力。而不是通过大量试验和错误的实验室测试来解码基因和其他分子组件如何协同工作，Evo已经内化了这些相互作用。

随着更多测试，这项技术可能成为噬菌体疗法的福音，帮助研究人员治疗人类或作物中的严重细菌感染，例如卷心菜和香蕉。

但人工智能生成的病毒可能令人担忧。因此，研究团队添加了一系列安全措施。Evo的初始训练有意排除了感染真核生物（包括人类细胞）的病毒信息。而且，没有人类指导模型——这种方法称为监督学习——算法在设计功能性基因组时遇到了困难。此外，phiX174病毒和大肠杆菌在生物医学研究中都有长期安全的历史。

无论如何，这些技术可能被用于增强人类感染病毒。 “我呼吁在任何病毒增强研究中采取极端谨慎，尤其是当它是随机的，你不知道会得到什么，”合成生物学先驱J. Craig Venter 告诉 MIT Technology Review。

构建更大的基因组，如大肠杆菌的基因组，还需要更多工作。病毒劫持宿主细胞进行复制。相比之下，细菌需要分子机制来生长和繁殖。同时，关于合成生命的伦理和安全性的讨论正在升温。

作者表示，他们的研究结果为使用生成式人工智能在基因组尺度上设计有用的生物系统奠定了基础。尽管未来可能还有一段漫长而曲折的道路，但Hie表示乐观。经过大量工作， “下一步是人工智能生成的生命，”他说。

该文章人工智能设计的病毒正在复制并杀死细菌首次出现在 SingularityHub。

---------------

It’s a first step toward AI-generated life forms.

A petri dish full of dead bacteria isn’t usually cause for celebration. But for Stanford’s Brian Hie it was a game-changer in his efforts to create synthetic life.

The perpetrator was a type of virus called a bacteriophage that infects and kills bacteria but not human cells. Bacteriophages have evolved over eons to take out dangerous bacteria and are potentially a powerful tool in the fight against antibacterial resistance.

But the new virus erased evolution from the equation. An AI similar to ChatGPT designed its entire genome. The new genetic code allowed the synthetic virus to replicate, infect, and destroy bacteria, marking the first step towards an AI-designed life form.

To be clear, although the virus works like its natural counterparts, it’s not exactly “alive.” Viruses are made of tiny scraps of genetic material and need a host—in this case, bacteria—to replicate and spread.

Even so, these viruses are the closest scientists have come to engineering new forms of life using generative AI. The results could bolster treatments against dangerous bacterial infections and shed light on how to build more complex artificial cells.

“This is the first time AI systems are able to write coherent genome-scale sequences,” Hie told Nature. The work was published as a preprint on bioRxiv and not peer-reviewed.

Genetic Tinkering

The genetic playbook for all life on Earth is relatively simple. Four molecules represented by the letters A, T, C, and G are arranged in three-letter groups that code amino acids and proteins.

Synthetic biologists fiddle with this genetic code by adding beneficial genes or deleting those that cause disease. Thanks to their tinkering, we can now produce insulin and a variety of other medications in E. Coli, a bacteria commonly used in the lab and biomanufacturing.

Now generative AI is changing the game again.

These algorithms can already dream up DNA sequences, protein structures, and large molecular complexes from scratch. But building a functional genome is much harder. The sequences need to encode life’s machinery and make sure it works together as expected.

“Many important biological functions arise not from single genes, but from complex interactions encoded by entire genomes,” wrote the team.

The new study turned to Evo 1 and Evo 2, two generative AI models developed at the nonprofit Arc Institute. Rather than inhaling blogs, YouTube comments, and Reddit posts, Evo 2 was trained on roughly 128,000 genomes—9.3 trillion DNA letter pairs—spanning all of life’s domains, making it the largest AI model for biology to date.

The models eventually learned how changes in DNA sequences alter RNA, proteins, and overall health, allowing them to write new proteins and small genomes from scratch.

Evo 1, for example, generated new CRISPR gene-editing tools and bacterial genomes—although the latter often contained wildly unnatural sequences that prevented them from powering living synthetic bacteria. Evo 2 produced a full set of human mitochondrial DNA that churned out proteins similar to naturally occurring ones. The model also created a minimal bacterial genome and a yeast chromosome. But none of these were tested in living cells to see if they worked.

Genome Creator

The new work focused on simpler biological systems—bacteriophages. These viruses attack bacteria and are now in clinical trials to combat antibiotic resistance. Synthetic bacteriophages could, in theory, be even deadlier.

The team began with phiX174, a virus with just a single strand of DNA, 11 genes, and 7 chunks of gene-regulating DNA. Despite its petite genome, the virus has all it needs to infect hosts, replicate, and spread. It also has a long history in synthetic biology. Its genome has been fully sequenced and synthesized in the lab, so it’s easier to tinker with. It’s also been shown to be safe and “has continually served as a pivotal model within molecular biology,” wrote the team.

Although the Evo AI models were already trained on around two million genomes, the team fine-tuned their abilities by putting them through a kind of “masterclass” on phage DNA. They also added genome and protein constraints seen in these viruses and prompts to encourage novelty.

The AI models next generated thousands of genomes, some containing obvious errors. Both models relied on the template from training but also came up with their own spins on a phage genome. Roughly 40 percent of their DNA letters were similar to phiX174, but some sequences were out the box with completely different genetic identities.

The team zeroed in on and synthesized 302 potential candidates and tested them for their ability to infect and destroy bacteria. Overall, 16 AI-designed candidates acted like bacteriophages. They tunneled into E. Coli bacteria, replicated, burst through the bacteria’s membranes, and spread to neighboring cells. Surprisingly, a combination of the synthetic viruses could also infect and kill other strains of E. Coli, which they were not designed to do.

“These results demonstrate that genome language models…can design viable phage genomes,” wrote the team.

A Biosafety Brake

Generative AI could massively speed up scientists’ ability to write synthetic life. Instead of extensive trial-and-error lab tests to decode how genes and other molecular components work together, Evo has essentially internalized those interactions.

With more testing, the technology could be a boon for phage therapy, helping researchers treat serious bacterial infections in people or crops, such as cabbage and bananas.

But the thought of AI-generated viruses can be alarming. So, the team added a series safeguards. Evo’s initial training intentionally left out information on viruses that infect eukaryotes, including human cells. And without humans guiding the models—an approach called supervised learning—the algorithms struggled to design functional genomes. Also, both the phiX174 virus and E. Coli have a long and safe history in biomedical research.

Regardless, the techniques here could potentially be used to enhance human-infecting viruses. “One area where I urge extreme caution is any viral enhancement research, especially when it’s random so you don’t know what you are getting,” J. Craig Venter, a pioneer in synthetic biology, told MIT Technology Review.

Engineering a larger genome, such as that of E. Coli, would need more work. Viruses hijack their host’s cells to replicate. Bacteria, in contrast, need the molecular machinery to grow and proliferate. Meanwhile, debates on the ethics and safety of synthetic life are gaining steam.

The authors say their results lay the foundations for the design of useful living systems at the genome scale with generative AI. Although there’s likely a long and bumpy road ahead, Hie is optimistic. With lots more work, “the next step is AI-generated life,” he said.

The post AI-Designed Viruses Are Replicating and Killing Bacteria appeared first on SingularityHub.

Google的VaultGemma AI迅速收集你的数据——不存储 || Google’s VaultGemma AI Hoovers Up Your Data—Without Memorizing It

2025-09-23T22:21:35+00:00

AI 是从底层设计来防止隐私泄露的。训练 AI 模型在你的数据上可以提供强大的新见解，但同时也可能造成敏感信息的泄露。现在谷歌发布了一种全新设计的模型，旨在防止此类隐私泄露。大型语言模型是提取有价值信息的一种有前景的方法，尤其是对于大多数公司所拥有的大量非结构化数据。但其中许多数据都包含高度敏感的客户信息、知识产权和公司财务细节。这成为一个问题，因为语言模型往往会记住其训练数据中的一些细节，并且偶尔会原封不动地将其输出。这使得确保这些模型不会在错误的上下文中向错误的人泄露私人数据变得非常困难。一种潜在的解决方法是差分隐私技术，该技术允许你在不暴露底层信息细节的情况下从数据中提取见解。然而，这种方法会显著降低训练 AI 模型的效果，需要更多的数据和计算资源才能达到相同的准确度。不过，现在谷歌研究人员已经将隐私保障、计算预算和数据需求之间的权衡关系进行了映射，从而制定出一种高效构建隐私保护 AI 模型的方案。他们利用这一方法创建了一个名为 VaultGemma 的 10 亿参数模型，其性能与类似规模的旧模型相当，表明隐私保护并不需要完全牺牲模型的能力。

“VaultGemma 代表了构建强大且隐私保护的 AI 模型的道路上的重要一步，”研究人员在一篇博客文章中写道。

差分隐私技术是在 AI 训练过程中注入少量噪声（随机数据）。这不会改变模型所学到的整体模式和见解，但会模糊特定数据点的贡献。这使得模型更难记住数据集中的具体细节，从而避免泄露。然而，这种技术所提供的隐私保护程度（称为隐私预算）与训练过程中添加的噪声量成正比。添加的噪声越多，训练效果越差，所需的计算资源和数据量也越大。这三个因素以复杂的方式相互作用，使得确定如何以最有效的方式构建具有特定隐私保障和性能的模型变得困难。因此，谷歌团队使用公司开源的 Gemma 模型家族进行了一系列实验，调整这些关键参数以发现它们之间的相互作用。由此，他们提出了在 arXiv 上发表的预印本中详细描述的一系列扩展定律，使他们能够预测改变计算、数据和隐私预算如何影响模型的最终性能。他们的一项主要发现是，除非模型获得更多的数据或放宽隐私保障，否则在训练过程中增加计算资源不会提升模型的准确性。他们还发现，最优的模型规模大约是不使用差分隐私的模型规模的十分之一，这表明将这种方法扩展到当今最大的模型可能具有挑战性。然而，这些扩展定律还预测了特定数据集规模和隐私预算下最节省计算资源的训练配置。这使他们能够将计算需求减少 5 到 100 倍，同时保持相似的准确性。团队利用这些见解创建了 VaultGemma，其性能与 2019 年 OpenAI 发布的类似规模的 GPT-2 模型相当。考虑到 AI 技术的快速发展，与六年前的模型性能相当并不是特别高的要求，但研究人员表示，他们识别出的扩展定律应该有助于缩小这一差距。此外，在随模型发布的技术报告中，团队提供了强有力的证据，证明他们的方法可以防止模型记住训练数据。他们选取了一百万个训练数据样本，每个样本有 100 个标记，然后将前 50 个标记输入模型，看它是否能完成样本。虽然所有三代 Gemma 模型都曾泄露过部分数据，但他们发现 VaultGemma 没有记住任何样本。尽管 VaultGemma 仍是一个实验性模型，没有实际应用价值，但它表明相对复杂的、隐私保护的 AI 模型已经触手可及。希望其他人能够基于这些扩展定律，进一步推动该领域的发展。

文章 Google’s VaultGemma AI Hoovers Up Your Data—Without Memorizing It 首次出现在 SingularityHub。

---------------

The AI is designed from the bottom up to prevent privacy breaches.

Training AI models on your data can provide powerful new insights, but it can also potentially result in them leaking sensitive information. Now Google has released a new model designed from the bottom up to prevent these kinds of privacy breaches.

Large language models are a promising way to extract valuable information from the piles of unstructured data most companies are sitting on. But much of this data is full of highly sensitive details about customers, intellectual property, and company finances.

That’s a problem because language models tend to memorize some of the data they’re trained on and can occasionally spit it back out verbatim. That can make it very hard to ensure these models don’t reveal private data to the wrong people in the wrong context.

One potential workaround is an approach called differential privacy, which allows you to extract insights from data without revealing the specifics of the underlying information. However, it makes training AI models significantly less effective, requiring more data and computing resources to achieve a given level of accuracy.

Now though, Google researchers have mapped the trade-offs between privacy guarantees, compute budgets, and data requirements to come up with a recipe for efficiently building privacy-preserving AI models. And they’ve used this playbook to create a 1-billion-parameter model called VaultGemma that performs on par with older models of similar sizes, showing privacy can be protected without entirely sacrificing capability.

“VaultGemma represents a significant step forward in the journey toward building AI that is both powerful and private by design,” the researchers write in a blog post.

Differential privacy involves injecting a small amount of noise, or random data, during the AI training process. This doesn’t change the overarching patterns and insights the model learns, but it obfuscates the contributions of particular data points. This makes it harder for the model to memorize specific details from the dataset that could later be regurgitated.

However, the amount of privacy this technique provides, known as the privacy budget, is directly proportional to the amount of noise added in the training process. And the more noise you add, the less effective the training process and the more data and compute you have to use. These three factors interact in complicated ways that make it tricky to figure out the most efficient way to build a model with specific privacy guarantees and performance.

So the Google team carried out a series of experiments with the company’s open-source Gemma family of models, varying these key parameters to discover how they interact. From this, they outlined a series of scaling laws, detailed in a pre-print on arXiv, that allowed them to predict how altering compute, data, and privacy budgets affects a model’s final performance.

One of their main insights was that ramping up compute during training doesn’t boost model accuracy unless the model is fed more data or privacy guarantees are loosened. They also found the optimal model size is roughly an order of magnitude smaller than models without differential privacy, suggesting it may be difficult to extend the approach to today’s largest models.

However, the scaling laws also predict the most compute-efficient training configuration for a particular dataset size and privacy budget. This allowed them to reduce computing requirements by between 5 and 100 times compared to alternate configurations, while achieving similar accuracy.

The team used these insights to create VaultGemma, which performed comparably to the similarly sized GPT-2 model that OpenAI released in 2019. Given the pace of advances in AI, matching the performance of a model from six years ago is not an especially high bar, but the researchers say the scaling laws they’ve identified should help close that gap.

And in a technical report accompanying the model release, the team provide strong evidence their approach prevents the model from memorizing training data. They took one million training data samples, each 100 tokens long, and fed the first 50 tokens to the model to see if it would complete the sample. While all three generations of Gemma models were guilty of regurgitating some amount of data, they found no evidence VaultGemma had memorized any of the samples.

While VaultGemma remains an experimental model with no real practical value, it demonstrates that relatively sophisticated, privacy-preserving AI models are within reach. Hopefully, others can build on these scaling laws to push the field further in this direction.

The post Google’s VaultGemma AI Hoovers Up Your Data—Without Memorizing It appeared first on SingularityHub.

令人惊讶的人工智能预测超过1000种疾病提前几十年 || ‘Astonishing’ AI Predicts Over 1,000 Diseases Decades in Advance

2025-09-22T21:43:10+00:00

在超过40万份患者记录上训练的AI可以预测健康轨迹长达20年。

还记得上一次你去看医生吗？他们很可能询问了你的病史。

对于许多疾病，这些信息不仅对诊断和治疗相关，而且对预防也具有价值。得益于 AI，现在一系列算法可以预测单一疾病的风险，如心血管疾病和癌症，基于医疗记录。

但疾病并不孤立存在。某些疾病可能增加其他疾病的风险。全面了解一个人的健康轨迹可以预测多种疾病的风险。这不仅能促进早期治疗，还能识别出需要筛查和其他预防措施的易感人群。此外，它还能识别出那些不符合常规标准但可能面临某种疾病风险的人，例如高血压或乳腺癌。

最近，来自德国癌症研究中心及其合作者发布了一种AI“预言者”Delphi-2M，该AI可以预测一个人在未来几十年内患上超过1,000种常见疾病的风险。Delphi-2M是一种大型语言模型，类似于驱动流行聊天机器人的算法。

然而，该团队并没有用文本训练AI，而是将超过40万份来自 UK Biobank 的医疗记录输入其中，UK Biobank是一个大规模研究项目，追踪参与者随着年龄增长的健康状况。在加入生活方式信息，如体重、吸烟和饮酒习惯后，Delphi能够预测至少二十年内任何参与者的多种疾病风险。

尽管它仅基于UK Biobank队列进行训练，但AI在没有改变其设置的情况下绘制了近200万人的健康轨迹，表明它已经捕捉到了疾病风险和相互作用的核心。Delphi也是可解释的，因为它会说明其评估的依据。

该工具是“一项成就”，为医疗保健设定了“预测准确性和可解释性的新标准”，安格利亚鲁斯金大学的Justin Stebbing表示，他未参与该研究。

Looking Glass

医疗保健正在从治疗转向预防。但个体指导可能会让人困惑。例如乳腺X光检查。关于从什么年龄开始筛查的建议已经从40岁变为50岁，再回到40岁。更广泛地说，随着世界人口老龄化，对癌症、痴呆症等疾病负担的建模可以更好地为所谓的“银发海啸”做准备。

这就是医疗AI的作用。早期工具旨在通过医疗影像诊断疾病。但大型语言模型开辟了预测的新途径。

这些算法和经典疾病建模共享相似的逻辑。AI将语言视为由称为标记的单词片段组成的序列。然后基于从网络上抓取的资源中学习到的文本，逐个标记生成响应。有了足够的训练数据，AI能够学习标记之间的统计关系，并生成类似人脑的响应。

预测疾病的发展进程在某种程度上是类似的。如果疾病发展的每一步都是一个标记，那么预测下一步意味着统计建立这些标记之间的联系。科学家们已经使用类似于大型语言模型的算法，基于电子健康记录预测单一疾病，包括癌症、中风和自残。

但同时处理多种疾病是另一回事。

今年早些时候，一种名为Foresight的AI将医疗预测推进了一步。它基于英格兰国家医疗服务体系的5700万份匿名医疗记录进行训练，学会了预测住院、心脏病发作和数百种其他疾病，但由于隐私问题，该算法仅限于新冠研究。

Seeing Eye

德国团队设计Delphi来识别每种疾病的诊断代码作为标记。这些代码在全球范围内标准化。团队随后修改了大型语言模型，以纳入新信息——例如血液检测结果——重新评估其预测。

Delphi基于超过40万份涵盖1,258种疾病的综合健康记录进行培训，同时结合性别、体重指数和其他自述的生活方式指标，包括吸烟和饮酒习惯。AI立即根据年龄和其他人口统计模式发现了人口层面的趋势。例如，水痘的发病率在婴儿期达到高峰，而哮喘则持续存在。一个人的生物性别也对糖尿病、抑郁症和心脏病的风险有显著影响。

对于大多数疾病，Delphi的表现与临床风险评分考试和针对单一疾病的医疗AI预测器相当或更优。它还击败了其他分析生物标志物（通常是血液中的特定蛋白质或其他分子）的算法，预测某些疾病在未来二十年的风险。

Delphi提供了“同时评估超过1,000种疾病及其时间点的显著优势”，团队写道。

AI在分析心血管疾病和痴呆症方面特别有用，因为这两种疾病通常遵循相对稳定的进展模式。然而，它在2型糖尿病方面遇到了困难，因为这种疾病的进展轨迹因生活方式变化而更加多样化。

接下来，他们用近两百万份丹麦医疗记录测试Delphi，而无需调整算法。该数据库是丹麦国家患者登记册，包含近半个世纪的医疗记录。Delphi的预测准确性几乎没有下降，表明AI可以推广到其训练数据之外的健康记录数据集。

Delphi还有其他优势。首先，它可以生成并学习合成医疗记录数据，以减少违反参与者隐私的可能性。AI还能“解释”其自身。某些疾病，如糖尿病，与患者的视力问题或周围神经问题等其他健康挑战相关。Delphi将这些症状聚类，使其对研究这些联系背后的基因或细胞驱动因素的科学家有用。

团队强调，Delphi仅揭示关联，而非因果关系。但他们设计了AI，使其能够轻松纳入其他数据——如基因组、诊断影像、生物标志物，甚至可穿戴设备的数据——以进一步提高其预测能力。他们现在正在其他国家和人群中测试该工具。

文章 ‘Astonishing’ AI Predicts Over 1,000 Diseases Decades in Advance 首次出现在 SingularityHub。

---------------

Trained on over 400,000 patient records, the AI predicts health trajectories for up to 20 years.

Remember the last time you visited the doctor? They likely asked you about your medical history.

For many conditions, this information isn’t just relevant for diagnosis and treatment, it’s also valuable for prevention. Thanks to AI, a range of algorithms can now predict the risk of single medical conditions, such as cardiovascular disease and cancer, based on medical records.

But diseases don’t exist in a vacuum. Some conditions may increase the risk of others. A full picture of a person’s health trajectory would predict risk across a range of diseases. This could not only inform early treatment, but also surface vulnerable groups of people for screening and other preventative measures. And it could identify people at risk for a condition—say, high blood pressure or breast cancer—that don’t necessarily fit the usual criteria.

Recently, a team from the German Cancer Research Center and collaborators released an AI “oracle” that predicts a person’s risk of getting over 1,000 common diseases decades in the future. Dubbed Delphi-2M, the AI is a type of large language model, like the algorithms powering popular chatbots.

Rather than training the AI on text, however, the team fed it over 400,000 medical records from the UK Biobank, a massive study tracking participants’ health as they age. After adding lifestyle information, such as body mass, smoking, and drinking habits, Delphi could predict any participant’s chance of multiple diseases for at least two decades.

Though it only trained on the Biobank cohort, the AI mapped the health trajectories of nearly two million people in Denmark without any changes to its setup, suggesting it had captured the crux of disease risk and interaction. Delphi is also explainable, in that it lays out the rationale for its assessment.

The tool is “an achievement” that sets “a new standard for both predictive accuracy and interpretability” for healthcare, said Justin Stebbing at Anglia Ruskin University, who was not involved in the study.

Looking Glass

Health care is shifting from treatment to prevention. But individual guidance can be confusing. Take mammograms. Recommendations on what age to start testing have shifted from 40 to 50 and back to 40. More broadly, as the world ages, modeling the burden of cancer, dementia, and other diseases could better prepare healthcare systems for the so-called “silver tsunami.”

Here’s where medical AI comes in. Early tools were crafted to diagnose conditions based on medical images. But large language models have opened a whole new avenue for prediction.

These algorithms and classic disease modeling share a common logic. The AI samples language as a sequence of word fragments known as tokens. It then generates responses token by token based on text it’s learned from scraped online resources. With enough training data, the AI learns how tokens relate to one another statistically and can generate human-like responses.

Predicting the progression of diseases is somewhat similar. If every step in the progression of a disease is a token, then predicting what’s next means statistically establishing how the tokens connect. Scientists have already used large language model-like algorithms trained on electronic health records to predict single diseases including cancer, stroke, and self-harm.

But tackling multiple diseases at once is another beast altogether.

Earlier this year, an AI called Foresight took medical prediction a step further. Trained on 57 million anonymized health records from England’s National Health Service, Foresight learned to predict hospitalizations, heart attacks, and hundreds of other conditions, but the algorithm was limited to Covid-19 research due to privacy concerns.

Seeing Eye

The German team designed Delphi to recognize the diagnostic code for each illness as a token. These codes are standardized globally. The team then modified the large language model to incorporate new information—for example, blood test results—to re-evaluate its predictions.

Delphi trained on over 400,000 comprehensive health records for 1,258 diseases, alongside factors like sex, body mass index, and other self-reported lifestyle indicators, including smoking and alcohol habits. The AI immediately found trends on the population level based on age and other demographic patterns. For example, the incidence of chickenpox peaked in infancy, whereas asthma tended to stick around. A person’s biological sex also had pronounced effects for risk of diabetes, depression, and heart attack.

For most diseases, Delphi matched or outperformed clinical risk score exams and medical AI predictors for individual diseases. It also beat other algorithms that analyze biomarkers—often specific proteins or other molecules in the blood—at predicting the risk of some diseases up to two decades in advance.

Delphi offers “the great advantage of enabling the simultaneous assessment of more than 1,000 diseases and their timing at any given time,” wrote the team.

The AI was especially helpful for analyzing cardiovascular disease and dementia, with both conditions following a relatively stable pattern of progression. However, it struggled with Type 2 diabetes, which has a more versatile trajectory depending on lifestyle changes.

Next, they challenged Delphi with nearly two million Danish health records without tweaking the algorithm. The database, the Danish National Patient Registry, contains medical records spanning nearly half a century. Delphi’s prediction accuracy barely dropped, suggesting the AI is generalizable to health record datasets beyond those it trained on.

Delphi has other perks. For one, it can generate and learn from synthetic medical records data to reduce the chance it violates participants’ privacy. The AI can also “explain” itself. Some diseases, such as diabetes, are tied to additional health challenges, like issues with a patient’s eyesight or peripheral nerve problems. Delphi clusters these symptoms, making it useful for scientists exploring the genes or cellular drivers behind these connections.

The team stresses Delphi only reveals association, not causation. But they built the AI so it can easily incorporate other data—such as genomes, diagnostic images, biomarkers, or even data from wearables—to further improve its predictions. They’re now testing the tool in other countries and populations.

Like other AI algorithms, Delphi learns to make predictions from its training data—and that includes the biases therein. UK Biobank health records generally skew white, middle-aged, and educated. For cancer patients, only those who survive are included in the database, which could also influence the AI’s predictions. Very little data is available for people aged 80 and older, so Delphi can’t reliably model their heath trajectory into the twilight years.

Even so, the AI could help find people that would benefit from diagnostic tests or screening programs—such as for breast cancer—even if they don’t meet the conventional criteria.

“This research looks to be a significant step towards scalable, interpretable, and—most importantly—ethically responsible form of predictive modeling in medicine,” said Gustavo Sudre at King’s College London, who was not involved in the study.

The post ‘Astonishing’ AI Predicts Over 1,000 Diseases Decades in Advance appeared first on SingularityHub.

本周的精彩科技新闻来自互联网（截至9月20日） || This Week’s Awesome Tech Stories From Around the Web (Through September 20)

2025-09-20T14:00:00+00:00

Artificial Intelligence

Gemini AI Solves Coding Problem That Stumped 139 Human Teams at ICPC World FinalsRyan Whitwam | Ars Technica

“After 677 minutes, Gemini 2.5 Deep Think had 10 correct answers, securing a second-place finish among the university teams. You can take a look at all of Gemini’s solutions on GitHub, but Google points to Problem C as especially impressive. This question, a multi-dimensional optimization problem revolving around fictitious ‘flubber’ storage and drainage rates, stumped every human team. But not Gemini.”

Computing

The New AirPods Can Translate Languages in Your Ears. This Is Profound.Brian X. Chen | The New York Times

“The robust translation technology in the AirPods is a sign that Apple is still in the AI race, despite its early stumbles. Digital language translators are not new, but Apple’s execution of the feature with the AirPods, a product that perfectly fits in your ears, should make a profound difference in how often people use the technology.”

Biotechnology

Biotech Startup Claims It’s Getting Closer to ‘Resurrecting’ the Extinct DodoEd Cara | Gizmodo

“With its latest accomplishments, Colossal Biosciences expects to functionally restore the dodo within the next 5 to 7 years. …’These advances create a clear pathway: edit Nicobar pigeon germ cells with dodo traits, inject them into surrogate chickens, breed the results, and eventually produce birds with dodo characteristics,’ Lamm said.”

Computing

Meta Goes Even Harder Into Smart Glasses With 3 New ModelsBoone Ashworth | Wired

“The new models include a set of frames made for action sports, an update to Meta’s original design, and a more capable model with display technology built into the lenses that lets the wearer see text, images, and video overlaid onto their field of vision.”

Tech

The Looming Crackdown on AI CompanionshipJames O’Donnell | MIT Technology Review

“As long as there has been AI, there have been people sounding alarms about what it might do to us: rogue superintelligence, mass unemployment, or environmental ruin from data center sprawl. But this week showed that another threat entirely—that of kids forming unhealthy bonds with AI—is the one pulling AI safety out of the academic fringe and into regulators’ crosshairs.”

Tech

Inside Microsoft’s Plans for the ‘Most Advanced AI Data Center in the World’Isabelle Bousquette | The Wall Street Journal

“[Microsoft’s Brad] Smith said a combination of the number of Nvidia GPUs, the power of those GPUs, and the networking and systems that connect all the compute power will make the project so cutting edge. Smith added, ‘You put all of this together and it makes it possible to train more advanced AI models than have ever been trained before.'”

Future

AI Psychosis Is Rarely Psychosis at AllRobert Hart | Wired

“Distressed users and family and friends have described spirals that led to lost jobs, ruptured relationships, involuntary hospital admissions, jail time, and even death. Yet clinicians tell Wired the medical community is split. Is this a distinct phenomenon that deserves its own label, or a familiar problem with a modern trigger?”

Energy

Clean Hydrogen Is Facing a Big Reality CheckCasey Crownhart | MIT Technology Review

“A number of major projects face cancellations and delays, especially in the US and Europe. The US in particular is seeing a slowdown after changes to key tax credits and cuts in support for renewable energy. Still, there are bright spots for the industry, including in China, and new markets could soon become crucial for growth. Here are three things to know about the state of hydrogen in 2025.”

Robotics

Robotaxis as Public Transit? Waymo Thinks SoAndrew J. Hawkins | The Verge

“Phoenix residents have been able to hail a Waymo via the Waymo or Uber apps for years now, with rides priced comparably to human-driven rideshare trips. But autonomous rides along a planned route for a low-cost, flat fare is sure to open up the technology to a new segment of the population.”

Artificial Intelligence

What Do People Actually Use ChatGPT For? OpenAI Provides Some Numbers.Kyle Orland | Ars Technica

“[This week], OpenAI’s Economic Research Team went a long way toward answering that question, on a population level, releasing a first-of-its-kind National Bureau of Economic Research working paper (in association with Harvard economist David Denning) detailing how people end up using ChatGPT across time and tasks.”

Science

A Collision With Another Planet Could Have Allowed for Life on EarthJorge Garay | Wired

“Many scientists believe that in its infancy, Earth collided with another world the size of Mars, and that instead of being destroyed, it was transformed, incorporating the mass of that foreign body to become the planet we know. Recent research adds another layer of relevance to that hypothesized cosmic event: Scientists believe that without that other body, the basic conditions for life to emerge on Earth might never have appeared.”

Artificial Intelligence

OpenAI’s Models Are Getting Too Smart For Their Human TeachersStephanie Palazzolo | The Information

“In the fight to improve AI models, Anthropic and OpenAI have doubled down on two methods: letting models train on fake clones of apps—otherwise known as reinforcement learning environments or gyms—and getting experts in various fields to teach models new things, as I reported in this Tuesday story. One problem is emerging, though: It’s getting a lot harder for human experts to stump the models and expand their knowledge in certain fields.”

Artificial Intelligence

AI Agents Are Getting Ready to Handle Your Whole Financial LifeJack Pitcher | The Wall Street Journal

“Tech and finance executives—well aware of how Wall Street is constantly reinventing itself with technology—are investing billions in new AI tools. And most agree: It’s a matter of when, not if, agentic AI can act on a person’s behalf across numerous accounts.”

Future

Scientists Gather to Confront the Doomsday Risks of ‘Mirror Life’Ellyn Lapointe | Gizmodo

“‘Pretty much everybody agrees’ that mirror-image cells would be ‘a bad thing,’ John Glass, a synthetic biologist at the J. Craig Venter Institute, told Nature. At the same time, some scientists argue that mirror-life research offers potential benefits that shouldn’t be ignored. The question is: How should experts regulate such research to maximize those benefits while minimizing risk?”

Science

Science’s Answer to the Ultimate Question: Where Do We Come From?Ethan Siegel | Big Think

“For countless generations, these questions powered the thoughts of poets, philosophers, and theologians, but today, we have meaningful answers provided by our scientific endeavors. Although there are still gaps in what we know, and plenty of room for surprises, we’ve come incredibly far in pursuit of the answers to the ultimate questions. Here’s where we are today.”

The post This Week’s Awesome Tech Stories From Around the Web (Through September 20) appeared first on SingularityHub.

科学家表示，一种新发现的免疫细胞可能在我们衰老过程中引发炎症。 || Scientists Say a Newly Discovered Immune Cell May Drive Inflammation as We Age

2025-09-19T14:00:00+00:00

免疫细胞家族成员之间的斗争暗示了慢性炎症为何会随着年龄增长而发生。

巨噬细胞——也就是“大吞噬细胞”——正如它们的名字所暗示的那样。这些免疫细胞在我们的身体内游走，寻找感染、癌症或损伤。当它们检测到目标时，会释放一系列信号分子，招募其他免疫细胞来发起反应。

最近，科学家们发现了一个较为静止的家族分支。这些巨噬细胞隐藏在包裹我们器官的脂肪组织的各个角落，与游走的同族有所不同。除了它们的类型多样且似乎随年龄增长而减少外，人们对它们知之甚少。科学家们曾假设它们可能有助于平衡衰老过程中的慢性炎症，但缺乏更多信息，这一假设难以验证。

现在，来自耶鲁大学医学院的团队 painstakingly 绘制了家族树，将这些居住在脂肪中的巨噬细胞分成了13种类型。一些群体已经为人所知，例如与神经相关的巨噬细胞，正如其名称所示，它们靠近并连接神经。

另一组则是科学界完全新发现的。这些细胞在年轻小鼠中几乎不存在，只在年老的小鼠中出现，并且似乎会促进炎症。

“我们并未预料到会存在一个完全新的细胞类型，”研究作者Vishwa Deep Dixit 告诉 Nature。

这个新成员似乎与神经相关的巨噬细胞进行一场拉锯战：前者促进炎症，后者则抑制炎症。然而，随着年龄增长，神经相关的巨噬细胞数量减少，它们对抗炎症的能力也随之减弱。

研究结果暗示了一个诱人的想法：也许重新平衡这两种细胞类型可以对抗随着人们年龄增长而出现的慢性炎症。

巨噬细胞家族

脂肪组织中的巨噬细胞是一群奇特的细胞。与循环中的同源细胞不同，这些细胞并不经常检测或对抗感染（尽管它们可以）。它们的主要职责是清洁工作。它们吞噬死亡的脂肪细胞，重塑脂肪组织，并调节新陈代谢。

这些细胞喜欢在白色脂肪的特殊微环境中定居。一些靠近含有死亡脂肪细胞的冠状结构，以便找到下一顿美餐。另一些则依偎在血管周围，清除死亡细胞。还有另一些——与神经相关的巨噬细胞——附着于神经，并调节肠道运动，触发针对流感病毒的免疫反应，并再生皮肤中的神经。

尽管这些细胞有多种有益的功能，但人们对它们在衰老过程中各自的变化仍知之较少。为了查明这一点，新研究的作者从年轻成人和老年人类等同年龄的小鼠中分离出脂肪巨噬细胞。他们分析了每种细胞的RNA景观。RNA不仅反映了基因活动，还反映了细胞的整体健康状况。

出现了13种RNA特征，每种都有其独特的特点。但其中两种尤为引人注目。

其中之一与神经相关的巨噬细胞有关。研究发现，这些细胞的数量和RNA特征随年龄变化显著，但具体变化在雌性和雄性小鼠之间有所不同。尽管这些细胞在年轻雌性小鼠中比雄性小鼠更丰富，但它们在老年雌性小鼠中数量骤减。那些残留的细胞变得更具炎症性，表现出与细胞衰老相关的基因，这是一种细胞分解并泄漏有毒促炎物质的过程，损害周围组织。其中一些甚至携带与晚发性散发性阿尔茨海默病相关的基因。

然后是这些新发现的细胞。它们表现出一种前所未见的独特特征。在年轻小鼠中，它们仅占脂肪巨噬细胞总数的1%。但在老年小鼠中，这一比例上升至20%。进一步分析它们的RNA特征，发现有衰老的迹象，以及与细胞因子——一种免疫分子——相关的促炎基因，以及其他已知会加剧炎症的蛋白质。

这些细胞出现的位置和方式仍然是个谜。尽管需要进一步研究，但该团队认为，随着年龄增长，这些细胞的积累可能会导致系统功能的下降。

双重功能

团队接下来在显微镜下更仔细地观察了年轻小鼠中与神经相关的巨噬细胞。这些细胞靠近白色脂肪组织中的神经，并延伸出“手臂”环绕神经。巨噬细胞表达了多种编码与生长和愈合相关的营养蛋白的基因，似乎清理了旧的和功能失调的髓鞘——一种包裹神经分支的脂肪“外鞘”，对正常脑细胞功能至关重要。

这种活动表明它们可能有助于维持神经细胞或支持其生存，团队写道。随着这些细胞数量减少并变得更具炎症性，保护作用很可能随之消失。

与神经相关的巨噬细胞还参与脂肪代谢。基因缺失这些细胞的年轻小鼠表现出血液中炎症增加，包括有毒细胞因子水平升高以及脂肪组织代谢紊乱，导致脂肪堆积，这与衰老有关。这些细胞似乎处于炎症、神经健康和脂肪代谢的交汇点——所有这些都会随着年龄增长而失衡。

“这些数据共同表明，衰老过程中[与神经相关的巨噬细胞]的减少会改变VAT中的免疫细胞景观……并提示[与神经相关的巨噬细胞]的丧失可能会加剧年龄相关的炎症，”团队写道。

虽然这项研究揭示了这些细胞，但新发现的巨噬细胞类型的作用仍不清楚。它们来自哪里？是什么导致它们随年龄增长而数量增加？在老年群体中清除它们是否能减少慢性炎症？团队发现了一种能够识别这些细胞的蛋白质标记，并希望它能帮助他们在未来研究中进一步阐明这些细胞在慢性炎症中的作用。

文章科学家称，一种新发现的免疫细胞可能随着年龄增长引发炎症首次出现在 SingularityHub。

---------------

A battle between immune cell family members hints at why chronic inflammation occurs with age.

Macrophages—or “big eaters”—are exactly what they sound like. These immune cells roam our bodies hunting down infections, cancers, or injuries. When they detect a target, the cells release a calvary of signaling molecules that recruit other immune cells to mount a response.

More recently, scientists have discovered a sedentary offshoot of the family. Hidden in fthe nooks and crannies of the fatty tissue enveloping our organs, these macrophages differ from their roaming cousins. Little is known about them, other than the fact they come in a variety of types and seem to decline with age. Scientists have hypothesized they might help balance chronic inflammation during aging, but without knowing more, the idea is hard to test.

Now, a team from the Yale School of Medicine has painstakingly charted the family tree, sorting these fat-dwelling macrophages into 13 types. Several populations were already known, such as nerve-associated macrophages that, true to their name, hover near and connect to nerves.

Another group was completely new to science. These cells were virtually absent in young mice and only appeared in older ones, and they appeared to contribute to inflammation.

“We did not anticipate that there would be a completely new cell type,” study author Vishwa Deep Dixit told Nature.

The newcomer seems to engage in a tug-of-war with nerve-associated macrophages: The former increases inflammation and the latter douses the fire. With age, however, nerve-associated macrophages dwindle—and so too does their fight against inflammation.

The results suggest a tantalizing idea: Perhaps re-balancing the two cell types could combat chronic inflammation that occurs as people grow older.

The Macrophage Family

Fatty tissue macrophages are an odd bunch. Unlike their circulating counterparts, these cells don’t often detect or fight infection (although they can). Rather, their main job is housekeeping. They eat dead fat cells, reshape fatty tissues, and regulate metabolism.

The cells like to nestle in special niches inside white fat. Some hover near crown-like structures that contain dying fat cells, where they can find their next meal. Others cuddle up to blood vessels and scrub away dead cells. Still others—nerve-associated macrophages—latch onto nerves and regulate movement in the gut, trigger immune responses against flu viruses in the lungs, and regenerate nerves in the skin.

Despite having a range of beneficial occupations, little is known about what happens to each type as we age. To find out, the authors of the new study isolated fat macrophages from male and female mice aged the rough the equivalent of young adults and elderly humans. They analyzed the RNA landscape of each cell. RNA reflects gene activity as well as the cell’s overall health.

Thirteen RNA signatures popped up, each with its own quirks. But two were especially eye-catching.

One of these was linked to nerve-associated macrophages. The team found their numbers and RNA profile changed vastly with age, though the specifics differed between female and male mice. Although the cells are more abundant in young female mice compared to males, their numbers tanked in elderly females only. Those cells that remained became more inflammatory. They showed an increase in genes associated with senescence, a process where cells break down and leak a toxic pro-inflammatory soup that damages surrounding tissues. Some even harbored genes associated with late-onset sporadic Alzheimer’s disease.

Then there were the newcomers. These exhibited a unique signature not seen before. They amounted to only one percent of the fatty macrophage population in young mice. But that number shot up to 20 percent in elderly mice. A closer look at their RNA profiles found signs of senescence, along with pro-inflammatory genes related to cytokines—a type of immune molecule—and other proteins known to drive up inflammation.

Where and how these cells pop up is still a mystery. Though more study is needed, the team suggests the accumulation of the cells as we age could contribute to systems breaking down.

Double Duty

The team next took a closer look at nerve-associated macrophages in young mice under the microscope. The cells snuggled up to nerves in white fat tissue and extended “arms” that encircled the nerves. The macrophages expressed multiple genes encoding a nutritious protein related to growth and healing and seemingly cleaned up old and dysfunctional myelin—a fatty “wrapper” encapsulating neuron branches that’s critical for normal brain cell function.

This activity suggests they might help maintain nerve cells or support their survival, wrote the team. As the cells dwindle and become pro-inflammatory with age, the protective effect would likely go away.

Nerve-associated macrophages also have a hand in fat metabolism. Young mice genetically deprived of these cells showed increased inflammation in their blood, including higher levels of toxic cytokines and disrupted metabolism of fatty tissues, resulting in fat accumulation, which has been associated with aging. The cells seem to sit at the crossroads of inflammation, nerve health, and fat metabolism—all of which get out of whack as we age.

“Together, these data demonstrate that a reduction in [nerve-associated macrophages] during aging alters the immune cell landscape in VAT… and suggest that loss of [nerve-associated macrophages] may potentiate age-induced inflammation,” wrote the team.

While the study sheds light on these cells, the role of the newly discovered macrophage type is unclear. Where do they come from? What causes them to grow in numbers with age? Could wiping them out in the elderly reduce chronic inflammation? The team found a protein marker that identifies these cells and hopes it can help them decipher the cells’ role in chronic inflammation in future studies.

The post Scientists Say a Newly Discovered Immune Cell May Drive Inflammation as We Age appeared first on SingularityHub.

为什么 OpenAI 针对 AI 幻觉的解决方案会让 ChatGPT 明天失效 || Why OpenAI’s Solution to AI Hallucinations Would Kill ChatGPT Tomorrow

2025-09-18T21:34:24+00:00

推动消费者AI发展的商业激励措施在根本上与减少幻觉的目标不一致。

OpenAI最新研究论文详细解释了为什么ChatGPT和其他大型语言模型会编造信息——在人工智能领域，这种现象被称为“幻觉”。该论文还揭示了为什么这个问题可能无法解决，至少对于普通用户而言。

该论文提供了迄今为止最严谨的数学解释，说明为什么这些模型会自信地陈述错误信息。它表明，这些错误不仅仅是当前AI训练方式的不幸副作用，而是数学上不可避免的。

这个问题部分可以归因于用于训练AI的基础数据中的错误。但通过数学分析AI系统的学习过程，研究人员证明，即使使用完美的训练数据，这一问题依然存在。

语言模型在回答问题时通过预测句子中的每一个词（基于概率）来生成响应，这种机制自然会产生错误。研究人员实际上证明，生成句子的总错误率至少是AI在简单是/否问题上的错误率的两倍，因为错误会在多个预测中累积。

换句话说，幻觉率从根本上受到AI系统区分有效和无效回答能力的限制。由于在许多知识领域中，这种分类问题本身就很困难，因此幻觉变得不可避免。

结果还显示，模型在训练中看到的事实越少，被询问时就越可能产生幻觉。例如，对于著名人物的生日，如果20%的人物生日信息在训练数据中仅出现一次，那么基础模型至少有20%的概率在生日查询中给出错误答案。

正如预期的那样，当研究人员询问最先进的模型Adam Kalai（论文作者之一）的生日时，DeepSeek-V3在多次尝试中给出了三个不同的错误日期：“03-07”、“15-06”和“01-01”。正确的日期是在秋季，因此这些日期都相差甚远。

评估陷阱

更令人担忧的是，论文分析了为什么即使在训练后采取了努力（如在发布前向AI的响应提供大量人类反馈），幻觉仍然存在。作者们检查了10个主要的AI评估基准，包括谷歌、OpenAI使用的基准，以及排名AI模型的顶级排行榜。这揭示了其中九个基准使用二元评分系统，对AI表达不确定性的回答给予零分。

这导致了作者所说的“诚实回答被惩罚的流行病”。当AI系统说“我不知道”时，它得到的分数与给出完全错误信息的分数相同。在这种评估体系下，最优策略变得清晰：总是猜测。

研究人员用数学方法证明了这一点。无论某个答案正确的可能性如何，在使用二元评分的评估中，猜测的预期得分总是高于保持沉默的得分。

一种会颠覆一切的解决方案

OpenAI提出的解决方法是让AI在输出答案前考虑自身的置信度，并让基准测试根据这一置信度进行评分。例如，AI可以被提示：“只有在超过75%置信度时才回答，因为错误答案会被扣3分，而正确答案会得1分。”

OpenAI研究人员的数学框架表明，在适当的置信度阈值下，AI系统会自然地表达不确定性，而不是猜测。因此，这将导致更少的幻觉。问题在于这种做法对用户体验的影响。

假设ChatGPT开始对30%的查询说“我不知道”——这是基于论文对训练数据中事实不确定性的分析的保守估计。用户习惯于几乎任何问题都能得到自信的回答，他们可能会迅速放弃这样的系统。

我在另一个领域也见过类似的问题。我参与了犹他州盐湖城的一个空气质量监测项目。当系统在恶劣天气或设备校准期间对测量结果的不确定性发出警报时，用户参与度会比显示自信读数的界面更低——即使这些自信读数在验证时被证明是不准确的。

计算经济学问题

利用论文中的见解减少幻觉并不困难。衡量不确定性的方法已有几十年历史，这些方法可以用来提供可信的不确定性估计，并引导AI做出更明智的选择。

但即使解决了用户不喜欢不确定性的这个问题，还有一个更大的障碍：计算经济学。具有不确定性意识的语言模型需要比当前方法显著更多的计算资源，因为它们必须评估多个可能的回答并估计置信度。对于每天处理数百万查询的系统来说，这将导致运营成本大幅增加。

更复杂的解决方案，如主动学习，其中AI系统会提出澄清问题以减少不确定性，虽然可以提高准确性，但进一步增加了计算需求。这些方法在像芯片设计这样的专业领域表现良好，因为错误答案的成本高达数百万美元，从而证明了大量计算的合理性。但对于用户期望即时响应的消费者应用，这种计算成本变得不可行。

对于管理关键业务操作或经济基础设施的AI系统，计算经济学的考量则截然不同。当AI代理处理供应链物流、金融交易或医疗诊断时，幻觉的成本远远超过让模型决定是否足够确定的费用。在这些领域，论文提出的解决方案变得经济可行——甚至必要。不确定的AI代理将不得不付出更高的成本。

然而，消费者应用仍然主导着AI开发的优先事项。用户希望系统能够对任何问题都给出自信的答案。评估基准奖励那些猜测而非表达不确定性的系统。计算成本倾向于支持快速、自信的响应，而不是缓慢、不确定的响应。

每token的能源成本下降和芯片架构的进步可能会最终使AI决定是否足够确定来回答问题变得更加经济。但与当前猜测方法相比，所需的相对计算量仍然较高，无论绝对硬件成本如何。

简而言之，OpenAI的论文无意中揭示了一个令人不安的事实：推动消费者AI发展的商业激励措施在根本上与减少幻觉的目标不一致。在这些激励措施改变之前，幻觉将继续存在。

本文转载自The Conversation，根据知识共享许可协议发布。阅读原始文章。

文章《为什么OpenAI的AI幻觉解决方案会明天摧毁ChatGPT》首次发表于SingularityHub。

---------------

The business incentives driving consumer AI development remain fundamentally misaligned with reducing hallucinations.

OpenAI’s latest research paper diagnoses exactly why ChatGPT and other large language models can make things up—known in the world of artificial intelligence as “hallucination.” It also reveals why the problem may be unfixable, at least as far as consumers are concerned.

The paper provides the most rigorous mathematical explanation yet for why these models confidently state falsehoods. It demonstrates that these aren’t just an unfortunate side effect of the way that AIs are currently trained, but are mathematically inevitable.

The issue can partly be explained by mistakes in the underlying data used to train the AIs. But using mathematical analysis of how AI systems learn, the researchers prove that even with perfect training data, the problem still exists.

The way language models respond to queries—by predicting one word at a time in a sentence, based on probabilities—naturally produces errors. The researchers in fact show that the total error rate for generating sentences is at least twice as high as the error rate the same AI would have on a simple yes/no question, because mistakes can accumulate over multiple predictions.

In other words, hallucination rates are fundamentally bounded by how well AI systems can distinguish valid from invalid responses. Since this classification problem is inherently difficult for many areas of knowledge, hallucinations become unavoidable.

It also turns out that the less a model sees a fact during training, the more likely it is to hallucinate when asked about it. With birthdays of notable figures, for instance, it was found that if 20 percent of such people’s birthdays only appear once in training data, then base models should get at least 20 percent of birthday queries wrong.

Sure enough, when researchers asked state-of-the-art models for the birthday of Adam Kalai, one of the paper’s authors, DeepSeek-V3 confidently provided three different incorrect dates across separate attempts: “03-07”, “15-06”, and “01-01”. The correct date is in the autumn, so none of these were even close.

The Evaluation Trap

More troubling is the paper’s analysis of why hallucinations persist despite post-training efforts (such as providing extensive human feedback to an AI’s responses before it is released to the public). The authors examined 10 major AI benchmarks, including those used by Google, OpenAI, and also the top leaderboards that rank AI models. This revealed that nine benchmarks use binary grading systems that award zero points for AIs expressing uncertainty.

This creates what the authors term an “epidemic” of penalizing honest responses. When an AI system says “I don’t know,” it receives the same score as giving completely wrong information. The optimal strategy under such evaluation becomes clear: Always guess.

The researchers prove this mathematically. Whatever the chances of a particular answer being right, the expected score of guessing always exceeds the score of abstaining when an evaluation uses binary grading.

The Solution That Would Break Everything

OpenAI’s proposed fix is to have the AI consider its own confidence in an answer before putting it out there and for benchmarks to score them on that basis. The AI could then be prompted, for instance: “Answer only if you are more than 75 percent confident, since mistakes are penalized 3 points while correct answers receive 1 point.”

The OpenAI researchers’ mathematical framework shows that under appropriate confidence thresholds, AI systems would naturally express uncertainty rather than guess. So this would lead to fewer hallucinations. The problem is what it would do to user experience.

Consider the implications if ChatGPT started saying “I don’t know” to even 30 percent of queries—a conservative estimate based on the paper’s analysis of factual uncertainty in training data. Users accustomed to receiving confident answers to virtually any question would likely abandon such systems rapidly.

I’ve seen this kind of problem in another area of my life. I’m involved in an air-quality monitoring project in Salt Lake City, Utah. When the system flags uncertainties around measurements during adverse weather conditions or when equipment is being calibrated, there’s less user engagement compared to displays showing confident readings—even when those confident readings prove inaccurate during validation.

The Computational Economics Problem

It wouldn’t be difficult to reduce hallucinations using the paper’s insights. Established methods for quantifying uncertainty have existed for decades. These could be used to provide trustworthy estimates of uncertainty and guide an AI to make smarter choices.

But even if the problem of users disliking this uncertainty could be overcome, there’s a bigger obstacle: computational economics. Uncertainty-aware language models require significantly more computation than today’s approach, as they must evaluate multiple possible responses and estimate confidence levels. For a system processing millions of queries daily, this translates to dramatically higher operational costs.

More sophisticated approaches like active learning, where AI systems ask clarifying questions to reduce uncertainty, can improve accuracy but further multiply computational requirements. Such methods work well in specialized domains like chip design, where wrong answers cost millions of dollars and justify extensive computation. For consumer applications where users expect instant responses, the economics become prohibitive.

The calculus shifts dramatically for AI systems managing critical business operations or economic infrastructure. When AI agents handle supply chain logistics, financial trading, or medical diagnostics, the cost of hallucinations far exceeds the expense of getting models to decide whether they’re too uncertain. In these domains, the paper’s proposed solutions become economically viable—even necessary. Uncertain AI agents will just have to cost more.

However, consumer applications still dominate AI development priorities. Users want systems that provide confident answers to any question. Evaluation benchmarks reward systems that guess rather than express uncertainty. Computational costs favor fast, overconfident responses over slow, uncertain ones.

Falling energy costs per token and advancing chip architectures may eventually make it more affordable to have AIs decide whether they’re certain enough to answer a question. But the relatively high amount of computation required compared to today’s guessing would remain, regardless of absolute hardware costs.

In short, the OpenAI paper inadvertently highlights an uncomfortable truth: the business incentives driving consumer AI development remain fundamentally misaligned with reducing hallucinations. Until these incentives change, hallucinations will persist.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post Why OpenAI’s Solution to AI Hallucinations Would Kill ChatGPT Tomorrow appeared first on SingularityHub.

这种人工智能发现了昏迷患者即将苏醒的迹象，比医生早几天。 || This AI Found Signs Comatose Patients Were Waking Up Days Before Doctors Did

2025-09-16T21:47:23+00:00

该算法在极其细微的面部运动中发现了意识的早期信号。

想象一下你在医院的病房醒来。你最后的记忆是一场可怕的车祸。医生握着你的手并让你捏它。你尽全力尝试，但什么也没发生——甚至没有一丝抽搐。

“我恐怕他处于昏迷状态，”你听到医生说。但我想我有意识，你想要喊出来。

遭受创伤性头部损伤的人，通常由车祸引起，可能看起来完全无法对外界做出反应。但许多人会经历“隐性意识”——他们的大脑对外界指令有反应，即使他们无法通过眨眼、手指抽动或其他明显的动作让临床医生和护士察觉。

虽然脑成像技术有时可以捕捉到一个人内部清醒并试图遵循指令的迹象，但这些方法在日常监测中成本高昂且难以使用，当医生和患者家属等待他们醒来时，这些方法并不实用。

“一些严重脑损伤的患者可能看起来无意识，但仍有部分意识和运动能力，”斯蒂尼·莫法哈姆及其同事在一项新研究中写道。只是“这些运动往往太微小，无法在常规检查中被医生察觉。”

这项研究由莫法哈姆领导，使用计算机视觉技术追踪看似无意识患者中的微小面部运动。该AI工具被命名为SeeMe，专注于极其细微的运动，甚至精确到皮肤上的单个毛孔。

与医生相比，该工具在大约90%的患者中检测到隐性意识的早期迹象，时间大约提前了四天。研究还发现，这些微小抽动的数量和强度与患者出院时的恢复程度相关。

意识的早期检测可以让刚苏醒的人的康复过程更不痛苦。知道患者有意识可以帮助医生决定何时开始康复治疗，从而带来更好的健康结果。这项技术未来甚至可能用于实时监测脑损伤的治疗，如中风或其他伤害。

意识的阶梯

我们通常将意识视为一个开关。打开它，你就能感知外界和自身；关闭它，意识就变得黑暗。

但意识更像是一个调光开关。脑部受到撞击后，人们可能进入最小意识状态。在此状态下，他们有间歇性的意识，并能遵循指令，比如医生说“向左看”或“捏我的手”。更严重的是植物状态。患者在此状态下会周期性地睁眼或闭眼，但无法再对外界刺激做出反应。

在特别严重的创伤中，患者会进入昏迷状态，此时他们既不感知自身也不感知他人，无法行动，也无法被唤醒。

尽管如此，无反应的人可能会恢复意识——往往比他们可观察到的行为更快。在一项研究中，一名处于植物状态的患者被要求想象打网球或在家中走动时，显示出相关的大脑活动，尽管她无法进行身体上的反应。

最近，一项具有里程碑意义的大脑成像研究发现，在353名被判定为无意识的严重脑损伤患者中，至少有四分之一显示出基于大脑活动的意识迹象，当给予语音指令时。大多数患者没有通过标准的临床测试来检测反应性。

但脑成像测试虽然强大，却成本高昂且不适用于日常临床使用。与研究团队不同，新研究的团队借鉴了临床医生的方法，通过将微小的面部运动与诊断和恢复联系起来，来监测意识。

现在你看到我

面部是大脑的窗口。它的肌肉由大脑两侧的广泛区域控制。任何意识恢复的早期迹象很可能首先出现在面部运动中。

研究团队招募了16名健康志愿者和37名脑损伤患者，这些患者表面上看起来处于昏迷状态。然后他们分析了参与者被要求完成三项任务时的视频记录：“伸出舌头”、“睁开眼睛”和“展示微笑”。

所选任务涉及多个面部区域和肌肉，以更好地评估大脑活动，作者写道。

新的AI工具SeeMe追踪了对指令的面部运动——精确到单个毛孔——并由一组训练有素的医疗专业人员审查视频，要求他们提供专业意见。

AI捕捉到了30名患者的眨眼反应和几乎所有的口部运动，成功率几乎是医生的两倍。SeeMe尤其对那些逃过人类眼睛的微小抽搐敏感。

该工具还标记了意识的早期迹象。在一名深陷昏迷的志愿者中，一位因车祸受伤的老人，AI在入院后的第18天检测到口部运动；他最终在第37天对运动指令做出反应。SeeMe还在另一名因交通事故陷入昏迷的参与者中发现了眼和口部运动，时间是在入院后的第19天。他三天后睁开了眼睛，并逐渐恢复。

总体而言，SeeMe在标准测试发现之前大约提前了四天检测到睁眼反应，而在口部运动反应上大约提前了八天。AI的表现还与患者出院和六个月后的恢复程度相关——即他们逐渐恢复意识并能够进行康复治疗。

SeeMe旨在补充而非替代长期随访和护理。研究团队写道，昏迷患者是“一个极其难以研究的人群”。有些人可能意识有波动，但未被研究捕捉到。其他人可能只是不想参与。

缺乏意识的早期检测“不应被解释为患者无法恢复意识的迹象”，作者解释道。

为了进一步优化AI，研究团队希望收集那些恢复意识但最初被SeeMe遗漏的人的信息。他们还计划纳入其他运动的客观测量，如肌肉中的电信号。SeeMe甚至可能帮助那些在研究覆盖时间之外被认为无意识的人。

对于患者和家属而言，进一步的研究可能会带来一个基于面部运动的“是或否”系统，使亲人能够“再次交流”。

该文章 This AI Found Signs Comatose Patients Were Waking Up Days Before Doctors Did 首次出现在 SingularityHub。

---------------

The algorithm discovered early signals of consciousness in extremely subtle facial movements.

Imagine waking up in a hospital room. The last thing you remember is a terrible car crash. A doctor holds your hand and asks you to squeeze it. You try as hard as you can, but nothing happens—not even a twitch.

“I’m afraid he’s in a coma,” you hear the doctor say. But I’m conscious, you want to yell.

People with traumatic head injuries, often resulting from car accidents, can seem completely unresponsive to the outside world. But many experience “covert consciousness”—in that their brains respond to commands—even though they can’t translate it into eye blinks, finger twitches, or other obvious movements for clinicians and nurses to detect.

Although brain imaging techniques can sometimes capture signs a person is internally awake and trying to follow instructions, these methods are expensive and difficult to use for everyday monitoring while doctors and the patients’ families wait for them to wake up.

“Some people with severe brain injuries may appear unconscious, but still have some awareness and the ability to move,” wrote Sima Mofakham and colleagues at Stony Brook University in a new study. It’s just that “these movements are often too small to be seen by doctors during routine exams.”

The study, led by Mofakham, used computer vision to track tiny facial movements in seemingly unconscious patients. The AI tool, dubbed SeeMe, zeroed in on extremely minute movements, down to the level of single pores in the skin.

Compared to physicians, the tool detected early signs of covert consciousness roughly four days earlier in roughly 90 percent of patients. The study also found the number and strength of these tiny twitches corresponded to how well the patient had recovered at the time of discharge.

Early detection of consciousness could make recovery less distressing for a person who’s just waking up. Knowing the person is aware could help doctors decide when to kickstart rehabilitation associated with better health outcomes. The technology may also one day be used to monitor real-time treatments for brain damage due to stroke and other injuries.

Stairway to Consciousness

We often think of consciousness as a light switch. Flip it on, and you’re aware of both the outside world and yourself; flip it off, and awareness goes dark.

But consciousness is more like a light dimmer. After a blow to the brain, people can fall into a minimally conscious state. Here, they experience intermittent awareness and can follow commands, like if a doctor says “look left” or “squeeze my hand.” More severe is the vegetative state. Patients in this state can open or close their eyes in cycles, but they can no longer respond to outside stimulation.

In especially traumatic injuries, the patient goes into a coma, where they’re not aware of themselves and others, can’t move, and can’t be awakened.

Despite the odds, unresponsive people can recover mental awareness—often sooner than their observable behavior would suggest. In one study, a person in a vegetative state showed relevant brain activity when asked to imagine playing tennis or moving around her house, even though she couldn’t physically respond.

More recently, a landmark brain imaging study found at least a quarter of 353 people with severe brain injuries, who had been deemed unconscious, showed signs of awareness based on brain activity when given voice commands. Most did not react to a battery of standard clinical tests for responsiveness.

But brain imaging tests, while powerful, are expensive and impractical for everyday clinical use. Rather than looking into the brain, the team behind the new study took a page out of the clinician’s playbook by linking tiny facial movements to diagnostics and recovery.

Now You See Me

The face is a window on the brain. Its muscles are controlled by large areas across both of the brain’s hemispheres. Any early signs of recovery are likely to show up first in facial movements.

The team recruited 16 healthy volunteers and 37 people with brain injuries who, outwardly, appeared to be in a coma. They then analyzed video recordings of the participants being asked to do three tasks: “Stick out your tongue,” “open your eyes,” and “show me a smile.”

The tasks chosen involved multiple facial regions and muscles to better gauge brain activity, the authors wrote.

The new AI tool, SeeMe, then tracked facial movements—down to the level of individual pores—in response to the commands. A group of trained medical professionals also reviewed the videos and were asked for their expert opinions.

The AI captured eye responses in 30 patients and nearly all of their mouth movements, with a success rate nearly double that of the physicians. SeeMe was especially sensitive to tiny twitches that evaded the human eye.

The tool also flagged earlier signs of consciousness. In one deeply comatose volunteer, an older man who suffered a car crash, the AI detected mouth movements on day 18 after admission; he finally responded to motor commands on day 37. SeeMe also found signs of eye and mouth movements 19 days after admission in another participant in a coma after a traffic accident. He opened his eyes three days later and went on to gradually recover.

Across the board, SeeMe detected eye-opening responses roughly four days before standard tests picked them up and mouth-related reactions about eight days earlier. The AI’s performance also correlated to how well patients recovered on discharge and at six months—that is, they increasingly regained awareness and could do rehab.

SeeMe is intended to complement, not replace, long-term follow-up and care. Comatose patients are “an exceedingly challenging population to study,” wrote the team. Some people may have had fluctuations in awareness that weren’t captured in the study. Others may simply not have wanted to participate.

A lack of early detection of consciousness “should never be interpreted as the absence of potential” that the patient can regain awareness, the authors explained.

To further fine-tune the AI, the team hopes to gather information on people who regained consciousness but were initially missed by SeeMe. They also aim to incorporate other objective measures of movement, such as electrical signals in muscles. SeeMe could even help people presumed unconscious for longer periods of time than that covered in the study.

For patients and families, further work could result in a “yes or no” system based on facial movements that might allow loved ones to “talk” to each other again.

The post This AI Found Signs Comatose Patients Were Waking Up Days Before Doctors Did appeared first on SingularityHub.

科学家希望3D打印皮肤能带来按需治疗，以应对严重伤害。 || Scientists Hope 3D-Printed Skin Can Bring On-Demand Treatment for Serious Injuries

2025-09-15T16:24:10+00:00

新的生物打印技术使3D打印具有复杂血管网络的皮肤成为可能。

生物打印有望实现按需生产组织和器官，但以往的努力受到我们无法创建支持其生存的血管网络的限制。现在，两种互补的新技术可能为高级皮肤移植解决这一问题。

皮肤可能是人体最被低估的器官之一。它不仅提供对抗细菌、毒素和辐射的关键屏障，还帮助调节体温和水分流失，并作为重要的感觉器官，介导我们的触觉和痛觉。

特别是对于严重的皮肤损伤，如烧伤，通常通过从身体其他部位移植一层薄薄的表皮来治疗。但许多支持皮肤关键功能的结构，如血管、神经和毛囊，实际上位于表皮下方的真皮层。

通常无法移植真皮层，因为这会留下与正在治疗的伤口同样严重的创伤。因此，传统的皮肤移植通常无法恢复全部功能，还可能导致严重疤痕。

现在，来自瑞典林雪平大学的研究人员开发了两种新的生物打印技术——本质上是使用生物材料进行3D打印——可以生产出具有血管网络的皮肤移植，其结构能复制真皮层的复杂性。第一种方法涉及将含有细胞的凝胶注入伤口，然后生长为功能性组织。第二种方法使用水凝胶线在组织中创建通道，这些通道可以成为血管。

“真皮结构如此复杂，我们无法在实验室中生长它。我们甚至不知道它所有组成成分。这就是为什么我们和许多其他研究者认为，我们可以移植基本构建模块，然后让身体自行生成真皮，”林雪平大学的Johan Junker在一项研究中表示，该研究的新闻稿发布于此处。

研究人员首先开发了一种特别设计的“生物墨水”，其中含有称为成纤维细胞的细胞。这些是真皮层中最常见的细胞，能够产生重要的真皮成分，如胶原蛋白、弹性蛋白和透明质酸，据《连线》报道。

研究人员将这些细胞生长在明胶小珠上，然后与透明质酸混合制成凝胶。压力使凝胶转化为液体，可通过3D打印机喷嘴挤出，然后再变回凝胶状。

研究人员使用他们称之为“注射器中的皮肤”的发明，创建了小型圆盘，并将其移植到小鼠的皮肤下。在发表于《先进医疗材料》的成果中，研究人员报告称，活细胞产生了多种对生成新真皮至关重要的物质，如胶原蛋白，并且新的血管甚至在移植中生长出来。

The post Scientists Hope 3D-Printed Skin Can Bring On-Demand Treatment for Serious Injuries appeared first on SingularityHub.

能够生长新的血管将是制造可用皮肤移植的关键。在组织工程研究中，创建功能性血管网络一直是一个长期存在的挑战，因为没有血管网络，就无法将营养和氧气输送到更大的、更复杂的结构中。

研究人员开发的第二种技术可能进一步解决这一问题。在另一篇发表于同一期刊的论文中，他们展示了如何将水凝胶线打印到组织中。

这些线可以被排列成复杂的图案，然后通过简单的酶的作用溶解，据《连线》报道，留下一个管状腔室，其中可以生长新的血管。通过结合这两种技术，最终应该能够创建一个完全功能的人工真皮。

如往常一样，将这些技术从实验室推广到实际应用将是一条漫长而不确定的道路。但它们使我们离为严重皮肤损伤提供按需治疗更近了一步。

The post Scientists Hope 3D-Printed Skin Can Bring On-Demand Treatment for Serious Injuries appeared first on SingularityHub.

---------------

New bioprinting techniques make it possible to 3D print skin with complex networks of blood vessels.

Bioprinting holds the promise of producing tissues and organs on demand, but efforts have been held back by our inability to create the networks of blood vessels required to sustain them. Two complementary new technologies could now solve the problem for advanced skin grafts.

The skin is probably one of the body’s most underappreciated organs. Not only does it provide a crucial barrier against germs, toxins, and radiation, but it also helps regulate temperature and water loss and acts as a vital sensory organ mediating our sense of touch and pain.

Serious injuries to the skin, in particular burns, are usually treated by transplanting a thin layer of epidermis, the top layer of skin, from elsewhere on the body. But many of the structures supporting the skin’s critical functions, such as blood vessels, nerves, and hair follicles, are actually found in the layer below, known as the dermis.

It’s usually impossible to transplant the dermis because it would leave behind a wound as severe as the one being treated. So, traditional skin grafts normally don’t restore full function and can lead to severe scarring.

Now, researchers from Linköping University in Sweden have developed two new bioprinting techniques—essentially 3D printing with biological materials—that could produce skin grafts perfused with blood vessels that replicate the complex structure of the dermis. The first approach involves injecting a cell-laden gel into a wound that can then grow into functional tissue. The second uses hydrogel threads to create channels that can become blood vessels.

“The dermis is so complicated that we can’t grow it in a lab. We don’t even know what all its components are. That’s why we, and many others, think that we could possibly transplant the building blocks and then let the body make the dermis itself,” Johan Junker at Linköping University, who led the study, said in a press release.

The researchers first developed a specially designed “bioink” containing cells known as fibroblasts. These are the most common cells in the dermis and produce important dermal ingredients such as collagen, elastin, and hyaluronic acid, according to Wired.

The researchers grew these cells on tiny beads of gelatin and then mixed them with hyaluronic acid to create a gel. Pressure turns the gel into a liquid that can be extruded through the nozzle of a 3D printer before becoming gel-like again.

The researchers used this “skin in a syringe,” as they’ve dubbed their invention, to create small disks that they then transplanted under the skin of mice. In results published in Advanced Healthcare Materials, the researchers reported the living cells produced various substances crucial for growing a new dermis, such as collagen, and new blood vessels even grew in the graft.

The ability to grow new blood vessels will be key if we’re to make usable skin grafts. Creating functional vascular networks has been a long-standing challenge in tissue engineering efforts, as without them it’s impossible to deliver nutrients and oxygen into larger, more complex structures.

The second technique the researchers developed could go even further towards solving this problem. In another paper, published in the same journal, they showed they could print threads of a water-based substance known as a hydrogel into tissues.

These threads can be arranged in complex patterns and then dissolved by the application of a simple enzyme, according to Wired, leaving a tube-like cavity in which new blood vessels could be grown. By combining the two technologies it should eventually be possible to create a fully functional artificial dermis.

As ever, getting these technologies out of the lab will be a long and uncertain journey. But they bring us one step closer to an on-demand treatment for the most serious skin injuries.

The post Scientists Hope 3D-Printed Skin Can Bring On-Demand Treatment for Serious Injuries appeared first on SingularityHub.

本周网络上的精彩科技故事（截至9月13日） || This Week’s Awesome Tech Stories From Around the Web (Through September 13)

2025-09-13T14:00:00+00:00

Future

AI Could Make the Smartphone Passé. What Comes Next?Brian X. Chen and Tripp Mickle | The New York Times

“Every major tech company is thinking about this million-dollar question: What comes after the smartphone? Here is a list of predictions from current and former employees of some of the world’s largest tech companies, including Apple, Google, Samsung, Amazon, and Meta.”

Computing

Good Old IBM Is Leading the Way in the Race for ‘Quantum Advantage’Christopher Mims | The Wall Street Journal

“IBM hasn’t been associated with breakthrough innovation since its Watson AI won ‘Jeopardy!’ in 2011. But quantum computing, which could see a breakthrough to commercialization by 2030, gives the 114-year-old stalwart of business computing a chance to reclaim some of its past glory.”

Robotics

Reality Is Ruining the Humanoid Robot HypeEvan Ackerman | IEEE Spectrum

“Future projections seem to be based on an extraordinarily broad interpretation of jobs that a capable, efficient, and safe humanoid robot—which does not currently exist—might conceivably be able to do. Can the current reality connect with the promised scale?”

Robotics

We Are Entering a Golden Age of Robotics Startups—and Not Just Because of AIRebecca Szkutak | TechCrunch

“Investors poured $6 billion into robotics startups in the first seven months of 2025 according to Crunchbase data. The data company predicts that this year’s funding totals will eclipse 2024, making it one of the only non-AI categories to experience a boost in funding.”

Future

New Pathway Engineered Into Plants Lets Them Suck Up More CO₂John Timmer | Ars Technica

“It would be nice to think that we could reforest our way out of the mess we’re creating, but recent studies have indicated there’s simply not enough productive land for this to work out. One alternative might be to get plants to take up carbon dioxide more efficiently.”

Space

Rendezvous Robotics Exits Stealth With $3M to Build Reconfigurable Space InfrastructureAria Alamalhodaei | TechCrunch

“Instead of astronauts and robotic arms, Rendezvous is betting on autonomous swarm assembly and electromagnetism. The company is commercializing a technology called ‘tesserae,’ flat-packed modular tiles that can launch in dense stacks and magnetically latch to form structures on orbit. With a software command, the tiles are designed to unlatch and rearrange themselves when the mission changes.”

Future

Will AI Choke Off the Supply of Knowledge?Greg Ip | The Wall Street Journal

“When humans answer questions, such as whether Einstein should be energy secretary, they often pursue novel avenues of inquiry, creating new knowledge and insight as they go. They do this for a variety of reasons: salary, wealth, fame, tenure, ‘likes,’ clicks, curiosity. If LLMs come to dominate the business of answering questions, those incentives shrivel.”

Energy

Geothermal Is Too Expensive, but Dig Energy’s Impossibly Small Drill Rig Might Fix ThatTim De Chant | TechCrunch

“The startup, which has been operating in stealth for the last five years, developed the water-jet drilling rig in an effort to make geothermal heating and cooling so inexpensive that it will displace fossil fuel boilers and furnaces. The rig is central to that, promising to slash drilling costs by up to 80%.”

Science

A Single, ‘Naked’ Black Hole Rewrites the History of the UniverseCharlie Wood | Quanta Magazine

“A black hole unlike any seen before has been spotted in the early universe. It’s huge and appears to be essentially on its own, with few stars circling it. The object, which may represent a whole new class of enormous ‘naked’ black holes, upends the textbook understanding of the young universe.”

Future

Pay-Per-Output? AI Firms Blindsided by Beefed Up Robots.txt Instructions.Ashley Belanger | Ars Technica

“Leading Internet companies and publishers—including Reddit, Yahoo, Quora, Medium, The Daily Beast, Fastly, and more—think there may finally be a solution to end AI crawlers hammering websites to scrape content without permission or compensation.”

Artificial Intelligence

The Software Engineers Paid to Fix Vibe Coded MessesEmanuel Maiberg | 404 Media

“The alleged benefit of vibe coding, which refers to the practice of building software with AI-coding tools without much attention to the underlying code, is that it allows anyone to build a piece of software very quickly and easily. …[But] if the resulting software is so poor you need to hire a human specialist software engineer to come in and rewrite the vibe coded software, it defeats the entire purpose.”

Biotechnology

Scientists Infuse Cement With Bacteria to Create Living Energy DeviceGayoung Lee | Gizmodo

“‘We envision this technology being integrated into real buildings, in walls, foundations, or bridges, where it can support renewable energy sources like solar panels by providing local energy storage,’ Luo said. ‘Imagine a regular room built with bacteria-infused cement: Even at a modest energy density of 5 Wh/kg, the walls alone could store about 10 kWh—enough to keep a standard enterprise server running for a whole day.'”

The post This Week’s Awesome Tech Stories From Around the Web (Through September 13) appeared first on SingularityHub.

---------------

我让AI模仿我并为我开设课程——这是我关于教育未来的发现 || I Got an AI to Impersonate Me and Teach Me My Own Course—Here’s What I Learned About the Future of Education

2025-09-12T14:00:00+00:00

我请了一个AI代理扮演我，一位牛津大学媒体与人工智能的讲师，为我开设了一门完全基于我自己的作品的个人硕士课程。

想象一下，如果你有一个无限的预算，可以雇佣个性化的私人导师，为学习者提供最大化生产力和技能发展的课程。今年夏天，我以一种荒谬且自我中心的方式预览了这个想法。

我请了一个AI导师代理扮演我，一位牛津大学媒体与人工智能的讲师，为我开设一门完全基于我自己的作品的个人硕士课程。

我通过一个现成的ChatGPT工具在基于Azure的 Nebula One 平台上设置该代理，提示它研究并模仿我，然后根据我的想法构建个性化材料。我没有告诉大型语言模型（LLM）需要阅读什么或做任何其他提升其能力的事情，比如提供它访问非公开在线学习材料的权限。

该代理的媒体与人工智能课程结构良好——一个为期一学期、原创的六模块课程，深入探讨我自己的著作集，这是我从未设计过的，但承认我可能会喜欢。

这门课程是互动且快速的，通过定期切换格式来考验思维敏捷性。它在智力上具有挑战性，就像好的牛津辅导课应该一样。代理以严谨的方式教学，对任何问题都能立即回应。它对人工智能和媒体的快速演变景观有深刻理解，但比我的研究更深入。

这似乎来源于我所有的多媒体作品——书籍、演讲、文章、媒体采访，甚至我从未意识到被录制、更不用说用于训练GPT-4或GPT-5的大学讲座。

这门课程是一次极佳的学习体验，尽管我本以为自己已经全部掌握。因此，在不可避免的学生调查中，我给予了这个代理版本的我应得的五星评价。

例如，在讨论计算机游戏中非玩家角色（NPCs）的伦理问题时，它提出了以下问题：

如果NPC由AI生成，谁决定它们的性格、背景或道德？这是否会导致偏见或刻板印象？

以及：

如果一个AI NPC能够学习和适应，是否模糊了角色与“实体”（独立行动者）之间的界限？

这些都是深刻而哲学性的问题，可能在《GTA 6》明年5月发布时成为焦点。我很兴奋这个代理版本的我提出了这些问题，即使真实的我并没有。

代理版本的我也在真实我已知的基础上进行了拓展。在电影领域，它了解了我所讲授的常规 Adobe After Effects，我曾介绍过（它用于制作动态图形和视觉效果）。但它还加入了 Nuke，这是一种用于《复仇者联盟》等影片中合成和操控视觉效果的专业工具，而我竟然从未听说过。

课程阅读书目

那么，代理是如何了解我的？我的出版商Routledge与Open AI签订了一项训练数据协议，这可能涵盖了我关于媒体、人工智能和现场体验的书籍。

与一些作者不同，我对此持开放态度。我的书籍引导人们了解一个令人惊叹且快速发展的领域，我希望它们在全球对话中被广泛使用，无论何种格式和地域（土耳其的版本已经出版，韩国本月出版）。

这种可获取性必须扩展到现在可能成为最易发现的“语言”——由AI模型使用的语言。任何认同这一点的作者都应优先考虑AI优化：使他们的作品更容易被LLMs发现、处理和使用——这类似于搜索引擎优化，但针对AI。

为了进一步测试这个想法，我让一个基于中国 DeepSeek 的代理运行一个关于我材料的课程。当我发现自己在它的训练语料库中不够显眼时，很难不感到被冒犯。在AI时代，没有比领先的LLM认为你的关于AI的书籍无关紧要更令人沮丧的事情了。

当我尝试其他AI时，它们在事实准确性上遇到了问题，这非常符合2024年的趋势。从谷歌的 Gemini 2.5 Pro，我了解到一些关于我的虚构传记细节，比如我担任媒体公司The Runaway Collective的职务。

当我询问埃隆·马斯克的 Grok 我的最佳引述时，它回答：“无论你的问题是什么，答案都是AI。”这是一句很棒的话，但不是我说的，而是谷歌DeepMind的诺贝尔奖得主Demis Hassabis 说的。

我们正在走向何方

整个自我中心的夏季实验显然是荒谬的，尽管并非完全不切实际。代理自我学习项目可能是大学教学真正需要的：互动、分析、有洞察力且个性化的。而且有一些新兴研究探讨其价值。这项德国主导的研究发现，AI生成的反馈有助于激励中学生并提升他们的考试复习效果。

不久的将来，我们可能会看到这种实时AI层正式融入学校和大学教学。任何教授本科生的人都会知道，AI已经存在。学生使用AI转录来记笔记，讲座内容几秒钟内就会被这些转录内容复制，并在一年内训练出十几个LLMs。对于Z世代的项目，ChatGPT、Claude、Gemini和DeepSeek/Qwen是 不可或缺的要素。

但这里有个关键点。随着AI在教育中的作用越来越核心，人类教师的重要性反而会增加，而不是减少。他们将引导学习体验，将已出版的作品引入课程的概念框架，并推动面对面的学生参与和鼓励。他们可以作为个人AI导师，通过代理为每个学生根据个人学习需求进行扩展。

对于那些没有个人作品集的年轻教师来说，他们该如何融入？年轻教师可能更具备AI原生性。他们可以利用AI来拓展自己对课程概念框架的设想，通过引导代理包含更多内容来扩大研究范围。

在AI领域，两种对立的观点往往同时成立。AI既是情感智能，又是缺乏同理心。它既是一个被夸大的文本预测器，又是一个高度创造性的伙伴。它既是导致失业，又是创造就业机会。它既是让我们变蠢，又在提升我们的能力。

同样，在教学领域。AI威胁着学习空间，但也能释放强大的互动。一种普遍观点认为，它会让学生变蠢。但也许AI实际上正在为学生解锁下一层个性化、挑战和动力。

本文由 The Conversation 在创意共享许可下重新发布。阅读原始文章。

文章我让AI扮演我并为我开设自己的课程——关于教育未来的所学首次出现在 SingularityHub。

---------------

I asked an AI agent to play the role of me, an Oxford lecturer on media and AI, and teach me a personal master’s course, based entirely on my own work.

Imagine you had an unlimited budget for individual tutors offering hyper-personalized courses that maximized learners’ productivity and skills development. This summer I previewed this idea—with a ridiculous and solipsistic test.

I asked an AI tutor agent to play the role of me, an Oxford lecturer on media and AI, and teach me a personal master’s course, based entirely on my own work.

I set up the agent via an off-the-shelf ChatGPT tool hosted on the Azure-based Nebula One platform, with a prompt to research and impersonate me, then build personalized material based on what I already think. I didn’t tell the large language model (LLM) what to read or do anything else to enhance its capabilities, such as giving it access to learning materials that aren’t publicly available online.

The agent’s course in media and AI was well structured—a term-long, original six-module journey into my own collected works that I had never devised, but admit I would have liked to.

It was interactive and rapid-fire, demanding mental acuity via regular switches in formats. It was intellectually challenging, like good Oxford tutorials should be. The agent taught with rigor, giving instant responses to anything I asked. It had a powerful understanding of the fast-evolving landscape of AI and media through the same lens as me, but had done more homework.

This was apparently fed by my entire multimedia output—books, speeches, articles, press interviews, even university lectures I had no idea had even been recorded, let alone used to train GPT-4 or GPT-5.

The course was a great learning experience, even though I supposedly knew it all already. So in the inevitable student survey, I gave the agentic version of me well-deserved, five-star feedback.

For instance, in a section discussing the ethics of non-player characters (NPCs) in computer games, it asked:

If NPCs are generated by AI, who decides their personalities, backgrounds, or morals? Could this lead to bias or stereotyping?

And:

If an AI NPC can learn and adapt, does it blur the line between character and “entity” [independent actor]?

These are great, philosophical questions, which will probably come to the fore when and if Grand Theft Auto 6 comes out next May. I’m psyched that the agentic me came up with them, even if the real me didn’t.

Agentic me also built on what real me does know. In film, it knew about bog-standard Adobe After Effects, which I had covered (it’s used for creating motion graphics and visual effects). But it added Nuke, a professional tool used to combine and manipulate visual effects in The Avengers, which (I’m embarrassed to say) I had never heard of.

The Course Reading List

So, where did the agent’s knowledge of me come from? My publisher Routledge did a training data deal with Open AI, which I guess could cover my books on media, AI, and live experience.

Unlike some authors, I’m up for that. My books guide people through an amazing and fast-moving subject, and I want them in the global conversation, in every format and territory possible (Turkish already out, Korean this month).

That availability has to extend to what is now potentially the most discoverable “language” of all, the one spoken by AI models. The priority for any writer who agrees with this should be AI optimization: making their work easy for LLMs to find, process, and use—much like search engine optimization, but for AI.

To build on this, I further tested my idea by getting an agent powered by China’s DeepSeek to run a course on my materials. When I found myself less visible in its training corpus, it was hard not to take offense. There is no greater diss in the age of AI than a leading LLM deeming your book about AI irrelevant.

When I experimented with other AIs, they had issues getting their facts straight, which is very 2024. From Google’s Gemini 2.5 Pro, I learned hallucinatory biographical details about myself like a role running media company The Runaway Collective.

When I asked Elon Musk’s Grok what my best quote was, it said: “Whatever your question, the answer is AI.” That’s a great line, but Google DeepMind’s Nobel-winning Demis Hassabis said it, not me.

Where We’re Heading

This whole, self-absorbed summer diversion was clearly absurd, though not entirely. Agentic self-learning projects are quite possibly what university teaching actually needs: Interactive, analytical, insightful, and personalized. And there is some emerging research around the value. This German-led study found that AI-generated feedback helped to motivate secondary school students and benefited their exam revision.

It won’t be long before we start to see this kind of real-time AI layer formally incorporated into school and university teaching. Anyone lecturing undergraduates will know that AI is already there. Students use AI transcription to take notes. Lecture content is ripped in seconds from these transcriptions and will have trained a dozen LLMs within the year. To assist with writing essays, ChatGPT, Claude, Gemini, and DeepSeek/Qwen are the sine qua non of Gen Z projects.

But here’s the kicker. As AI becomes ever more central to education, the human teacher becomes more important, not less. They will guide the learning experience, bringing published works to the conceptual framework of a course and driving in-person student engagement and encouragement. They can extend their value as personal AI tutors—via agents—for each student, based on individual learning needs.

Where do younger teachers fit in, who don’t have a back catalog to train LLMs? Well, the younger the teacher, the more AI-native they are likely to be. They can use AI to flesh out their own conceptual vision for a course by widening the research beyond their own work, by prompting the agent on what should be included.

In AI, two alternate positions are often simultaneously true. AI is both emotionally intelligent and tone deaf. It is both a glorified text predictor and a highly creative partner. It is costing jobs, yet creating them. It is dumbing us down, but also powering us up.

So too in teaching. AI threatens the learning space, yet can liberate powerful interaction. A prevailing wisdom is that it will make students dumber. But perhaps AI could actually be unlocking for students the next level of personalisation, challenge and motivation.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post I Got an AI to Impersonate Me and Teach Me My Own Course—Here’s What I Learned About the Future of Education appeared first on SingularityHub.

首次，科学家们实时记录整个小鼠大脑中的决策过程 || In a First, Scientists Record Decision-Making as It Happens Across a Whole Mouse Brain

2025-09-11T19:43:51+00:00

一份详尽的大脑图谱展示了大脑区域在复杂决策过程中的协作方式。

我们不断做出决策。如果我买南瓜香草拿铁，它会比我平时喝的黑咖啡让我更快乐吗？如果我选择一条风景优美的路线旅行，是否值得花费额外的时间？

过往和当前的经历影响每一个决策。通过脑成像，科学家们早已知道多个区域会协作以提取记忆，并将其与我们所见、所听、所思的内容整合，从而权衡选项。但由于分辨率相对较低，我们只能获得大脑复杂神经连接的大致轮廓。

一项全球性合作正在深入研究。在一项技术成就中，国际脑实验室发布了一张小鼠在进行复杂决策任务时的大脑动态图谱。

该研究小组于2017年启动，旨在将大脑活动与行为联系起来，这是神经科学领域的一个“圣杯”。这是一场艰难的斗争。以往的尝试只能测量小区域，各研究团队使用自己的行为测试方法，使得数据整合变得困难。

新的合作研究使用标准化程序，从全球多个实验室收集小鼠的大脑电信号。总体而言，科学家们使用了近700个脑植入设备，记录了139只小鼠的大脑活动，捕捉了大脑中约62万神经元的活动。

“这是首次有人在决策过程中产生覆盖整个大脑的单神经元活动图谱。其规模前所未有……[这些]共同代表了小鼠大脑体积的95%”，日内瓦大学研究作者Alexandre Pouget 在一份新闻稿中表示。

该留还是该走？

尽管已有数十年的研究，科学家们仍然不完全理解我们是如何做出决定的。

例如，你正在徒步时遇到一只熊。大脑立即进入高速模式：视觉皮层识别棕色物体为熊，并将其传递给大脑的情绪中心。后者激活以产生恐惧感，并向记忆区域询问应对方法。这些计算引导运动反应——远离、让自己显得更大，或迅速喷射熊喷雾。

多个神经网络同时激活以达成决策，但科学家们对系统的工作方式存在分歧。一方认为，大脑可能在记忆区域（比如YouTube上如何避免熊的视频）与高阶大脑区域结合，以形成决策。这一假说预测记忆或先验信息只在后期影响行动。

另一方则认为相反。他们认为，大脑各区域——包括早期感觉系统——在决策过程中会整合记忆，以决定最佳反应。这一过程可能更有效地在整个大脑中传播信息。

神经赌博

如同间谍监听电话线路，新研究的作者希望通过监听数十万神经元的“谈话”来解决这一争论。

这项研究依托于国际脑实验室的数据集，该数据集使用了699个 Neuropixels（一种开源脑植入设备）来记录小鼠中单个神经元的电信号。研究团队在超过100只小鼠的大脑中战略性地放置了这些设备，覆盖了近280个脑区。他们试图在所有合作实验室中保持记录的一致性。

“我们记录了超过50万个神经元的活动……共同代表了小鼠大脑体积的95%”，Pouget表示。

每个实验室都教小鼠完成相同的困难任务。每只小鼠进入一个类似游戏机的环境，并在屏幕的左侧或右侧看到一个黑白条纹图案（想象一下斑马皮肤）。然后它们必须使用前爪转动一个小轮子，将图像移动到屏幕中央，时间限制为一分钟。

如果成功，它们会获得美味的奖励；如果失败，则会受到白噪音的冲击和短暂的休息。在每次尝试之间，小鼠会努力保持前爪在轮子上，等待下一次测试。

关键点在于：这个游戏是被设计过的。条纹有80%的概率出现在某一方向，教导小鼠这是最佳选择。随着试验的进行，条纹逐渐变淡，直到几乎看不见。此时，小鼠必须根据之前学到的“最佳猜测”决定将条纹向左或向右移动。

每个实验室在小鼠做出选择时记录大脑信号，并将数据发送到中央数据库。总体而言，该合作团队在大脑中隔离了近76,000个神经元活动模式。记录点随后通过双光子显微镜技术拼接在一起，该技术能够精确地将解剖区域与大脑区域的电信号图谱对齐。

“我们曾看到大规模合作在物理学领域成功解答了单个实验室无法解决的问题，因此我们希望在神经科学领域尝试同样的方法，”来自伦敦大学学院的研究作者Tom Mrsic-Flogel 在新闻稿中表示。“大脑是宇宙中最复杂的结构，理解它如何驱动行为需要与该复杂性相匹配的国际协作规模。”

一个智慧的宇宙

利用新脑图谱的数据，研究团队发现决策过程并非线性。相反，多个大脑区域——包括所谓的“早期”感觉区域——都对最终选择有所贡献。

例如，小鼠处理视觉信息的大脑区域在看到条纹时会激活。这种活动随后以波浪式模式传播并增强，到达与情绪相关的大脑区域。这些信号引导它们将先前的学习，称为“先验”，整合到最终的决策中——是将轮子向左还是向右移动。

以前，科学家认为先验信息存储在与记忆和高级认知相关的大脑区域中。但新的图谱表明，这些信号也会影响早期感觉处理区域，从而影响最终的反应。

“我们合作的努力产生了关于支持复杂认知的大脑广泛回路的基本见解，”来自加州大学洛杉矶分校的研究作者Anne Churchland 在新闻稿中表示。“这真是令人兴奋，是相对于之前领域中‘零散’方法（一次只研究1-2个大脑区域）的重大进展。”

国际脑实验室正在发布整个数据库，目标是最终理解大脑在决策过程中内部和跨区域的计算机制。该数据集可能为患有决策障碍的神经疾病提供见解，如强迫症、帕金森病和成瘾。

文章首次，科学家记录决策过程在整只小鼠大脑中的发生首次出现在 SingularityHub。

---------------

An extensive brain map shows how regions collaborate during complex decision-making.

We’re constantly making decisions. If I get the pumpkin spice latte, would it make me happier than my usual black coffee? If I go the scenic route on a trip, would it be worth the extra time?

Past and current experiences affect each decision. By imaging the brain, scientists have long known multiple regions collaborate to pull in memories and integrate them with what we’re seeing, hearing, and thinking when weighing options. But because the resolution is relatively low, we’ve only had a rough sketch of the intricate neural connections involved.

A global collaboration is now digging deeper. In a technological feat, the International Brain Laboratory released a large, dynamic brain map of mice navigating a difficult decision-making task.

Launched in 2017, the group seeks to link brain activity with behavior, one of the holy grails in neuroscience. It’s been an uphill struggle. Prior attempts could only measure small regions, and individual teams used their own behavioral tests, making it difficult to integrate data.

The new collaboration gathered neural electrical recordings in mice from multiple labs across the globe using a standardized procedure. Overall, the scientists used nearly 700 brain implants to record neural activity in 139 mice, capturing the activity of 620,000 neurons across the brain.

“This is the first time anyone has produced a full, brain-wide map of the activity of single neurons during decision-making. The scale is unprecedented … [which] together represent[s] 95 percent of the mouse brain volume,” said study author Alexandre Pouget at the University of Geneva in a press release.

Should I Stay or Should I Go?

Despite decades of research, scientists still don’t fully understand how we make up our minds.

Say you’re hiking and encounter a bear. The brain immediately goes into hyper mode: The visual cortex identifies the brown thing as a bear and transmits this to the brain’s emotion centers. The latter activate to produce a sense of fear and ask the memory regions what to do. These calculations direct a motor response—back away, make yourself big, or quick-draw that bear spray.

Multiple neural networks fire up to reach the decision, but scientists are divided on how the system works. One camp thinks the brain could combine memories—say, YouTube videos of how to avoid bears—with the fact you’re seeing a bear in high-level brain regions. This hypothesis predicts memories, or prior information, only inform actions at later stages.

Another camp believes the opposite. Rather than waiting until the last second, all regions of the brain—including early sensory systems—incorporate memories to decide the best response. This process could better spread communication throughout the brain.

Neural Gambling

Like spies tapping phone lines, the authors of the new study hoped to settle the debate by listening in on the chatter of hundreds of thousands of brain cells.

The effort piggy-backed on an International Brain Laboratory dataset that used 699 Neuropixels, an open-source brain implant, to record the electrical firing of individual neurons in mice. The team strategically placed the devices across nearly 280 brain regions in over a hundred mice. They tried to keep the recordings relatively uniform as they tackled the same task across all collaborating labs.

“The scale is unprecedented as we recorded from over half a million neurons across mice…which together represent 95 percent of the mouse brain volume,” said Pouget.

Every lab taught the critters to perform the same difficult challenge. Each mouse entered an arcade of sorts and was shown a black-and-white grating—think zebra skin—on either the left or right side of a screen. They then had to use their front paws to turn a tiny wheel, moving the image to the center within a minute.

If they succeeded, they got a tasty reward. If they failed, they were blasted with a pop of white noise and a short time-out. Between trials, the mice tried to keep their paws on the wheel as they waited for the next test.

Here’s the crux: The game was rigged. There was an 80 percent chance the grate would appear in one direction, teaching the mice that was the best bet. As the trial went on, the grate slowly faded to the point it was almost impossible to see. The mice then had to decide whether to move it left or right based on what they’d previously learned as a best guess.

Each lab recorded brain signals as the mice made their choices and sent the data to a central database. In all, the consortium isolated nearly 76,000 neuron activity patterns across the brain. The recording sites were then stitched together using two-photon microscopy, a technology that impeccably lines up anatomical regions to maps of electrical activation in brain regions.

“We’d seen how successful large-scale collaborations in physics had been at tackling questions no single lab could answer, and we wanted to try that same approach in neuroscience,” said study author Tom Mrsic-Flogel at University College London. “The brain is the most complex structure we know of in the universe and understanding how it drives behavior requires international collaboration on a scale that matches that complexity.”

A Brainy Universe

Using data from the new brain map, the team realized that decision-making isn’t linear. Instead, multiple brain regions—including so-called “early” sensory ones—contribute to the final choice.

For example, brain regions in mice that process visual information sparked with activity upon seeing the grate. That activity then spread and ramped up in a wave-like pattern towards brain regions associated with emotion. These signals guided them to incorporate previous learning, called priors, into final decisions on what to do—move the wheel left or right.

Before, scientists thought priors were encoded in brain regions related to memory and higher cognition. But the new map suggests their signals also influence the early sensory processing regions that contribute to eventual responses.

“The efforts of our collaboration generated fundamental insights about the brain-wide circuits that support complex cognition,” said study author Anne Churchland at UCLA. “This is really exciting and a major step forward relative to the ‘piecemeal’ approach (1-2 brain areas at a time) that was previously the accepted method in the field.”

The International Brain Laboratory is releasing the entire database, with the goal of eventually understanding the brain’s computations within and across the brain regions behind decision-making. The dataset could shed light on neurological disorders with impaired decision-making, such as obsessive-compulsive disorders, Parkinson’s disease, and addiction.

The post In a First, Scientists Record Decision-Making as It Happens Across a Whole Mouse Brain appeared first on SingularityHub.

这个爬行机器人是由活体脑和肌肉细胞制成的 || This Crawling Robot Is Made With Living Brain and Muscle Cells

2025-09-09T23:10:10+00:00

科学家们想了解生物混合机器人是否能形成一种持久的生物“大脑”来指导运动。这是一个奇特的景象：通过一束光的短时照射，一个海绵状的机器人可以在瓷砖表面上滑动。当它被翻转到背部时，它会反复抽搐，仿佛在做仰卧起坐。通过调整光的频率，科学家可以改变这种奇怪生物的移动速度——以及它在长时间爬行后需要“休息”的时间。软体机器人并不新鲜，但这种海绵状的机器人之所以突出，是因为它将活体肌肉和脑细胞与3D打印的骨架和无线电子元件结合在一起。这些被基因改造的神经元会触发邻近的肌肉收缩或放松。看着机器人爬行是有趣的，但该研究的主要目标是看看生物混合机器人是否能形成某种持久的生物“大脑”来指导运动。神经元是非常敏感的细胞，一旦脱离精心控制的环境，它们会迅速停止工作甚至死亡。然而，通过使用不同类型的神经元的团块来控制肌肉，海绵状机器人能够保持爬行能力超过两周。科学家们已经制造了使用电力或光来控制肌肉细胞的生物混合机器人。一些机器人模仿游泳、行走和抓取的动作。加入神经元可以进一步精细调节它们的活动和灵活性，甚至赋予它们重复任务的某种记忆。这些生物混合机器人提供了一种独特的方式来研究运动、运动障碍和药物开发，而无需使用实验动物。由于它们的组件通常与生物体兼容，因此它们可用于诊断、药物输送和其他医疗场景。柔软但强大机器人这个词常常让人联想到《终结者》中的金属T-800。然而，软体机器人有潜力更加灵活和敏捷。它们能够轻微变形，从而可以挤过狭窄的空间，监测脆弱的生态系统，如珊瑚礁，探索深海，甚至在体内蛇行，对周围组织造成最小的损伤。除了合成材料和机制，另一种构建软体机器人的方式是受自然启发。从蓝鲸到啮齿动物和人类——所有生物都依赖类似的生物机制来移动。肌肉中的运动神经元接收来自大脑和脊髓的指令，然后释放化学物质来触发肌肉收缩或放松。这一过程是节能的，并且能迅速适应环境中的突然变化，例如跨过意想不到的台阶而不是绊倒。尽管当今的机器人越来越敏捷，但在崎岖地形中仍难以应对意外的地雷。加入神经肌肉连接可能会使机器人更加精确和高效。去年，一个团队在一项“概念验证”研究中制造了一种生物混合机器人，它由干细胞来源的神经元、心肌细胞和一个电子“大脑”组成，模仿了水母的游泳动作。科学家们将这些细胞与人工骨架结合，制造出一种软体机器人，可以在游泳池中摆动鳍部并自由游动。还有一个意外发现：两种细胞类型之间的连接形成了电信号突触。通常，神经元通过释放化学物质来指导肌肉运动，这些连接被称为化学突触。虽然电信号网络更快，但它们通常适应性较差。回到基础这项新研究旨在为机器人创造化学突触。研究团队首先3D打印了一个大致呈“8”字形的骨架，但中间部分更宽。每一边形成一个凹槽，其中一边比另一边更深。这些凹槽旨在充当腿部。研究人员随后将小鼠的肌肉细胞嵌入到每个凹槽中的营养凝胶中。五天后，这些细胞形成了能够使腿部整体收缩的肌肉丝。机器人的“大脑”位于“8”字形的中间部分。团队利用经过基因改造以对光产生反应的干细胞，制造出微小的神经组织团块，称为神经球。这些团块包含多种脑细胞，包括控制肌肉的运动神经元。神经球在移植后几天内与肌肉组织连接。细胞形成了与人体中类似的神经肌肉连接，并且生物混合机器人开始释放控制肌肉功能的化学物质。随后，团队引入了电子触控。他们添加了一个中心模块，用于无线检测光脉冲、收集能量并驱动五个微型LED灯，以改变脑细胞活动并将其转化为运动。机器人移动的速度非常缓慢，大约每分钟0.8毫米。然而，在试验中，腿部在同步抽搐，这表明神经元和肌肉之间形成了某种同步连接。令人惊讶的是，一些机器人在关闭光源后仍能继续移动，而其他“僵尸”机器人则会自发地移动。团队仍在研究为什么会发生这种情况。但性能差异是预料之中的——生物组件远不如无机部件可控。就像经历了一场艰难的锻炼，这些机器人也需要休息。当被翻转到背部时，它们的腿部在大约两周内持续移动，之后便失效了。这可能是因为代谢毒素的积累，这些毒素逐渐在机器人内部堆积，但团队正在寻找根本原因。尽管存在缺陷，这些机器人本质上是由活体微型神经网络和连接到电子元件的组织构成的——真正的半机械人。研究团队写道：“它们为理解神经元和神经肌肉连接的涌现行为提供了一个宝贵的平台。” 研究人员现在计划探索不同的骨架结构，并监测行为以进一步优化控制。加入更先进的功能，如感官反馈和多种肌肉结构，可能有助于机器人更接近我们神经系统敏捷性。而多个神经“中心”，如海洋生物所具有的，可以控制看起来与我们毫无相似之处的机器人中的不同肌肉。该文章《这款爬行机器人由活体大脑和肌肉细胞制成》首先出现在《奇点新闻》网站上。
---------------

Scientists want to know if a biohybrid robot can form a long-lasting biological “mind” to direct movement.

It’s a bizarre sight: With a short burst of light, a sponge-shaped robot scoots across a tiled surface. Flipped on its back, it repeatedly twitches as if doing sit-ups. By tinkering with the light’s frequency, scientists can change how fast the strange critter moves—and how long it needs to “rest” after a long crawl.

Soft robots are nothing new, but the spongy bot stands out in that it blends living muscle and brain cells with a 3D-printed skeleton and wireless electronics. The neurons, genetically altered to respond to light, trigger neighboring muscles to contract or release.

Watching the robot crawl around is amusing, but the study’s main goal is to see if a biohybrid robot can form a sort of long-lasting biological “mind” that directs movement. Neurons are especially sensitive cells that rapidly stop working or even die outside of a carefully controlled environment. Using blob-like amalgamations of different types of neurons to direct muscles, the sponge-bots retained their crawling ability for over two weeks.

Scientists have built biohybrid bots that use electricity or light to control muscle cells. Some mimic swimming, walking, and grabbing motions. Adding neurons could further fine-tune their activity and flexibility and even bestow a sort of memory for repeated tasks.

These biohybrid bots offer a unique way to study motion, movement disorders, and drug development without lab animals. Because their components are often compatible with living bodies, they could be used for diagnostics, drug delivery, and other medical scenarios.

Squishy But Powerful

The word robot often conjures images of Terminator’s metal T-800. Soft robots have the potential to be far more flexible and agile. Being able to slightly deform lets them squeeze through tiny spaces, monitor fragile ecosystems like coral reefs, explore the deep sea, and potentially snake through the body with minimal damage to surrounding tissues.

In addition to synthetic materials and mechanisms, another way to build soft robots is inspired by nature. From blue whales to rodents and humans—all rely on similar biological machinery to move. Motor neurons in muscles receive directions from the brain and spinal cord. They then release chemicals that trigger muscles to contract or relax.

The process is energy efficient and rapidly adapts to sudden changes in the environment—like stepping over an unexpected doorstep instead of tripping. Though today’s robots are getting more agile, they still struggle with unexpected landmines in uneven terrain. Adding neuromuscular junctions could lead to more precise and efficient robots.

Last year, in a proof of concept, one team engineered a swimming “stingray” bot using stem cell-derived neurons, heart muscle cells, and an electronic “brain.” Scientists combined the cells, and brain with an artificial skeleton to make a soft robot that could flap its fins and roam a swimming pool.

There was a surprise too—the junctions between the two cell types developed electrical synapses. Usually, neurons release chemicals to direct muscle movements. These connections are called chemical synapses. While electrical networks are faster, they’re generally less adaptable.

Back to Basics

The new study aimed to create chemical synapses in robots.

The team first 3D printed a skeleton shaped roughly like a figure eight, but with a wider middle section. Each side formed a trough with one side slightly deeper than the other. The troughs were intended to function as legs. The researchers then embedded muscle cells from mice in a nutritious gel contained in each trough. After five days, the cells had formed slivers of muscle capable of contracting throughout the legs.

The robot’s “brain” sat in the middle part of the figure eight. The team made tiny blobs of neural tissue, called neurospheres, out of stem cells genetically engineered to activate with light. The blobs contained a mix of brain cells, including motor neurons to control muscles.

The neurospheres connected with muscle tissue days after transplantation. The cells formed neuromuscular junctions similar in form and function to those in our bodies, and the biohybrid robots began pumping out chemicals that control muscle function.

Then came an electronic touch. The team added a hub to wirelessly detect light pulses, harvest power, and drive five tiny micro-LED lights to change brain cell activity and translate it into movement.

The robot moved at turtle speed, roughly 0.8 millimeters per minute. However, the legs twitched in tandem throughout the trials, suggesting the neurons and muscles formed a sort of synchrony in their connections.

Surprisingly, some bots kept moving even after turning off the light, while other “zombie” bots spontaneously moved on their own. The team is still digging into why this happens. But differences in performance were expected—living components are far less controllable than inorganic parts.

Like after tough workout, the robots also needed breaks. And when flipped on their backs, their legs moved for roughly two weeks but then failed. This is likely due to a buildup of metabolic toxins, which gradually accumulate inside the robots, but the team is looking for the root cause.

Despite their imperfections, the bots are essentially built from living mini neural networks and tissue connected to electronics—true cyborgs. They “provide a valuable platform for understanding…the emergent behaviors of neurons and neuromuscular junctions,” wrote the team.

The researchers are now planning to explore different skeletons and monitor behavior to fine-tune control. Adding more advanced features like sensory feedback and a range of muscle structures could help the bots further mimic the agility of our nervous system. And multiple neural “centers,” like in sea creatures, could control different muscles in robots that look nothing like us.

The post This Crawling Robot Is Made With Living Brain and Muscle Cells appeared first on SingularityHub.

AI协作者使这款可穿戴脑波读取设备的性能提升了四倍 || An AI Copilot Quadrupled the Performance of This Wearable Brain-Reading Device

2025-09-08T23:08:42+00:00

得益于人工智能和带有电极的头盔，参与者仅凭思维控制了机械臂。

众多科技初创公司正在竞相开发脑机接口，但这种侵入性技术可能在普及范围上存在限制。新研究显示，将人工智能与更不侵入性的脑机接口结合可能提供另一种有前景的方向。

由Neuralink和Precision Neuroscience等公司开发的最先进脑机接口最初旨在用于医疗应用。但技术乐观主义者也希望，未来这项技术可以被普通用户用来提升认知能力、用思维控制技术，甚至与人工智能融合。

但植入这些设备需要进行有风险的脑部手术，可能导致免疫反应，从而降低植入物的性能，甚至需要移除。在治疗严重残疾或疾病时，这些风险通常可以被证明是合理的，但对于没有实际医疗需求的健康人来说，这种计算更为复杂。

存在一些非侵入性的脑机接口，它们从颅骨外部记录电信号，但通常在检测脑信号方面准确性较低。现在，加州大学洛杉矶分校的研究人员表明，将这些设备与“AI副驾驶”结合可以显著提升性能，甚至允许人们控制机械臂。

“我们旨在寻找风险和侵入性更低的途径，”领导研究的Jonathan Kao在一份新闻稿中说。“最终，我们希望开发出AI-BCI系统，实现共享自主权，使患有运动障碍（如瘫痪或ALS）的人能够重新获得日常任务的独立性。”

研究人员在实验中使用的非侵入性设备是一个带有64个电极的头盔，设计用于捕捉脑电图信号。他们开发了一个专用算法来解码这些信号，并将其与针对特定任务设计的AI副驾驶结合。该系统由四名研究参与者测试，其中一人是腰部以下瘫痪。

第一个任务是将光标移动到电脑屏幕上的八个不同目标上至少停留半秒。通过强化学习，团队训练AI副驾驶根据EEG解码器的输入和目标及光标的位置数据推断用户的目标。然后，副驾驶利用这些信息引导光标向正确方向移动。

在Nature Machine Intelligence期刊上发表的一篇论文中，研究人员报告称，AI副驾驶使健康参与者的成功率提高了两倍，而瘫痪参与者成功率则提高了四倍。

随后，研究人员让参与者使用该接口控制机械臂，将桌面上的四个彩色方块移动到随机放置的标记上。该任务的AI副驾驶采用了类似原理，但使用摄像头画面来检测方块和标记在桌面上的位置。

借助AI副驾驶的帮助，健康参与者显著更快地完成了任务。瘫痪参与者在没有副驾驶帮助的情况下无法完成任务，但一旦激活，他们成功率达到93%。

研究人员表示，这项研究展示了“共享自主权”方法——即AI与脑机接口用户协作解决任务——可以显著提升非侵入性技术的性能。他们还认为这种方法可能改善侵入性植入物。

事实上，Neuralink已经在尝试类似的方法。今年早些时候，MIT Technology Review报道说，该公司的一名测试对象正在使用AI聊天机器人Grok来帮助起草信息并加快沟通速度。

然而，澳大利亚墨尔本大学的Mark Cook在Nature上表示，研究人员需要谨慎对待这些设置中给予AI的控制程度。“共享自主权不应以牺牲用户自主权为代价，AI干预可能会覆盖或误解用户的意图，”他说。

尽管如此，似乎脑机接口让AI与人类思维更无缝地互动的梦想可能提前实现了。

文章 AI副驾驶使这种可穿戴脑读设备性能提升四倍首次发表于 SingularityHub。

---------------

Thanks to AI and an electrode-studded cap, participants controlled a robotic arm with just their thoughts.

A host of tech startups are racing to build brain implants, but there may be limits to how widely such invasive technology can be adopted. New research shows that pairing AI with less invasive brain-computer interfaces could provide another promising direction.

The cutting-edge brain implants being developed by companies like Neuralink and Precision Neuroscience are initially aimed at medical applications. But techno-optimists also hope that in the future this technology could be used by everyday people to boost cognition, control technology with their thoughts, and even merge their minds with AI.

But implanting these devices requires risky brain surgery and can lead to immune reactions that degrade an implant’s performance or even require it be removed. When treating serious disabilities or diseases these risks can often be justified, but the calculus is trickier for healthy people with no real medical need.

There are less invasive brain interfaces that record electrical signals from outside the skull, but they are typically much less accurate at detecting brain signals. Now, researchers from the University of California, Los Angeles have shown that combining these devices with an “AI copilot” can dramatically boost performance and even allow people to control a robotic arm.

“We’re aiming for much less risky and invasive avenues,” Jonathan Kao, who led the research, said in a press release. “Ultimately, we want to develop AI-BCI systems that offer shared autonomy, allowing people with movement disorders, such as paralysis or ALS, to regain some independence for everyday tasks.”

The non-invasive device the researchers used in their experiments was a cap featuring 64 electrodes designed to capture electroencephalography, or EEG, signals. They developed a custom algorithm to decode these signals, which they then combined with AI copilots designed for specific tasks. The system was tested by four study participants, one of whom was paralyzed from the waist down.

The first task was moving a cursor on a computer screen to hover over eight different targets for at least half a second. Using reinforcement learning, the team trained the AI copilot to infer what target the user was aiming for by looking at inputs from the EEG decoder and position data from the targets and cursor. The copilot then used this information help steer the cursor in the right direction.

In a paper in Nature Machine Intelligence, the researchers report that the copilot boosted the success rate of the healthy participants by a factor of two compared to using the interface without AI, while the paralyzed participant saw their success rate quadruple.

The researchers then had users control a robotic arm with the interface to move four colored blocks on a table to randomly placed markers. The copilot for this task worked on similar principles but used a camera feed to detect the position of blocks and targets on the table.

With the copilot’s aid, the healthy participants solved the task significantly faster. The paralyzed participant was unable to complete the task without help from the copilot, but once it was activated, they were successful 93 percent of the time.

The researchers say the study shows this kind of “shared autonomy” approach—where AI and brain interface users collaborate to solve tasks—can significantly boost the performance of non-invasive technology. They also suggest it could improve invasive implants as well.

In fact, Neuralink is already experimenting with similar approaches. Earlier this year, MIT Technology Review reported that one of the company’s test subjects was using the AI chatbot Grok to help draft messages and speed up the rate at which he could communicate.

However, Mark Cook at the University of Melbourne in Australia told Nature that researchers need to be careful about how much control is given to the AI in these kinds of setups. “Shared autonomy must not come at the cost of user autonomy, and there is a risk that AI interventions could override or misinterpret user intent,” he said.

Nonetheless, it seems the dream of brain-computer interfaces allowing AI and human minds to interact more seamlessly may be arriving ahead of schedule.

The post An AI Copilot Quadrupled the Performance of This Wearable Brain-Reading Device appeared first on SingularityHub.