内容来源：https://www.theverge.com/column/935021/google-io-gemini-for-science-alphafold-alphagenome-ai-health

内容总结：

谷歌DeepMind CEO放言“AI将治愈所有疾病”，专家呼吁警惕“科学洗白”

在本年度谷歌I/O开发者大会的主题演讲尾声，谷歌DeepMind首席执行官德米斯·哈萨比斯面无表情地宣称，公司希望“重新构想药物发现过程，目标是有朝一日解决所有疾病”。这一惊人言论迅速引发争议。

哈萨比斯实际介绍的是“科学双子座”（Gemini for Science）——一系列旨在鼓励研究人员探索和发现新成果的实验性AI工具。然而，对于普通观众而言，这句话极易被误解为“双子座AI将凭借其力量治愈所有疾病”。批评者指出，这种表述忽略了真实的医学突破逻辑。

事实上，AI在医学研究领域已应用数十年。可穿戴设备的算法、无创检测技术背后都离不开机器学习。一项荟萃分析显示，AI在缩短新冠疫苗研发时间上发挥了关键作用，但同时也暴露出算法偏见、数据隐私和全球公平获取等重大伦理与监管挑战。

哈萨比斯提到的AlphaFold和AlphaGenome项目确有实际价值：前者帮助科学家理解蛋白质结构，已用于研发疟疾疫苗、发现“坏胆固醇”关键蛋白等；后者则能预测人类DNA序列突变，但谷歌在《自然》期刊论文中也坦承，该模型尚未被验证可用于个人基因组预测，且在捕捉细胞和组织特异性模式方面存在局限。

真正的问题在于，这类宏大声明在短视频和注意力碎片化的时代极易被曲解。此前，美国卫生与公众服务部部长小罗伯特·F·肯尼迪就曾声称AI可能让FDA“变得无关紧要”。这种不负责任的类比与哈萨比斯的言论一旦被关联，可能加剧公众误解。

资深科技评论员指出，无论AI多么强大，它都无法替代数十年来建立的药物临床试验、动物测试等科学严谨流程。从“AI能解决所有疾病”到“追踪生物数据、服用补剂、战胜死亡”的逻辑跳跃，正是眼下流行的“科学洗白”话术——用几个时髦词汇和夸张声明营造高科技合法性的假象。

科学家们强调，即便AI终将助力攻克顽疾，这个过程也绝不会简单明了。考虑到地缘政治、社会文化变迁对临床研究能力的深远影响，未来20年充满变数。正如一位评论者所言：“对哈萨比斯的乐观，我暂时持保留态度。”

中文翻译：

这是《优化指南》（Optimizer）——由《边缘》（Verge）资深评论员维多利亚·宋（Victoria Song）每周撰写的一封新闻简报，旨在剖析和讨论那些声称能改变你生活的最新科技产品与神奇方案。本周的这期是配合《边缘》对谷歌I/O大会报道而推出的特别早刊。下一期将在下周五的常规时间与您见面。点击此处订阅《优化指南》。

“你说要‘治愈所有疾病’？”
谷歌DeepMind首席执行官戴米斯·哈萨比斯（Demis Hassabis）在今年I/O大会主题演讲中抛出了一个大胆的声明。先别急着信！

在今年谷歌I/O大会主题演讲接近尾声时，谷歌DeepMind首席执行官戴米斯·哈萨比斯面无表情地宣布，该公司希望“重塑药物研发流程，以期有朝一日能治愈所有疾病”。

这话正是“若真如此，意义重大”这句名言的绝佳注解。

哈萨比斯真正描述的是“Gemini for Science”——一套实验性AI工具集，旨在鼓励研究人员进行探索并实现新发现。

在《优化指南》中，我经常对AI健康领域持批评态度，但哈萨比斯的这句话确实需要更多的背景解读。良好的科学传播——既能让普通人理解，又不至于无意中助长错误信息——已变得越来越困难。想必I/O大会现场的研发人员都明白，这句话的本意是：AI的进步已大幅缩短了医学新发现所需的时间。但对于普通人（甚至可以说，对于科学传播者而言），这听起来很可能像是“Gemini能治愈所有疾病，因为这就是AI的力量”。然而，医学突破在现实世界中绝非如此运作。

几十年来，AI一直是医学研究与发现中不可或缺的一部分。可穿戴设备所用的算法？那是AI。非侵入式可穿戴检测功能的发现？那是机器学习，老铁。生成式AI相对较晚才进入这一研究领域，但它蕴含着巨大的潜力。作为工作的一部分，我经常与临床研究人员交流，近年来消费健康科技的许多突破，部分归功于AI的进步。例如，一项荟萃分析发现，AI在缩短新冠疫苗的研发时间线上发挥了重要作用。这是全世界都受益的成果。然而，该分析也指出，在算法偏见、数据隐私和全球公平可及性等方面，像这样使用AI仍面临着重大的伦理、后勤和监管挑战。

在主题演讲中，哈萨比斯提到了谷歌的AlphaFold和AlphaGenome项目。前者帮助研究人员更好地理解蛋白质结构。这之所以重要，是因为蛋白质在无数生物过程中扮演着多种角色。更深入地理解蛋白质——甚至设计新型合成蛋白质——可能是攻克癌症疗法的关键。（最近，科学家们发现了1700种可能做到这一点的蛋白质。）传统上，发现新蛋白质、了解其功能以及它们与其他分子的相互作用，是一个耗时数年的过程。像AlphaFold这样的工具有助于大幅缩短这一时间线。在实际案例中，研究人员已利用该模型帮助开发疟疾疫苗，发现低密度脂蛋白（即“坏胆固醇”）背后的关键蛋白质，并理解早发性帕金森病背后的另一种蛋白质等。

与此同时，AlphaGenome是另一个帮助研究人员预测人类DNA序列突变的模型。该模型的潜力在于，它可能帮助研究人员理解某些疾病发生的原因，尽管谷歌在一项《自然》杂志的研究中指出，它存在重要的局限性。例如，该模型尚未经过验证，甚至并非为个人基因组预测而设计，并且它在捕捉细胞和组织特异性模式方面存在困难。这些对研究人员来说是很重要的细微差别，但通常其他人对此会毫无察觉。

从很多方面来看，哈萨比斯在台上说的那些话，目标受众并不是你或我。此外，还有另一个重要的背景：这些AI模型和Gemini for Science工具，并不会在未来三年、五年甚至十年内神奇地根除癌症或所有以前“无法解决”的疾病。像这样的事情，更可能至少需要20年，甚至更久。你可能会觉得这时间很长——特别是考虑到这对你眼下生病的亲人，或对你自己的寿命意味着什么。但就严谨的科学研究而言，这已经是一个雄心勃勃且激进的估计了。

然而，在一场你要宣布四万十亿个其他AI代理和功能的主题演讲中，你并没有时间去解释这些。问题在于，这类言论传播甚广，影响范围极大。对我们大多数人来说，迄今为止的AI健康体验可谓糟糕透顶：无非是重复性的指标汇总、胡言乱语和令人厌烦的保姆式指引。我们不应该将面向研究人员的AI工具与面向消费者的AI健康功能混为一谈，但这么做又是人之常情。

我对哈萨比斯言论的本能反应，是想起卫生部长小罗伯特·F·肯尼迪（RFK Jr.）最近的一份声明。在一次国会听证会上，肯尼迪表示，AI可能让美国食品药品监督管理局“变得无关紧要”。他的逻辑是，AI可以帮助开发并批准新药。将这与哈萨比斯的言论——一个完全不同的背景——相比较，你就会看到普通人可能会如何做出误导性的联想。例如，认为谷歌在附和或认可肯尼迪的分析。

并非毫无根据，《边缘》此前曾报道过为什么肯尼迪关于AI在健康领域作用的观点是有缺陷的。但作为复习，去年在接受塔克·卡尔森（Tucker Carlson）采访时，肯尼迪表示，AI可以迅速加速药物审批流程。这是一个宽泛的说法，并非完全不正确。是的，AI工具长期以来一直用于这一领域。是的，更新、更强大的模型可能会让研究人员和制药公司的流程更简单、更高效。但这并不能消除对FDA药物试验、动物试验或各种已实施数十年的流程的需求。AI归根结底是一种需要专家输入和协作的工具，而且说了一百万遍了，科学严谨性绝不是一个可以随意跳过的步骤。

背景信息至关重要，而它往往是热门金句中最先被舍弃的东西。这就是为什么，当我最初勾勒“健康骗子套路手册”时，我指出第一步通常是将一个宽泛的事实与一个误导性的论断并列。要明确的是，我并不是说哈萨比斯在主题演讲中的声明犯下了弥天大错。谷歌（以及苹果）实际上做了大量的临床研究，并努力通过博客来传达这些努力。但是，就像传话游戏一样，在如今这个短视频流行、注意力缩短和媒体素养下降的时代，很多东西都在传播中丢失了。我没有什么解决方案，只能尽力在可能的地方多提供背景信息，并希望它能触及合适的受众。

“科学洗白”如今如此盛行，是有原因的。几个流行词或大胆声明就能营造出一种高科技的可信度，从而抹杀掉细微差别。在硅谷，你会看到那些参加“肽派对”或信奉布莱恩·约翰逊（Bryan Johnson）那套以延长寿命为核心的生物黑客理念的技术精英。“AI能治愈所有疾病”和“追踪你的生物指标，用这些补充剂优化身体，战胜死亡”之间，并没有太大的差距。

也许有朝一日，AI最终能帮助治愈所有疾病。但即便真的如此，这条路也不会是清晰或简单的。未来20年可能发生很多事情，尤其是在政治、社会和文化环境方面，这些也将影响临床研究的能力——所以，请原谅我，眼下我并没有哈萨比斯那么乐观。

英文来源：

This is Optimizer, a weekly newsletter sent from Verge senior reviewer Victoria Song that dissects and discusses the latest gizmos and potions that swear they’re going to change your life. This week’s issue is a special early edition tied to The Verge’s Google I/O coverage. You can expect our next issue at its usual time next Friday. Opt in for Optimizer here.
‘Solve all diseases,’ you say?
Google DeepMind CEO Demis Hassabis made a bold claim at this year’s I/O keynote. Not so fast!
Toward the end of this year’s Google I/O keynote, Google DeepMind CEO Demis Hassabis declared, with a completely deadpan face, that the company hopes to “reimagine the drug discovery process with the goal of one day solving all disease.”
This is the sort of statement that the phrase “big, if true” was coined for.
What Hassabis was really describing was Gemini for Science, a collection of experimental AI tools designed to encourage researchers to explore and make new discoveries.
I’m often critical of AI health in Optimizer, but Hassabis’ statement is one that deserves a lot more contextualization. Good science communication — something that is digestible enough for the layperson, that doesn’t unintentionally promote misinformation — has become increasingly difficult. Surely the researchers in the I/O audience understood the claim to mean that advances in AI have dramatically reduced the time it takes to make new medical discoveries. But for the average person (and arguably, even science communicators), it probably sounded like “Gemini is going to cure every disease because that is the power of AI.” This is just not how medical breakthroughs work in the real world.
For decades, AI has been an integral part of medical research and discovery. The algorithms that wearables use? That’s AI. Discoveries for noninvasive, wearable detection features? Machine learning, baby. Generative AI is a relatively newer entrant into this area of research, but it holds incredible promise. As part of my job, I often speak with clinical researchers, and many of the breakthroughs in consumer health tech over the years are due in part to AI advances. For example, this meta review found that AI played a major role in reducing the development timeline for the covid-19 vaccinations. That’s something that the entire world benefited from. However, the review also found that significant ethical, logistical, and regulatory challenges remain in using AI like this with regard to algorithmic bias, data privacy, and equitable global access.
In the keynote, Hassabis pointed to Google’s AlphaFold and AlphaGenome projects. The former helps researchers better understand protein structures. This is important because proteins play myriad roles in countless biological processes. Better understanding proteins — or even designing novel synthetic proteins — could be the key to unlocking cancer treatments. (Recently, scientists found 1,700 new proteins that might do just that.) Traditionally, to discover new proteins, what they do, and how they interact with other molecules was a yearslong process. Something like AlphaFold helps to dramatically reduce that timeline. In terms of real-life case studies, researchers have used this model to help develop malaria vaccines, discover a key protein behind LDL (or the “bad cholesterol”), and understand another protein behind early-onset Parkinson’s disease, among other applications.
Meanwhile, AlphaGenome is another model that helps researchers predict mutations in human DNA sequences. The potential for this model is that it may help researchers understand why certain diseases happen, though in a Nature study, Google has noted that there are important limitations. For instance, this model hasn’t been validated or even designed for personal genome prediction, and it struggles to capture cell- and tissue-specific patterns. These are important nuances for researchers, but something that typically will fly over the heads of everybody else.
In many respects, what Hassabis was saying onstage wasn’t directed at you or me. And, some other important context, these AI models and Gemini for Science tools are not going to magically eradicate cancer or every previously “unsolvable” disease in the next three, five, or even 10 years. Something like this is more likely to take at least 20 years, probably more. You might think that’s a long time — especially in terms of what that means for a currently sick relative, or your own lifespan. But as far as rigorous scientific research goes, that’s an ambitious, aggressive estimate.
But this isn’t exactly something you have time to explain at a keynote where you’re announcing forty bajillion other AI agents and features. The problem is that these statements travel far and have a wide-ranging impact. For the majority of us, AI health has been, thus far, a craptacular experience of regurgitated metric summaries, hallucinations, and tedious hand-holding. We shouldn’t necessarily conflate AI tools for researchers and consumer AI health features, but it’s extremely human to do so.
My gut reaction to Hassabis’ comment was remembering a recent statement from Health Secretary RFK Jr. In a congressional hearing, Kennedy said that AI might make the Food and Drug Administration “irrelevant.” His logic is that AI could help develop and approve new drugs. Compare that to Hassabis’ comment — something with a completely different context — and you can see how the layperson’s reaction may leap to misleading associations. For example, that Google is parroting or lending credence to Kennedy’s analysis.
Not for nothing, The Verge has previously reported on why Kennedy’s take on AI in the health space is flawed. But as a refresher, in an interview with Tucker Carlson last year, Kennedy stated that AI could rapidly accelerate the drug approval process. That’s a broad statement that isn’t wholly untrue. Yes, AI tools have long been used in this space. Yes, newer, more powerful models could make researchers’ and pharmaceutical companies’ processes a lot easier and more efficient. But it doesn’t eliminate the need for FDA drug trials, animal testing, or various processes that have been in place for decades. AI is ultimately a tool that requires expert input and collaboration, and for the millionth time, scientific rigor is not a step that can be skipped willy-nilly.
Context is king, and it’s usually the first thing to go in buzzy soundbites. This is why, when I first outlined the wellness grifter playbook, I said step one is generally to juxtapose a broad fact next to a misleading assertion. To be clear, I’m not saying that Hassabis has committed a colossal crime with his statement during the keynote. Google (and Apple) actually does a lot of clinical research and puts effort into communicating that effort in blogs. But, like a game of telephone, there is a lot that gets lost in this current age of short-form social videos, reduced attention spans, and declining media literacy. I have no solution, other than to try and plug in more context whenever, wherever possible and hope it finds the appropriate audiences.
There’s a reason why sciencewashing is so prevalent today. A few buzzwords or bold statements lend an air of high-tech legitimacy that erases nuance. In Silicon Valley, you can see it in tech bros who attend peptide parties or subscribe to Bryan Johnson’s brand of longevity-focused biohacking. It’s not a huge leap from “AI can solve all diseases” to “track your biometrics, optimize with these supplements, and defeat death.”
Maybe AI will eventually, one day, help solve all diseases. But if it does, the path won’t be clear-cut or simple. A lot can happen in the next 20 years, especially in the political, societal, and cultural milieu that’ll also impact clinical research capabilities — so forgive me if, right now, I’m not quite as optimistic as Hassabis.