Meta Has Its Muse

"Motley Fool Hidden Gems Investing" 播客，由 特拉维斯·霍伊 主持，卢·怀特曼 和 瑞秋·沃伦 共同参与，深入探讨了 Meta Platforms 和 Google DeepMind 在人工智能方面的重大进展，分别从投资角度讨论了它们对消费者和医疗领域的影响。 主要议题是 Meta 推出的 Muse 应用，该应用搭载了其新的 Muse Spark 模型。特拉维斯 指出 Meta 过去在消费者 AI 市场份额方面举步维艰，市场份额常被 Anthropic 和 ChatGPT 等公司占据。卢 对“虚拟朋友”的概念表示兴奋，但他质疑 Meta 在消费者领域取得成功的可能性，认为 Meta 的优势在于消费者领域，而非企业级 AI。然而，他怀疑 Muse 是否能证明 Meta 巨额数据中心投资的合理性，理由是过去的结果不尽如人意，以及用户可能面临的高昂成本。瑞秋 承认消费者购物格局正在演变，人们越来越习惯 AI 代理协助完成预订航班或酒店等任务。她强调了 Muse 的广泛功能，它可以在独立的虚拟环境（"Muse Secure VM"）中运行代理，从而能够操作网络工具并与 Google Workspace 和 Spotify 等服务集成。虽然 Muse 可能不会立即产生显著影响，但其订阅模式（20 美元和 100 美元的套餐，并提供免费层级）是未来实现盈利的先驱。一个主要担忧是数据隐私，因为 Muse 需要深度访问个人通信和金融工具。尽管 Meta 的 "Sentinel" 代理旨在保护敏感数据，但内部测试却发现了处理问题，导致消费者犹豫不决。卢 强调了 Meta 的信任赤字，并举了一个演示的例子，其中 AI 似乎在没有明确提示的情况下知道了孩子的名字。他认为，像 Gemini 或 Siri 这样熟悉的手机集成助手更有可能赢得消费者 AI 竞赛。特拉维斯 质疑了消费者采用曲线，将其与企业在编码领域迅速采用 AI 进行了对比，并指出消费者可能不愿连接个人数据。卢 建议投资者“押注现状”，并以信用卡普及缓慢（70 年才达到 30% 的支付份额）为例进行说明。他还批评了产品演示中婴儿车示例，强调此类决策的个人性质，而硅谷常常误解这一点。瑞秋 将 Meta 的努力视为“持续性创新”，大型科技公司正利用其资本和分销优势。她提到了 Hubspot 的一项预测，即两年内 95% 的买家旅程将始于 LLM 聊天，但强调了“信任瓶颈”和“监管瓶颈”。值得注意的是，电商巨头 Amazon 没有出现在讨论中，这表明其现状受益于不整合这些新的 AI 代理。 讨论随后转向 Google DeepMind 的 AlphaGenome Atlas，它被誉为下一代 AlphaFold。瑞秋 解释说，这个 AI 模型能够计算人类 DNA 中每一个可能变化的精确分子影响。尽管科学家们明白基因组中 2% 的部分编码蛋白质，但剩余的 98%（基因的“控制面板”）在很大程度上仍是个谜。AlphaGenome 分析数百万个碱基对，计算个体突变如何破坏生物过程并分配损害/影响评分。瑞秋 将人类 DNA 比作软件代码，AlphaGenome 则负责识别神秘的 98% 中的每一个可能的“拼写错误”，并为研究人员编制一个可搜索的数据库。最初免费，但制药公司很快将需要通过 Google Cloud 授权使用 Atlas 进行药物研发。早期验证包括 纪念斯隆凯特林癌症中心 和 斯坦福大学 使用它解决了复杂的医疗案例，例如发现了一种罕见癫痫病的突变。 然而，卢 告诫不要直接投资，称其“不可能”，并认为这可能更多地为 Google Cloud 带来“顺风”。他引用 礼来 首席执行官 David Ricks 的话，强调了人体的复杂性，指出很少有疾病是由单一基因引起的。尽管 AlphaGenome 是一个“很好的起点”，但这只是“第一步”。卢 强调 AI 模型“说的是英语，而不是生物学”，因为人类自己也未能完全理解生物学的语言。他估计在十年或更长时间内才能看到切实的投资回报，敦促听众为人类健康方面的进步喝彩，但要避免仅仅因为这项技术的授权而大量购买生物技术股票。特拉维斯 最后总结说，在更广泛的担忧中，他强调了人工智能对人类健康的积极影响。

The "Motley Fool Hidden Gems Investing" podcast, hosted by Travis Hoy with Lou Whiteman and Rachel Warren, delved into significant AI developments from Meta Platforms and Google DeepMind, discussing their consumer and medical implications, respectively, from an investment perspective. The primary topic was Meta's introduction of the Muse app, featuring its new Muse Spark model. Travis noted Meta's past struggle in consumer AI mindshare, often ceded to companies like Anthropic and ChatGPT. Lou expressed excitement over the concept of an "imaginary friend" but questioned Meta's ability to succeed in the consumer space, suggesting their strength lies there, unlike enterprise AI. However, he doubted if Muse could justify Meta's substantial data center investments, citing past underwhelming results and the potential high cost for users. Rachel acknowledged the evolving consumer shopping landscape, with more comfort in AI agents assisting with tasks like booking flights or making reservations. She highlighted Muse's broad capabilities, running agents in isolated virtual environments ("Muse Secure VM") that can navigate web tools and integrate with services like Google Workspace and Spotify. While Muse might not immediately move the needle, its subscription model ($20 and $100 plans with a free tier) is a precursor to future monetization. A significant concern raised was data privacy, as Muse requires deep access to personal communications and financial tools. Despite Meta's "Sentinel" agent designed to protect sensitive data, internal tests flagged handling issues, leading to consumer hesitancy. Lou emphasized Meta's trust deficit, citing a demo where the AI seemed to know a child's name without explicit prompting. He argued that familiar phone-integrated assistants like Gemini or Siri are more likely to win the consumer AI race. Travis questioned the consumer adoption curve, contrasting it with the rapid enterprise adoption of AI for coding, noting a likely reluctance to connect personal data. Lou advised investors to "bet on the status quo," illustrating with the slow adoption of credit cards (70 years to reach 30% of payments). He also critiqued the product demo's stroller example, highlighting the personal nature of such decisions, which Silicon Valley often misunderstands. Rachel viewed Meta's efforts as "sustaining innovations," with large tech companies leveraging their capital and distribution. She mentioned a Hubspot projection of 95% of buyer journeys starting in LLM chats within two years but stressed the "trust bottleneck" and "regulatory bottleneck." Notably, Amazon, the e-commerce giant, was absent from the conversation, suggesting its current status quo benefits from not integrating these new AI agents. The discussion then shifted to Google DeepMind's AlphaGenome Atlas, hailed as the next generation of AlphaFold. Rachel explained that this AI model computes the exact molecular impact of every possible change in human DNA. While scientists understood that 2% of the genome codes for proteins, the remaining 98% (a "control panel" for genes) was largely a mystery. AlphaGenome analyzes millions of base pairs, calculating how individual mutations disrupt processes and assigning harm/influence scores. Rachel used the analogy of human DNA as software code, with AlphaGenome identifying every possible "typo" in the mysterious 98% and compiling a searchable library for researchers. Initially free, pharmaceutical companies will soon need to license Atlas via Google Cloud for drug discovery. Early validation includes Memorial Sloan Kettering Cancer Center and Stanford University using it to solve complex medical cases, such as uncovering a mutation for a rare form of epilepsy. Lou, however, cautioned against direct investment, calling it "impossible" and suggesting it might offer more "tailwinds for Google Cloud." Citing Eli Lilly CEO David Ricks, he emphasized the complexity of the human body, noting that few diseases are caused by a single gene. While AlphaGenome is a "great starting point," it's just a "first step." Lou stressed that AI models "speak English, they don't speak biology" because humans themselves don't fully understand biology's language. He estimated a decade or more before tangible investment returns, urging listeners to celebrate the progress for human health but avoid loading up on biotech stocks based solely on licensing this technology. Travis concluded by highlighting the positive human health implications of AI amidst broader concerns.