What Would Make an AI Assistant Worth Paying For?

在一期A16Z节目中，主持人与“Assistant Bench”基准测试的作者大卫·保林（David Paulin）热烈讨论了个人代理和消费者AI的快速发展。他们指出AI演变中的几个关键时刻，始于ChatGPT的对话能力，随后是编码代理，最近则是Instinct、Muse和ChatGPT Voice等个人代理的爆发式增长。 保林强调了他的“Assistant Bench”——一个面向消费者的基准测试，它根据单次提示在16个维度上比较了122种不同的AI助手。他的网站在播客发布前16天上线，已吸引了超过10万访客，表明该领域引起了广泛的好奇心。他将代理分为B2C通用代理（如Instinct、Muse、Caddy、Ollie、Season）和B2B工作流助手（如Town、Catch、Vellum）。 讨论随后转向了消费者动机，保林同意本·汤普森（Ben Thompson）的观点，即大多数消费者优先考虑花时间而非节省时间。然而，他认为金融领域，尽管是不常发生但价值高的行为，可能成为代理的关键应用领域，因为它涉及行政管理开销以及通过节约成本行动实现“免费的钱”的潜力。例如，使用代理进行HSA报销、机票降价退款，甚至智能家居自动化，如将洒水器连接到天气预报以减少50%的水费。保林认为，这些主动的、节约成本的代理的“隐形”本质对大规模采用至关重要。 对话探讨了用户与代理的互动方式，对比了iMessage集成和独立应用程序。虽然有些人喜欢iMessage的个人化感受，但另一些人，比如保林的妻子，则更喜欢视觉应用界面。最近发布的Muse配件引发了保林的“犀利观点”：他认为这不关乎赢得硬件竞争，更多是Meta收集真实世界数据以驱动扎克伯格（Zuckerberg）的元宇宙，利用其摄像头和多个麦克风作为环境感知、全天候的伴侣。然而，主持人则思考Meta的仅限音频的眼镜是否会成为“黑马产品”，在不引起社交尴尬的情况下提供连接性。 保林分享了他在骑自行车通勤时使用ChatGPT Voice的“神奇时刻”，他通过语音指令免提地对邮件进行分类、发送邀请和回复，从而实现了“零收件箱”。这强调了用户在忙碌时代理的实用性。 关于社交互动，他们承认将代理整合到群聊中而不具侵入性是一个挑战。虽然有些代理作为传统聊天参与者加入，但Instinct使用“代理网络”进行幕后沟通。沙恩·麦克（Shane Mack）的“Doc”提出了一种新方法，其中一个沉默的代理会倾听、记录笔记，并独立向个人发送行动事项提醒，从而防止群聊中断。他们一致认为模型在“应对技巧”方面的进展已停滞不前，使得社交贡献变得棘手，并建议代理在完成目录整理等实用性任务时更具增益作用。 发言者们讨论了市场是会根据代理的个性（例如，俏皮型 vs. 谨慎型）进行细分，还是根据其“内在特性”（例如，高度冒失型 vs. 高度谨慎型）进行细分。保林认为，由于可配置性，个性并不是一个强大的护城河，但“主动性”是一个巨大的差异化因素。他们提醒，代理必须走一条“非常微妙的界限”：虽然起草邮件或获取航班积分等主动行动受到欢迎，但直接影响用户（例如更换保险）的决策需要获得许可才能维持信任。主持人开玩笑地预测代理会和伴侣分手，突显了过度冒失的风险。 展望未来，保林建议消费者代理将越来越趋向于主动性和高度专业化，与主持人关于“垂直型初创公司”的论点相符——为小型、特定用户群体构建有价值的软件。他们还讨论了“助手”（执行指令性任务）与“代理”（拥有自主发起行动的能力）之间的区别。 播客随后展望了一个未来，代理将深刻影响人类，自动化繁琐任务，为家人和朋友腾出时间，从而形成一个“以消费者为中心的互联网”。他们探索了代理间的互动，预见到商业和服务的新范式。亚马逊阻止Muse而Shopify接纳它的例子说明了这种转变：亚马逊担心代理驱动的购买会导致去中介化和广告收入损失，而Shopify则将代理视为为个体卖家实现商业民主化。他们还设想代理如何可能扰乱餐厅预订等供应受限市场，可能导致基于忠诚度的访问或竞价战。 最后，他们讨论了经济方面，指出观察到的122个代理中有65个是付费的。这与Muse和Instinct等占主导地位的免费产品形成对比，对长尾初创公司构成了挑战。他们希望浏览器使用成本的下降能让更多产品免费，但也对消费者愿意支付高额费用的高价值代理表示了兴趣，表明这是一种“可能性拓展”而非单纯的成本降低。

In a recent episode of the A16Z Show, the host and David Paulin, author of the "Assistant Bench" benchmark, enthusiastically discussed the rapid advancements in personal agents and consumer AI. They noted several pivotal moments in AI's evolution, starting with ChatGPT's conversational capabilities, followed by coding agents, and most recently, the explosion of personal agents like Instinct, Muse, and ChatGPT Voice. Paulin highlighted his "Assistant Bench," a consumer-facing benchmark comparing 122 different AI assistants across 16 dimensions based on single-shot prompts. His site, launched 16 days prior to the podcast, had already garnered over 100,000 visitors, indicating widespread curiosity in the space. He categorized agents into B2C generalized agents (like Instinct, Muse, Caddy, Ollie, Season) and B2B workflow assistants (like Town, Catch, Vellum). Discussion then turned to consumer motivations, with Paulin agreeing with Ben Thompson's sentiment that most consumers prioritize spending time over saving it. However, he argued that finance, despite being an infrequent high-value behavior, could be a key area for agents due to its administrative overhead and potential for "free money" through cost-saving actions. Examples included using agents for HSA reimbursements, flight price drop refunds, or even smart home automation like connecting sprinklers to weather for 50% water bill reduction. Paulin believes the "invisible" nature of these proactive, cost-saving agents will be crucial for mass adoption. The conversation explored how users interact with agents, contrasting iMessage integration with separate applications. While some prefer the personal feel of iMessage, others, like Paulin's wife, prefer visual app interfaces. The recent release of the Muse charm sparked a "hot take" from Paulin: he believes it's less about winning the hardware game and more about Meta collecting real-world data to fuel Zuckerberg's metaverse, leveraging its camera and multiple microphones as an ambient, 24/7 companion. The host, however, pondered if Meta's audio-only glasses might be a "sleeper hit," offering connectivity without social awkwardness. Paulin shared a "magic moment" with ChatGPT Voice during his bike commute, where he achieved "inbox zero" by categorizing emails, sending invites, and replying, all through voice commands while hands-free. This underscored the utility of agents when users are occupied. Regarding social interactions, they acknowledged the challenge of integrating agents into group chats without being intrusive. While some agents join as traditional chat participants, Instinct uses an "agent network" for behind-the-scenes communication. A new approach by Shane Mack's "Doc" was noted, where a silent agent listens, takes notes, and independently pings individuals for action items, preventing group chat disruption. They agreed that models' progress on "bedside manner" has plateaued, making social contributions tricky, and suggested agents are more additive when fulfilling utilitarian tasks like cataloging collections. The speakers debated whether the market segments by agent personality (e.g., sassy vs. cautious) or by "constitution" (e.g., highly presumptuous vs. highly cautious). Paulin argued personality isn't a strong moat due to configurability, but "proactivity" is a massive differentiator. They cautioned about the "very fine line" agents must walk: while proactive actions like drafting emails or securing flight credits are welcomed, decisions impacting users directly (like switching insurance) require permission to maintain trust. The host jokingly anticipated agents breaking up with partners, highlighting the risks of over-presumptuousness. Looking ahead, Paulin suggested consumer agents would increasingly move towards proactivity and hyper-specialization, aligning with the host's "narrow startups" thesis – building valuable software for small, specific user groups. They also discussed the distinction between "assistant" (performing directed tasks) and "agent" (possessing agency to initiate actions). The podcast then envisioned a future where agents profoundly impact humanity, automating mundane tasks to free up time for family and friends, leading to a "consumer-aligned internet." They explored agent-to-agent interactions, foreseeing a new paradigm for commerce and services. The example of Amazon blocking Muse while Shopify embraced it illustrated the shift: Amazon fears disintermediation and loss of ad revenue from agent-driven purchases, while Shopify sees agents as democratizing commerce for individual sellers. They also mused about how agents might disrupt supply-constrained markets like restaurant reservations, potentially leading to loyalty-based access or bidding wars. Finally, they addressed the economics, noting that 65 out of 122 observed agents were paid. This contrasts with dominant, free players like Muse and Instinct, posing a challenge for long-tail startups. They hoped that falling browser use costs could make more products free, but also expressed interest in high-value agents that consumers would be willing to pay significant sums for, indicating "possibility expansion" rather than mere cost reduction.