Listen Labs AI Sinks Research Costs to Predict Customer Intent

Listen Labs 利用 AI 降低调研成本并预测客户意图 AI 语音访谈比传统选择题调查具有更高的一致性和更丰富的预测准确性，因为开放式的口头推理能揭示人类的真实意图。 Alfred Wahlforss 详细阐释了客户调研从传统人工焦点小组向 AI 原生并行访谈及生成式智能体模拟的演进历程。他解释了 Listen Labs 如何通过深度对话文本记录实现高预测准确率，这种方式比选择题问卷或行为日志能更有效地挖掘出真实且不带评判色彩的客户反馈。Wahlforss 说明了为何他们将 80% 的工程精力投入于受众招募与细分，以构建不可或缺的“人类 API”。他最终认为，随着 AI 让产品开发变得轻而易举，掌握用以决定“该构建什么”的洞察闭环将成为终极的竞争护城河。 Listen Labs 打造了面向 3,000 万受众群体同时并发数千场语音访谈的 AI 客户调研平台，其 80% 的工程资源均集中在受众挖掘、资格筛选与人群细分上。 (0:00) 测试显示受访者在 AI 语音访谈中大声推理作答时的一致性远高于传统单选问卷，且面对无偏见、异步的 AI 往往比面对真人访谈员更为诚实。 (3:00) 平台采用双向音视频交互并借助面部表情与眼动追踪捕捉情绪细节，所有提炼观点均支持精确回溯至对应时间戳的原始视频，有效避免 AI 幻觉。 (5:00) 进入“市场调研 3.0”，Listen Labs 基于个体深度访谈构建生成式 Agent 模拟，对未来回答的预测准确率最高达 95%，并证实访谈文本比消费账单或行为日志更能预测用户决策。 (20:00) 随着自动化编程让构建功能变得轻而易举，Listen Labs 致力于成为连接真实用户偏好的“人类 API”，帮助自治 Agent 与企业高效决定该构建什么产品。 (25:00)

Alfred Wahlforss, co-founder and CEO of Listen Labs, argues that as AI makes product building trivial, knowing what to build has become the primary bottleneck in business. He contends that AI-driven voice interviews and generative agent simulations can overcome the systemic biases of traditional surveys to provide continuous, high-fidelity customer intelligence. Overcoming Survey Bias Through AI Voice Conversations (0:03) Traditional market research often fails because respondents exhibit selection bias and provide inconsistent answers on static, multiple-choice surveys. Wahlforss notes that internal consistency tests showed participants were significantly more consistent when forced to reason through answers aloud during AI-led voice interviews. Listen's two-way video calls track facial expressions and eye movements to detect emotional resonance, which correlates more reliably with marketing performance than standard Likert scales. Counterintuitively, participants frequently report preferring AI interviewers over humans. The asynchronous, non-judgmental interface creates an environment where respondents feel comfortable being candid about sensitive topics, eliminating the scheduling overhead and social friction typical of legacy focus groups. Sourcing Niche Demographics and Bypassing CRM Limits (14:48) Listen directs 80% of its engineering effort toward audience sourcing, segmentation, and qualification across a 30-million-person panel, with long-term plans to scale to 1 billion. This targeting accommodates power-law customer distributions, such as Sweetgreen deriving an estimated 80% of revenue from high-income urban users concerned with niche topics like seed oils. Automated screening mitigates the severe participant drop-off that legacy panels experience when qualifying low-incidence demographics. Wahlforss explains that external research panels bypass the deliverability risks and privacy constraints of corporate customer relationship management (CRM) systems. Major tech and consumer brands cannot safely email their entire customer base for feedback without harming email deliverability. A neutral third party also enables comparative testing between active power users and prospective non-customers. Generative Agent Simulations for Predictive Testing (20:00) Listen is expanding from parallel automated interviews into generative agent simulation, which Wahlforss terms "Market Research 3.0." By conducting an hour-long qualitative interview with a single participant, the platform captures open-ended behavioral data that Wahlforss claims can predict that individual's future survey responses with up to 95% accuracy. Simulated panels allow companies to conduct rapid message testing, such as evaluating conference titles or advertising copy, without recruiting fresh human cohorts for every iteration. Wahlforss notes that general-purpose large language models (LLMs) like ChatGPT fail these tests because they represent an aggregate average person rather than the hyper-specific niches that drive real-world purchasing decisions. Human APIs, Moats, and Developer Workflows (25:00) Listen envisions its platform operating as a "human API" that autonomous coding agents query via protocols like Anthropic's Model Context Protocol (MCP) to validate product decisions before writing code. Wahlforss expresses skepticism toward multi-agent simulation debates due to compounding unpredictability from chaos theory, choosing instead to focus on the predictive accuracy of individual persona models. To protect against commoditization, Wahlforss points to vertical AI moats centered on proprietary evaluation benchmarks, 30-million-user panel network effects, and rigorous interview methodology. Rather than prompting vanilla LLMs with questions that risk leading the witness, the system structures interview guides around established research standards, retaining direct links from synthesized claims to timestamped source video. Wahlforss maintains that human telemetry will remain essential even in an advanced AI economy because human behavior and cultural trends remain chaotic and irrational. An ongoing technical challenge involves advancing evaluation benchmarks: while early instruction-following scores improved from 20% to 85%, newer benchmarks for complex tasks like screen-recording analysis have reset to 20% baseline performance.

Listen Labs 联合创始人兼首席执行官 Alfred Wahlforss 认为，在通用人工智能与自动化智能体使产品构建变得低成本的背景下，明确“应该构建什么”正成为企业最核心的战略瓶颈。该公司正在构建能够并发执行海量语音访谈与生成式模拟的 AI 平台，旨在将传统低效、易失真的定性调研转变为实时、高保真且具预测性的客户行为数据系统。 重构定性调研：AI 视频语音对话与可追溯洞察 (5:00) 传统市场调研通常依赖僵化的问卷工具（如 Qualtrics）或耗资巨大的线下焦点小组，不仅流程拖延数月，还常因“你怎么看/愿意付多少钱”等诱导性问题导致样本失真。Wahlforss 指出，Listen Labs 的内部一致性测试显示，受访者在面对相同选择题时的回答往往高度前后矛盾；然而，当要求受访者在语音访谈中大声阐述推理过程时，其观点一致性会显著提升。此外，与静态量表相比，系统在两分钟双向视频互动中捕捉的面部表情与眼动微表情，能更真实地反映受访者的潜在情绪，与 Meta 和 LinkedIn 等渠道的实际广告投放效果高度吻合。 为消除企业对模型幻觉的担忧，平台引入了全链路数据溯源机制：企业客户点击分析报告中的任何结论或引用，均可直接跳转至对应时间戳的原始音视频片段。例如，服装品牌 Chubbies 曾通过 Listen Labs 发现用户胸毛会与某款衬衫面料产生摩擦不适，随后据此重新设计了产品，数月后的实际销售增长验证了该定性预测的准确性；男士理容品牌 Manscaped 亦根据该平台的反馈调整了超级碗广告策略。 消除受访阻力：机器对话的诚实红利与垂直受众沉淀 (9:50) 相较于真人访谈，受访者往往更倾向于与 AI 交流。Wahlforss 解释称，异步访谈消除了复杂的日程协调压力，更重要的是，受访者面对非评判性的 AI 智能体时感到更安全，愿意表达出近乎“残酷的诚实”。这种心理机制使得该模式在青少年调研等易因社交压力导致言不由衷的敏感场景中表现尤为突出。 在受众获取端，Listen 将 80% 的工程力量投入于受众寻访、细分与资格审查。传统调查面板常因严格的资格筛选导致极高的弃答率，而 Listen 目前管理着 3000 万规模的受访者池（长期目标扩至 10 亿），并通过多轮互动建立持续更新的画像档案（例如受访者随口提及球鞋爱好便会被打上相应标签）。相比于企业直接使用 CRM 群发可能触发的反垃圾邮件机制或送达率惩罚，中立平台既能保护企业声誉，又能精准触达极度细分的权力法则受众——例如 Sweetgreen 80% 的收入依赖关注“种子油”等极小众议题（仅占总人口约 1%）的都市高收入拥趸，亦或是分布于多个国家的专科医生。 迈向市场研究 3.0：生成式模拟与 Human API (20:00) Wahlforss 将市场调研的演化划分为三个阶段：1.0 时代依赖逐一人工致电与手动汇总；2.0 时代通过 AI 原生平台并行开展数千场语音访谈并自动提炼洞察；正在演化中的 3.0 时代则是基于深度访谈文本的“生成式智能体模拟”。Listen 的实测显示，深度访谈转录本包含了大量开放式旁白与行为追问，其预测价值远超信用卡账单或行为日志；系统仅需对单一目标进行一小时的深度访谈，就能以高达 95% 的准确率预测该对象对未来新问题的回答偏好。在演讲标题或广告文案的测试中，基于垂直客群模拟的筛选结果显著优于通用大模型，因为后者仅代表“平均大众”的认知。 基于该技术，Listen 推出了“增强响应”（Augmented Responses）功能，使之能够作为一种“Human API”供自主编码智能体或自动化组织直接调用，以在功能上线或资本投入前即刻获取用户偏好反馈。针对当前行业探讨的“多智能体互动辩论模拟”，Wahlforss 表达了审慎态度，认为混沌理论下多重交互的复合误差难以预测，因此系统当前聚焦于单个人物画像的高精度建模与代表性样本扩展。 决定造什么：战略决策闭环与垂直领域护城河 (30:00) 在现代商业闭环中，核心环节分为“决定造什么”与“实际造出来”。Wahlforss 指出，编码智能体正在指数级缩短构建软件的时间，物理硬件企业的试错成本更居高不下，因此 Listen 定位为“决定造什么”的基础设施。他援引商业史指出，宝洁通过用户调研发现人们对黏稠洗衣液的处理痛点从而推出汰渍洗衣凝珠（Tide Pods），玛氏（Mars）在 20 世纪 50 年代通过调研发现原为士兵防化设计的 M&Ms 在儿童防弄脏家具场景下的巨大商机，均证明了从非理性人类诉求中挖掘小众刚需的商业杠杆。 关于垂直 AI 企业的壁垒，Wahlforss 强调了专有评测集（Evals）的作用：其团队在早期 GPT-4 阶段将基础遵循度从 20% 优化至 85%，当前针对屏幕录制交互及跳过无关问题的高级评测则建立了新的质量门槛。此外，平台已通过模型上下文协议（MCP）接入 Anthropic 的 Claude，使用户能够与 Listen 进行持续的概念头脑风暴，推动调研从对既有创意的被动评估，演变为访谈中途根据用户痛点动态生成原型图像的“实时产品共创”。 Wahlforss 认为，人类行为的非理性特征以及受文化与社交媒体趋势影响的随机性，决定了即使在 AGI 时代，AI 决策依然需要人类的实时反馈作为校准遥测。然而，如何确保生成式模拟在脱离实时活体样本的情况下长期抵御漂移，以及单体高准确度模拟在多变量复杂社会决策中能否保持稳健，仍是该技术路径未来面临的持续考验。

## Background 本期内容由红杉资本（Sequoia Capital）于 2026 年 6 月 2 日在 RSS 及网络播客平台发布。节目围绕 Listen Labs 联合创始人兼 CEO Alfred Wahlforss 展开对谈，介绍其构建的 AI 客户调研 Agent 如何依托 3000 万人受众网络并发开展数千场真实语音访谈，并阐述为何访谈数据比传统行为日志更能精准预测客户决策，以及在 AGI 降低构建门槛后“弄清该造什么”如何成为最稀缺的战略闭环。 ## TL;DR 对谈系统拆解了 Listen Labs 如何颠覆传统问卷调查与高成本咨询，利用双向音视频 AI 对话在 5 分钟内召集精准人群并收集高真实度、无评判压力的深度反馈。创始人 Alfred 阐明了公司将 80% 的工程资源投入到受众精准画像匹配的原因，并展示了平台如何基于海量真实访谈记录打造生成式 Agent“模拟预测”（Simulation）与“人类 API”。节目进一步展望了商业闭环的未来重构：随着编码 Agent 让产品落地变得轻而易举，“与用户交谈、明确构建方向”的策略侧成为垂直 AI 最具价值的护城河，调研也正从被动打分升级为由 AI 驱动的实时概念协同共创。 ## Atmosphere & style 整场对话兼具一线创业者的务实敏锐与深度的商业哲学思辨，语调轻松幽默却直击行业痛点。从男装衬衫摩擦胸毛、超级碗广告调整等接地气的具体案例，层层递进推演至因果推断、多 Agent 混沌效应与企业护城河构建，展现出清晰的技术推演逻辑。主持人的犀利追问与 Alfred 的坦诚作答形成良好共振，不仅生动还原了初创团队自救探索的生机，更让听众切实感受到垂直应用层 AI 正在重塑现代商业决策范式的爆发力。 ## Bulletin - `00:00` Listen Labs 打造由 AI Agent 驱动的并发访谈平台，当前覆盖 3000 万真实受众池并计划拓展至 10 亿人，成立仅一年已渗透微软、Anthropic、Sweetgreen 等 20 家《财富》500 强与头部科技企业。 - `12:15` 公司将 80% 的工程资源投入到受众画像与招募网络建设上，利用幂律受众识别与跨访谈画像沉淀，攻克传统调研筛选发生率低、触达潜在客户困难及大企业 CRM 反垃圾合规等痛点。 - `20:00` 基于海量深度访谈对话数据，Listen Labs 研发生成式 Agent 模拟预测引擎（Simulation），在细分垂直人群偏好测试中准确率可达 95%，并提供可供下游 Coding Agent 直接调用的“人类 API”。 - `30:00` 随着 AGI 推动代码编写与工程实现的成本与门槛趋近于零，企业的核心稀缺资产正在转向策略侧的“明确该造什么”，形成“倾听洞察—策略生成—自动编码交付”的闭环。 ## Key takeaways - `00:00` Listen Labs 端到端闭环涵盖自动生成访谈大纲、定向招募精准人群、AI 并发数百场 1 对 1 语音视频访谈，以及最终的跨对话洞察提炼与未来行为仿真预测。 - `10:00` 受访者在面对无预设立场且不带评判色彩的 AI 时往往展现出更高的坦诚度，且碎片化异步对话极大降低了受访门槛与企业报酬成本。 - `03:40` 单选问卷的重测一致性极差，而 AI 深度多轮探问要求受访者完整推演因果逻辑，能产出一致性更高、更可信的用户输入。 - `05:00` Listen 采用类似 Zoom 的双向视频交互，实时追踪受访者的面部微表情与眼神兴奋度，以此更精准地预测真实效果营销与购买转化，同时提供全链路可追溯引述以杜绝 AI 幻觉。 - `12:15` 企业绝大部分收入由符合幂律法则的核心垂直用户贡献，Listen 将 80% 的工程精力用于跨访谈受众标签沉淀，以高精度召回极度稀缺的小众人群。 - `12:15` 企业无法直接在内部 CRM 中探寻未转化的潜在客户，且面临混乱的数据和严格的邮件垃圾拦截机制，外部智能化受众池成为刚需。 - `15:47` AI 访谈将促使麦肯锡、贝恩等传统咨询服务解耦，虽然单次调研交付成本显著下降，但全行业的用户洞察需求总量将迎来两个数量级的爆发式增长。 - `22:15` 相比信用卡消费流水和行为日志，高质量的深度访谈是训练模拟人类思维轨道的最佳基准数据集，能够完整补全细分利基客群的因果逻辑链条。 - `22:15` 在特定垂直利基场景下，通用 ChatGPT 的大众平均人语料与粗糙 Prompt 人设无法准确预测垂直决策，而基于真实访谈沉淀的定制模型表现显著更优。 - `25:12` 构建面向真实人类偏好的“人类 API”，允许下游 Coding Agent 等自主 Agent 在编码、做产品规划或战略决策时直接请求调用并对齐真实意图。 - `27:08` 针对多 Agent 自主辩论可能引发的混沌理论失控风险，Listen 采用精细化个体建模结合“AI 委员会”裁判机制，确保模拟预测的可控性与稳定性。 - `27:08` 垂直 AI 企业的真正壁垒在于持续迭代针对垂直业务流的严苛评估体系（Evals），不断解决模型重复提问、屏幕录制感知与多模态逻辑跳转等深水区难题。 - `30:00` 未来可能出现仅需极少员工的十亿美元公司，核心由左侧负责客户倾听与需求提炼的策略 Agent，与右侧负责自动写代码及交付的执行 Agent 共同闭环驱动。 - `35:00` 人类本质上充满非理性与混乱变化，AI 逼近人类智能渐近线过程中未能完全覆盖的那个微小“差值”（Delta），是未来最稀缺、最决定胜负的资产。 - `38:00` 借助 MCP 扩展接口与多模态即时生成能力，市场调研正从过去事后拿方案请用户评价的单向模式，演进为实时生成概念图与用户协同共创的原型工作流。 ## Detailed walkthrough ### `00:00` Listen Labs 的愿景与核心工作流：从十亿受众到仿真预测 Listen Labs 创始人兼首席执行官 Alfred Wahlforss 介绍了公司的宏大目标：构建一个规模达到 10 亿人的全球受众库，并对每个人的专业领域进行深度分层与画像沉淀。即使是在像球鞋（sneakers）这样垂直的消费细分赛道中，人群里也存在大量具备潮流引领力（influencers）和早期尝鲜特质（early adopters）的核心用户。传统的调研很难精准捕捉到这群人，而 Listen 能够在每一次访谈中持续沉淀受众认知，根据过往累积的跨领域对话动态建立完整的受众档案（profiles），从而通过算法随时定向检索并触达最匹配的目标人选。 成立仅约一年，Listen Labs 已经成长为服务微软（Microsoft）、Anthropic、Sweetgreen、NBC 环球（NBC Universal）等 20 家《财富》500 强企业和前沿科技先锋的 AI 原生客户研究平台。在解释“用 AI 规模化赋能市场调研”的实际含义时，Alfred 表示其核心是打造一个比企业自身“更懂其客户”的 AI Agent。具体的工作流程分为端到端的几个步骤：首先，企业提出一个具体的业务困惑（例如：“如何优化 Cursor 的新手引导流程？”）；系统自动生成结构化的访谈大纲（interview guide）作为 Agent 的行动指令；紧接着依托平台现有的 3000 万真实受众库，在几小时内定向召集从肿瘤科医生（oncologist）到软件工程师等各类群体；AI 并发进行数百场一对一深度访谈并提取洞察与建议。 Alfred 透露，Listen Labs 即将在两个月后推出终极功能——“仿真预测”（simulation）。当平台内沉淀了来自某一品牌或领域的成千上万场深度访谈数据后，系统将能够直接模拟并预测客户对未来潜在问题的反应。正如 Alfred 对 AI 时代的底层判断：“随着我们越来越接近 AGI（通用人工智能），构建产品的工程门槛会变得极低，而真正稀缺且艰难的环节将变成‘弄清楚到底该造什么’（knowing what to build）。这正是 Listen 所专注并试图牢牢掌控的闭环。” ```mermaid flowchart LR accTitle: Listen Labs 调研与仿真预测全流程 accDescr: 展示从提出业务问题到生成大纲、受众招募、并发访谈、洞察生成及未来仿真预测的完整闭环。 subgraph Input["1. 业务输入与规划"] Q["业务问题<br/>(例如：如何优化 Cursor 引导？)"] --> Guide["AI 自动生成<br/>访谈大纲 (Guide)"] end subgraph Execution["2. 受众匹配与并发访谈"] Audience["3000 万受众网络<br/>(医生、工程师等)"] --> Agent["AI Agent 并发语音访谈<br/>(数百场同时进行)"] Guide --> Agent end subgraph Insight["3. 洞察生成与预测闭环"] Agent --> Analyze["跨对话数据分析<br/>与业务改进建议"] Analyze --> Simulation["仿真引擎 (Simulation)<br/>预测未来客户反应"] end ``` - **核心受众规模**：当前受众池已覆盖 3000 万参与者，长期战略目标是拓展至 10 亿人。 - **商业化表现**：上线仅一年左右，已打入 20 家《财富》500 强企业，并被 Anthropic、Cursor 等前沿 AI 团队深度使用。 --- ### `02:30` 真实客户洞察：从男装胸毛痛点到超级碗广告 讨论进入具体的客户落地场景时，Alfred 列举了几个极具反差感且生动的品牌案例。男士休闲服饰品牌 Chubbies（以短裤闻名）是 Listen 最早期的客户之一，他们将 Listen 应用于全方位的营销测试以及衬衫产品的面料与版型评估。在一次产品测试中，调研收集到了一个极其具体却极度影响复购的细节：他们某款衬衫的特定面料会与男性的胸毛（chest hair）发生摩擦拉扯，导致穿着体验极其不适。Chubbies 迅速根据访谈反馈调整了衬衫面料，瞬间让舒适度实现了质的飞跃。 除了 Chubbies，男士理容品牌 Manscaped 也借助 Listen 的访谈洞察，对他们价值连城的超级碗（Super Bowl）广告进行了关键调整；知名塑形内衣品牌 Skims 同样是其标杆客户。主持人幽默地调侃 Listen 的利基市场仿佛是“从男士体毛修剪一路做到服饰面料”，Alfred 笑着认可了这个接地气的定位，并强调正是这些往往在传统大颗粒度问卷中被完全忽略的细微体验，决定了消费品在现实商业环境中的成败。 --- ### `03:40` 问卷失真与数据遥测的争论：为什么 AI 深度对话更具一致性？ 主持人提出了一个长久以来对市场调研的根本性质疑：传统的问卷调查（surveys）往往存在严重的样本选择偏差（selection bias），因为填问卷的人大多是为了拿小额报酬；更致命的是，“人们嘴上说会做什么”与“他们在真实世界中实际的行为”之间存在巨大鸿沟。主持人直言自己倾向于“真实数据遥测派”（telemetry），即只相信真实行为日志和 A/B 测试，对用户自述持深度怀疑态度，并询问 AI 如何跨越这道信任鸿沟。 Alfred 直面了这一尖锐问题，并分享了 Listen 的对照实验结果。他们曾让同一批受访者在不同时间重复填写相同的单选问卷（multiple choice surveys），结果发现受访者的答案呈现出惊人的不一致性与随机性；但在使用 Listen 的 AI 深度对话时，由于系统会进行多轮探问，受访者必须真正调动思维、完整梳理推演自己的逻辑（reason through your answer），这使得他们在回答同一问题时表现出了显著更高的一致性与真实度。 谈及与真实销售的对齐，Alfred 指出 Listen 会对 Chubbies 等客户的调研预测进行长达数月的真实销售数据回溯校验。他坦言，A/B 测试固然是业务验证的“圣杯”（holy grail），但在现实场景中，运行严谨的 A/B 测试不仅周期冗长，而且需要极其庞大的用户体量作为统计学支撑，对于大多数没有海量流量的决策场景而言，完全没有客户输入直接盲猜，远不如拥有一套经过强推理验证的对话输入来得有效。 --- ### `05:00` 双向视频面试体验与微表情捕捉：如何做到可追溯与可量化？ 针对 Listen 的具体交互形式，Alfred 澄清这绝非冷冰冰的“文本转语音”或填表工具，而是一个端到端的“双向视频通话”（two-way Zoom-like call）。受访者直接在视频环境中与 AI 虚拟面试官面对面交流。在这个过程中，AI 不仅能听懂语音内容，还能实时捕捉受访者的面部微表情和眼神动向（looks at your eyes）。 这种多模态情绪感知能力极大地缩短了“口头说辞”与“真实内心”之间的差距。尤其在广告创意测试中，这种差异格外明显：在传统的 5 分制李克特量表（Likert scale）上，受访者点选“极有可能购买”往往只是随手一按；但在视频中，受访者眼神中展现出的自发兴奋与高能量反应，才是该创意能否在 Meta 或 LinkedIn 的效果营销（performance marketing）中带来高转化的决定性指标。 对于采购 Listen 服务的企业客户，透明度和可信度同样至关重要。为了彻底消除企业对“AI 幻觉（hallucinations）”以及黑盒总结的担忧，Listen 围绕“全链路可追溯性”（traceability）构建了整个分析后台。客户在报告中点击任何一个结论或数据切片，系统都能直接反查到对应受访者的原始音视频切片或逐字稿引述，做到每一句洞察都言之有据。 | 调研维度 | 传统单选问卷 (Multiple Choice Surveys) | Listen Labs AI 视频双向访谈 | |:--- |:--- |:--- | | **交互形式** | 静态单选 / 李克特 5 分量表（易敷衍勾选） | 类似 Zoom 的双向音视频对话，由 AI Agent 动态深挖 | | **思维深度** | 无需深度思考，重测一致性极差 | 必须给出因果逻辑并推演，态度一致性高 | | **情绪验证** | 仅有干燥的数字分数，无法辨别真实意愿 | 实时追踪眼神与面部表情，精准捕捉真实兴奋度 | | **广告预测度**| 分数虚高，难以预测真实点击转化 | 情绪反应与 Meta / LinkedIn 效果广告转化强相关 | | **可追溯性** | 扁平的汇总百分比，背后原因不透明 | 任意结论一键反查原声视频片段与原始引用，杜绝幻觉 | --- ### `06:40` 灵感起源与旧工业范式的局限：从 BeFake 病毒式裂变到千亿调研咨询 Listen Labs 的诞生并非源于宏大的商业企划，而是来自创始人此前一次切肤之痛的创业自救。Alfred 与合伙人曾开发过一款名为“BeFake”的消费级 AI 图像社交应用（类似于 ChatGPT 图像生成与 Fine-tuned Stable Diffusion 的早期结合体），允许用户生成逼真的 AI 化身。该应用一经推出便迎来爆发式病毒传播，一夜之间涌入超过 20,000 名用户。然而随着流量激增，团队遭遇了极高的用户流失率（churn）。为了搞清楚用户究竟如何使用、为何流失、以及如何重构产品定位，两位创始人自制了一个能够自动向用户打电话访谈的早期 AI 原型工具。正是这个解决自身生存问题的自研脚本，展现出了惊人的业务洞察力，促成了 Listen Labs 的诞生。 在梳理市场格局时，Alfred 剖析了传统调研产业的割裂与低效：一边是像 Qualtrics 这样较为陈旧的纯问卷软件工具；另一边则是一个规模高达数百亿美元的人工咨询服务产业。对于大企业而言，访谈现有用户尚且不易，触达并访谈“潜在客户”（prospective customers）更是难上加难。 传统市场研究之所以高度依赖昂贵的人力，是因为它包含极其严苛的学术与实操门槛： - **提问的科学性**：提问本身是一门高深的方法论学问。你绝不能直白地问用户“你愿意为这个付多少钱”，因为这根本无法反映真实付费意愿，必须借助复杂的阶梯测试与情境投射模型。 - **样本采购的高门槛**：如何在全球范围内搜寻、清洗并组织符合特定画像的合格受访者，本身就是一个巨大的中介黑盒。 - **分析效率低下**：在宝洁（P&G）等快消巨头或微软这样的科技公司内部，每年需要耗费数千万美元把人们召集到实体房间里开展传统焦点小组（focus groups），随后分析师需花费数周逐一人工听录音。 Alfred 指出，AI 原生调研平台正在彻底重构这一切。相比于旧范式耗资巨大、耗时漫长且往往仅能服务于极少数重大战略决策，Listen Labs 不仅将成本和交付周期压缩了数个数量级，更从底层改变了企业的决策机制——在过去，大多数日常产品与运营决策因为无法承受焦点小组的高昂代价，实际上是在没有任何真实客户输入（customer input）的情况下拍脑袋做出的；而 Listen 正在将这种高质量的深度输入常态化。 ### `10:00` 异步 AI 访谈的魔法与非评判性倾听 直接与客户交谈在传统模式下存在极高摩擦力，而一旦降低了这道沟通门槛，企业就能做出明智得多的决策。Alfred Wahlforss 指出，Listen Labs 带来的核心优势首先是惊人的速度：短短 5 分钟内，企业就能收到来自真实人类的反馈，看到数百名受访者同时涌入访谈界面，这种体验甚至带有一种“魔法般”的震撼感。 更出人意料的是背后的经济学与心理学机制。很多人好奇普通用户是否愿意接受 AI 的访谈，客观事实不仅是“愿意”，而且企业付给受访者的报酬甚至可以远低于真人同步访谈。真人访谈需要约定固定时间、占用大块精力，而 AI 访谈是完全异步的，受访者在碎片化时间随开随停，心理压力几乎为零。更深层次的发现是：人们在面对 AI 时往往比面对人类更诚实。面对一个没有预设立场、不带评判色彩、却对自己表现出浓厚兴趣的非人类实体，受访者往往会完全敞开心扉，甚至觉得访谈具有某种“心理疗愈”的作用。这使得 Listen 能够顺利切入诸如“儿童对新产品的真实反应”等传统上极其敏感、复杂的对话领域。 ```mermaid flowchart TD accTitle: 传统同步访谈与 Listen Labs 异步 AI 访谈对比 accDescr: 对比传统访谈的高摩擦、高成本与 Listen Labs 异步访谈的低门槛、高诚实度与快速扩展性 subgraph TRAD["传统真人访谈模式"] T1["协调排期与严格筛选"] --> T2["高额受访者报酬与面对面心理压力"] T2 --> T3["受访者易迎合、耗时漫长"] end subgraph AI["Listen Labs 异步模式"] A1["受众画像精准匹配与即时分发"] --> A2["5分钟内数百人异步语音作答"] A2 --> A3["无评判环境触发极端诚实反馈"] end ``` - **5 分钟快速回流**：发起调研后 5 分钟即可开始看到数百名真实用户的详细语音输入。 - **更低报酬成本**：由于纯异步、无排期压力，受访者愿意接受显著低于传统真人定性访谈的费用。 - **心理安全感带来的坦诚**：非人类实体的非评判属性让用户毫无防备，展现出最真实、近乎残酷的坦率。 --- ### `12:15` 80% 的工程投入：幂律受众与跨访谈画像网络 谈到儿童等难以触达的市场时，主持人询问痛点究竟在于法规限制还是排期困难。Alfred 解释说，除了需要家长同意、孩子们上学及课外活动繁多导致无暇配合之外，最核心的挑战在于“找到真正对的那群孩子”。Listen Labs 内部将高达 80% 的工程资源全部投入到“构建精准受众”上。每个公司的商业模式本质上都由客户群体的幂律法则驱动。例如 Sweetgreen 沙拉，表面上似乎是大众产品，但核心用户其实高度集中在都市、高家庭收入、以女性为主的客群，且这些人往往需要懂得什么是“种子油”（seed oils）——而全社会大概只有 1% 的人关注这个概念；正是这群每天都吃 Sweetgreen 的重度用户，贡献了该品牌 80% 的收入。 为了精准捕捉到这类关键人群，Listen Labs 的长远目标是将受众池扩展至 10 亿人，并建立能够精准分层、识别用户在各个细分领域专业度的系统。平台具备跨访谈的持续学习能力：一个人可能在某场完全不相关的访谈中随口提了一句“我是个超级球鞋迷”（I'm a total sneakerhead），该标签就会沉淀在系统的动态档案库中；当 Nike 等品牌要发布新产品线时，系统便能瞬间召回这个人。传统研究机构通常采用极其粗暴的邮件群发，不仅体验糟糕，其筛选发生率（Incidence Rate）常常低至 10%（即每 10 个人中只有 1 人符合资质），导致受访者因连续被淘汰 10 次却拿不到一分钱而极速流失。 针对品牌为何不直接利用自身现有 CRM 系统的疑问，Alfred 指出： - **触达潜在客户（Prospects）**：企业固然可以通过接入 Listen 与自身存量客户对话，但品牌无法在自己的 CRM 里找到尚未成为用户的群体，也无法探究边缘用户与核心幂律用户之间的差异。 - **CRM 混乱与合规封禁风险**：现实中多数企业的 CRM 数据杂乱无章；更严重的是政策与反垃圾机制的制约——即便是 Google 内部团队，也绝对不能直接私自向 Gmail 用户群发邮件。盲目展开大规模冷启动调研，极易被各大邮件服务商列入垃圾邮件黑名单，陷入“微软炼狱”（Microsoft purgatory）。 --- ### `15:47` 咨询行业的解耦、高上限与模拟（Simulation）的未来 对于麦肯锡、贝恩这类依赖制作数百页 PPT、访谈数千人换取高额佣金的传统咨询巨头，AI 的介入是否会让他们失业？作为曾雇佣过这些咨询机构的前从业者，主持人与 Alfred 展开了热烈讨论。Alfred 透露贝恩本身就是 Listen 的重要客户，他们正通过 Listen 来加速传统工作流程。传统咨询在落地推行变革层面依然拥有独特价值，但行业的传统服务毛利率势必大幅下滑，迫使咨询公司将业务“解耦”（unbundle），把大量过往由初级分析师人工解决的问题转交给 AI Agent。 与单纯的“价格压低”不同，Alfred 认为用户洞察的需求上限极高。以往因为成本和精力限制，用户连鞋带系法这种细枝末节的偏好都无处反馈；而现在企业可以把原本用于单一项目的预算投入到对五到十个全新扩张领域的探索中。关于经济盈余的分配，Listen 曾做过一个面向 8 个国家、仅访谈 20 位专科医生的项目，收费依然高达数十万美元——因为速度的大幅提升和极端稀缺的触达能力，让快速交付拥有了极高的溢价空间。虽然单次访谈在长期来看成本会变低，但调研的总量将迎来两个数量级（2 orders of magnitude）的爆发式增长。 更加令 Alfred 兴奋的前沿方向是他们正在构建的“客户模拟”（Simulation）。它能解锁那 99% 过去根本来不及找真人交谈的边缘场景。例如在医疗系统内部，医生极度繁忙，平时绝不可能特意抽出时间参加冗长繁琐的焦点小组，导致电子健康记录（EHR）或手术室系统中的成千上万个微小体验缺陷（paper cuts）长期无人问津；通过手机端碎片化的异步反馈甚至模拟推演，这些关键痛点能被迅速捕捉。听取客户建议不仅止于生成报告，Alfred 指出，真正的闭环是让系统在获取这些细碎问题的瞬间，直接调度另一个下游 Agent 去自动修复问题，例如从流失客户的访谈中实时提炼原因并触发针对性的挽回与产品修正工作流。 ### `20:00` 市场研究 3.0：从 AI 访谈到生成式 Agent 模拟 行业的发展轨迹正清晰地展现出代际跨越：市场研究 1.0 时代依赖人工逐个拨打电话访谈上百人，再耗费大量时间手工整理归纳；2.0 时代进入 AI 原生阶段，模型能够自主设计提问路径、同时并发访谈数千名用户并自动生成洞察提炼。如今，行业正在向“生成式 Agent 模拟”（Generative Agent Simulation）的 3.0 阶段迈进。虽然自动驾驶等领域早已证明合成数据的巨大威力，但外界对于模拟真实用户也普遍抱有天然的怀疑：单纯依靠合成数据，是否只是在对基础模型预训练语料进行排列组合与重混？其中是否真的存在增量信息（Alpha）？ ```mermaid flowchart TD accTitle: 市场研究的代际演进路径 accDescr: 展示从 1.0 人工访谈到 2.0 AI 并发访谈再到 3.0 生成式 Agent 模拟的进化过程 subgraph S1["市场研究 1.0"] A1["人工逐一电话访谈 100 人"] --> A2["人工整理与归纳汇总"] end subgraph S2["市场研究 2.0 (AI 原生)"] B1["AI 动态设计提问轨道"] --> B2["数千人并发多轮语音访谈"] --> B3["AI 自动结构化提炼洞察"] end subgraph S3["市场研究 3.0 (生成式模拟)"] C1["深度访谈构建个体数字画像"] --> C2["合成人群模拟与预测评估"] --> C3["为下游决策与 Agent 提供 API"] end S1 --> S2 --> S3 ``` Listen Labs 打造模拟体系的切入点并非虚构数据，而是从极度细致地对单个个体进行深度访谈开始。Alfred Wahlforss 解释道，如果对一个人进行长达一小时的深入交流，AI 就能在相当程度上精准预测对方在各种情境下的偏好与行为。大语言模型非常擅长捕捉这种高维度的个人特征；当系统将尽可能详尽的个体信息输入模型后，在某些特定问题的回答预测上，准确率甚至能达到 95%。 --- ### `22:15` 文案测试实战与真实访谈数据的不可替代性 在模拟预测的落地场景中，最直接且高频的需求之一是信息与文案测试（Message Testing），例如评估户外广告牌上的标语效果。Alfred 分享了自己亲身经历的一个案例：他在筹备一场行业大会的演讲时，为了提高听众到场转化率，需要确定一个最具吸引力的演讲标题。他利用自己客户群体的数据创建了一个模拟受众池（Panel），一口气生成了 100 个不同的备选标题输入系统进行模拟打分，结果排名第一的标题得分比第二名高出整整一倍。Alfred 直言，即便无法 100% 验证其终极正确性，这种“将决策辅助外包”的指引在实际操作中极具价值。 为了验证模拟系统的有效性，他们将同样的问题输入通用 ChatGPT，并用 Alfred 过往一次不太成功的演讲标题与另一个更成功的竞品标题进行对照测试。通用 ChatGPT 做出了错误的判断，而 Listen Labs 的模拟系统则准确挑出了真正成功的那个标题。通用模型是在海量“大众平均人”的语料上训练出来的，而商业决策往往需要面向极度细分的垂直利基市场（Niche）。 针对“为何不能直接在通用 ChatGPT 中通过 Prompt 设定人设（例如‘设定你是一位 35 岁、脾气暴躁、习惯用终端的软件工程师’）”的疑问，Alfred 解释了背后的数据直觉： - **Prompt 设定的局限性**：预训练通用模型本质上缺乏针对特定用户画像（Persona）在现实中如何思考与行动的干净数据，仅靠角色扮演只是对平均分布的粗糙拟合。 - **访谈数据的优越性**：Listen Labs 曾尝试过输入信用卡消费记录、行为日志、购买习惯等多维度数据，但最终发现效果最好的基石数据集是“深度访谈”。 - **捕捉思维发散与因果**：访谈允许用户自由发散展开话题，能深度挖掘因果关系与底层行为动机。AI 的本质是对压缩数据的高维泛化，当在特定空间中积累了足够多高质量的真实访谈，模型便能精准“补全该类人群思维轨道的全貌”。 --- ### `25:12` “人类 API”（Human API）与 Agent 调用的未来 既然能通过成千上万次真实访谈构建出极高精度的模拟能力，Listen Labs 将这一能力产品化为“增强响应”（Augmented Responses）。客户在完成数百乃至上万次常规的 Listen Labs 真实用户访谈后，系统便能沉淀出一个专属的定制化 Bot，用户随时可以向它抛出临时性、突发性的细分问题，获得即时的模拟反馈。 这一体系更深远的价值在于它能够作为下游代码编写 Agent（Coding Agents）或其他自主 Agent 的上游输入。Alfred 描绘了未来的系统架构： - 打造一个面向真实世界用户偏好的“人类 API”（Human API）。 - 下游的各类业务 Agent 在编写代码、规划产品功能、决定投资标的或制定服务方案时，可以直接调用该 API 查询：“目标用户到底需要什么？他们最反感什么设计？”从而闭环指导开发与决策。 - 在底层技术实现上，Listen Labs 综合运用了后期后训练（Post-training）、垂直 RAG（检索增强生成）以及多项专有模型微调技术，将海量访谈对话深度融入模型系统。 --- ### `27:08` 多 Agent 辩论的混沌风险与评估基准（Eval）的持续演进 在探讨是否引入多 Agent 辩论机制（Multi-Agent System）来迭代得出更优答案时，Alfred 指出，目前 Listen Labs 的模拟方式是精确为单个用户建模，再将其横向扩展至 1,000 个具有代表性的样本池，目前并未让这 1,000 个虚拟 Agent 互相打擂台辩论。他对让 Agent 群体自主辩论的方案保持审慎：依据混沌理论，当多 Agent 交互的非线性因素不断复合叠加时，系统行为与输出结果将变得极难预测。相比之下，类似“AI 委员会”（将同一个查询发送给三个不同的 LLM，再由一个裁判 LLM 综合评估）的方法在平均表现上更为可控有效。 展望未来 2.0 与 3.0 模式的分工，Alfred 认为生成式模拟并不会彻底取代真人访谈： - **高风险重大决策**（如投放超级碗广告）：企业依然需要组织数千名真实用户进行实际访谈测试。 - **轻量级敏捷决策**（如小型路边广告牌的标语选择、日常文案 A/B 测试）：模拟系统能够极大降低决策成本并大幅拓宽使用场景。 在核心产品的模型迭代中，垂直领域 AI 公司的核心护城河在于其持续进化的评估体系（Evals）： | 发展阶段 | 核心攻坚指标（Eval 关注点） | 早期指标表现 | 攻坚后指标表现 | |:--- |:--- |:--- |:--- | | **初期基线** | 消除访谈中的重复提问，确保模型严格遵守访谈指令（解决 GPT-4 重复问同一问题几十遍的问题） | 约 20% | 攀升至 85% | | **高阶基线** | 多模态屏幕录制理解（感知用户在屏幕上的实时操作）、动态跳过已不相关的冗余问题 | 重置回约 20%（开启新一轮垂直优化） | 持续攻坚爬坡中 | Alfred 强调，随着业务深入，团队会不断构建出难度极高、极度贴合垂直业务流的全新 Eval。每当引入全新维度的能力测试，指标往往会瞬间跌回 20% 的低位，而不断跨越这些评估鸿沟，正是垂直 AI 公司的巨大价值与 Alpha 所在。 ### `30:00` 构建无限循环：从“写代码与聊客户”到自动化策略执行 主持人与 Alfred 探讨了企业运作的核心闭环：商业本质无非是“弄清要做什么，然后做出来；写代码，以及与用户交谈”（"figure out what to build, build it... Write code and talk to users"）。随着生成式 AI 与编码 Agent 呈指数级演进，**“做出来”（Execution/Build it）** 的门槛急剧降低、速度极快且成本骤减。相比之下，垂直 AI 企业的真正护城河在于掌控另一侧的策略端——通过专有的 Eval 系统不断攀登，掌握更好的数据，解决更难的问题，从而明确**“要做什么”（Strategy / Figure out what to build）**。 Alfred 展望，未来哪怕是仅有一名员工的“十亿美元公司”，也会运行在一个闭环中：左手是 Listen Labs（负责与真实客户交谈、提炼需求与策略），右手是 Coding Agent（负责写代码、修 Bug、交付上线），两者相互串联成完全自主运转的组织。即便对于大公司也是如此，在纯软件场景下，Agent 能直接与用户对话、定位 Bug、提交 PR 并合并上线；而对于宝洁（P&G）等实体原子世界（Atoms）的企业，这种前期决策的准确性尤为致命——宝洁进入一个新市场的试错成本动辄数千万美元甚至更高，一旦执行便覆水难收，因此必须在启动前百分之百把方向摸准。 谈及历史上真正擅长倾听客户的代表企业，Alfred 指出宝洁堪称市场调研机构的典范： - **宝洁与 Tide Pods（汰渍洗衣凝珠）**：宝洁通过深度客户访谈发现，用户普遍觉得倒传统洗衣液极其繁琐笨重，极度渴望一种更简便优雅的方案，最终通过访谈洞察催生了洗衣凝珠这一重磅爆款。 - **玛氏（Mars）与 M&M's 巧克力的历史转折**：正如播客《Acquired》中提到的经典案例，M&M's 在 1950 年代进行了商业史上最早的一批市场研究。该产品最初专为军方设计（“只溶在口，不溶在手”，放在行军口袋里不会融化）；但市场调研意外发现，小孩子也是极佳的受众群体，而且家长极度看重它“不会融化弄脏家具”的特性。玛氏随即全面调转广告策略，开辟了庞大的儿童消费市场。 ```mermaid flowchart LR accTitle: The Strategic Infinity Loop accDescr: 自动化企业运作的双轮闭环，从倾听用户到策略生成再到代码执行 subgraph Strategy["策略侧 (Listen Labs)"] A["真实用户访谈 / 需求洞察"] --> B["验证假设与策略决策"] end subgraph Execution["执行侧 (AI / Agent)"] C["Coding Agent / 产品研发"] --> D["自动化交付上线"] end B --> C D --> A ``` --- ### `35:00` 人性与 AI 渐近线的差值，以及垂直 AI 的护城河 在展望未来的确定性与未知数时，Alfred 坚信**人类输入（Human Input）在未来永远不可或缺**。即便未来 AGI 成为全知全能的完美理性实体，人类本身依然是充满混乱与非理性的——用户会突然沉迷于某种新产品，或是因 TikTok 上毫无征兆流行起来的迷因趋势而改变消费取向。因此，企业永远需要实时校准以捕捉这些非理性变化。至于合成数据模拟（Simulation）在预测人类行为上究竟能走多远，目前仍处于探索和验证阶段。主持人 Constantine 形象地打了个比方：商业的本质是服务人类，如果 AI 的智能正无限接近人类预期的那条“渐近线”，那么**人类脑海中存在、但 AI 认知里没有的那个微小“差值”（Delta），反而成了最稀缺、最决定胜负的资产**。 面对关于“垂直 AI 公司护城河（Moat）”的讨论，现场甚至幽默地调侃起畅销书《7 Powers》（七大战略力量）与播客《Acquired》的分析框架。Alfred 梳理了 Listen 的复合壁垒： 1. **网络效应与规模经济**：招募池端形成的供需双向动态壁垒。 2. **数据飞轮与高粘性**：访谈数据越丰富，模拟预测越精准；所有历史访谈与用户认知资产均沉淀在平台上，形成极高的迁移成本。 3. **“极简”的产品体验与专业方法论嵌入**：红杉合伙人 Bryan Schreier 曾指出，创始人总喜欢把产品做得很复杂，但客户想要的却是“简单到无脑且开箱即用”（"stupid, simple, and it just works"）的工具。例如撰写访谈提纲（Interview Guide）实际上是一门深奥的学术课题，稍有不慎就会“诱导证人”（lead the witness）。Listen 最初直接采用通用基础大模型，导致客户拿回来的访谈数据质量极其低下、怨声载道；Listen 随后系统性将顶尖市场调研方法论融入专有 Agent 中，确保任何普通企业都能稳定产出高质量结论。 --- ### `38:00` 调研模式的进化：从单向被动反馈到实时协同共创 对话最后探讨了市场调研范式的剧烈变革。传统的用户研究通常是“把方案拿给用户看，请他们打分评价”，但生成式 AI 的“幻觉”与即时生成能力正在让调研转化为**实时产品构想（Live Product Ideation）**。 Alfred 透露，目前很多客户已经开始借助 AI 实时生成不同方案的概念图并投入访谈测试。更进一步，Listen 已经推出了支持 MCP（Model Context Protocol）的扩展接口，用户可以直接将 Listen 挂载给 Claude，下达指令：“让 Listen 在后台循环跑一圈，针对某种营销方向或概念生成一整套创意方案并迅速验证”。主持人 Sonia 更是设想了一种未来场景：当用户在访谈中抱怨“洗衣液太笨重”的刹那，AI 不仅仅是倾听记录，而是当场调用文生图工具开始与用户头脑风暴：“你看，如果做成这种小包装凝珠，是不是你想要的？” Alfred 笑着肯定了这个设想，打趣道：“这个想法太棒了，Sonia 你应该加入我们的产品团队！” 对谈在双方对垂直应用层 AI 演进路径的探讨与感谢声中圆满落幕。 ## Notable quotes > `00:00` 随着我们越来越接近 AGI（通用人工智能），构建产品的工程门槛会变得极低，而真正稀缺且艰难的环节将变成‘弄清楚到底该造什么’（knowing what to build）。这正是 Listen 所专注并试图牢牢掌控的闭环。 — Alfred Wahlforss > `10:00` 人们在面对 AI 时往往比面对人类更诚实。面对一个没有预设立场、不带评判色彩、却对自己表现出浓厚兴趣的非人类实体，受访者往往会完全敞开心扉，甚至觉得访谈具有某种‘心理疗愈’的作用。 — Alfred Wahlforss > `12:15` 每个公司的商业模式本质上都由客户群体的幂律法则驱动。 — Alfred Wahlforss > `20:00` 如果对一个人进行长达一小时的深入交流，AI 就能在相当程度上精准预测对方在各种情境下的偏好与行为。大语言模型非常擅长捕捉这种高维度的个人特征；当系统将尽可能详尽的个体信息输入模型后，在某些特定问题的回答预测上，准确率甚至能达到 95%。 — Alfred Wahlforss > `22:15` AI 的本质是对压缩数据的高维泛化，当在特定空间中积累了足够多高质量的真实访谈，模型便能精准‘补全该类人群思维轨道的全貌’。 — Alfred Wahlforss > `30:00` figure out what to build, build it... Write code and talk to users — Alfred Wahlforss > `35:00` 人类脑海中存在、但 AI 认知里没有的那个微小‘差值’（Delta），反而成了最稀缺、最决定胜负的资产。 — Constantine > `35:00` stupid, simple, and it just works — Bryan Schreier > `38:00` 这个想法太棒了，Sonia 你应该加入我们的产品团队！ — Alfred Wahlforss ## Mentioned Alfred Wahlforss, Listen Labs, Microsoft, Anthropic, Cursor, Sweetgreen, NBC Universal, Chubbies, Manscaped, Skims, Super Bowl, Meta, LinkedIn, BeFake, Qualtrics, P&G, Google, Bain & Company, McKinsey & Company, Nike, ChatGPT, Tide Pods, Mars, M&M's, Acquired, TikTok, 7 Powers, Bryan Schreier, Claude, Model Context Protocol (MCP), Constantine, Sonia ## Video description Alfred Wahlforss, co-founder and CEO of Listen Labs, is building an AI agent that interviews your customers at a scale no focus group ever could—thousands of voice conversations at once, drawn from an audience of 30 million people. A year after launch, Listen serves hundreds of Fortune 100s to Startups including Microsoft, Google, NBC Universal, P&G, Anthropic, Cursor, and Cognition. Alfred explains the counterintuitive finding underneath it all: people are often more honest with an AI than a human interviewer, opening up to a non-judgmental entity that costs less and never makes them feel rushed. He walks through why interview transcripts—not credit card data or behavioral logs—turn out to be the richest fuel for predicting how customers will behave, how Listen back-tests its simulations to know which questions it can and can't answer, and why 80% of the company's engineering goes into building the right audience. As AGI makes building trivial, Alfred argues the scarce resource becomes knowing what to build. That's the loop Listen wants to own. [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI4652106703.mp3)

[0:00] Alfred Wahlforss: 随着我们越来越接近 AGI（通用人工智能），构建产品的工程门槛会变得极低，而真正稀缺且艰难的环节将变成‘弄清楚到底该造什么’（knowing what to build）。这正是 Listen 所专注并试图牢牢掌控的闭环。 [10:00] Alfred Wahlforss: 人们在面对 AI 时往往比面对人类更诚实。面对一个没有预设立场、不带评判色彩、却对自己表现出浓厚兴趣的非人类实体，受访者往往会完全敞开心扉，甚至觉得访谈具有某种‘心理疗愈’的作用。 [12:15] Alfred Wahlforss: 每个公司的商业模式本质上都由客户群体的幂律法则驱动。 [20:00] Alfred Wahlforss: 如果对一个人进行长达一小时的深入交流，AI 就能在相当程度上精准预测对方在各种情境下的偏好与行为。大语言模型非常擅长捕捉这种高维度的个人特征；当系统将尽可能详尽的个体信息输入模型后，在某些特定问题的回答预测上，准确率甚至能达到 95%。 [22:15] Alfred Wahlforss: AI 的本质是对压缩数据的高维泛化，当在特定空间中积累了足够多高质量的真实访谈，模型便能精准‘补全该类人群思维轨道的全貌’。 [30:00] Alfred Wahlforss: figure out what to build, build it... Write code and talk to users [35:00] Constantine: 人类脑海中存在、但 AI 认知里没有的那个微小‘差值’（Delta），反而成了最稀缺、最决定胜负的资产。 [35:00] Bryan Schreier: stupid, simple, and it just works [38:00] Alfred Wahlforss: 这个想法太棒了，Sonia 你应该加入我们的产品团队！

[3:00] Alfred Wahlforss: Even if you go back to the same person and ask them a survey question in a multiple choice fashion, they're much more inconsistent. But we did the same thing with Listen... when you actually have to think and you have to really reason through your answer... you're much more consistent. [30:00] Alfred Wahlforss: Procter & Gamble is kind of the archetype of the best market research organization... they're essentially marketing companies trying to figure out what niches people really care about, then building specific brands to solve those problems.

[0:00] Our goal is to get to a billion people in our audience and then to be able to stratify and know what exactly is this person an expert on? And it might be even something like sneakers. You have some people who are influencers and kind of early adopters. And if you're able to find that audience and interview them first, the insights are much more valuable. And we can learn across all of the interviews that we do. We build profiles of people as we do more interviews in the platform, and then we can search and find the right person. Okay, today we're sitting down with Alfred Walfors, founder and CEO of Listen Labs. Listen is an AI-first customer research platform that can run thousands of voice interviews simultaneously. You launched about a year ago and you now serve 20 of the Fortune 500 including iconic brands like Microsoft Anthropic Sweetgreen NBC and others And Constantine I very excited to sit down with you today and talk about market research and how it getting transformed with AI Yeah thank you for having me Maybe just to get started. So you are building an AI-enabled platform that scales market research. What does that mean? Yeah, so we have this AI agent that can understand your customers better than you can. And the way we do that is by talking to them. So to give you an example, you can ask a question like how can you improve cursors onboarding? And then Listen will create an interview guide, which is an instructions for the agent to make the interviews. And then we have an audience. We have 30 million participants. We can find pretty much anyone from an oncologist to a software engineer. And we'll go and actually talk to them and have hundreds of those interviews and then analyze the data, give you recommendations. And now the final step that we're just launching in a couple months is simulation. So after you've done tens of thousands of interviews in the platform, can you predict how your customers will answer questions in the future? Put it another way, as we get closer to AGI, it will be easier to build things, but the hard part will know what to build. And that's what we're building at Listen. Awesome. Do you have any favorite customer stories? Yeah so Chubbies is one of our customers Like the Shorts brand Yeah they been like one of our early customers what they use you for they use us for everything so a lot of marketing testing for testing shirts to understand um what products perform well and what doesn't and one of my favorite examples is they discovered that chest hair interferes really poorly with one of the materials they have so it's like really uncomfortable to wear one of their shirts and they changed the shirt and it became like radically more comfortable um so we saw you know the small things so the big things manscaped um changed their super bowl ad with insights from from so never heard of that but i'm not gonna ask uh so you you got the men's hair yes that's our niche shaping to clothing that's right wow we do other things skims is one of our Don't know what you're talking about, but I can I know context clues. So that's awesome. I'd love to understand as you framed it, as we get closer to this AGI future. One of the questions I have is you know traditionally I always been very skeptical actually of surveys because people get paid to take surveys So you already got a selection bias issue the things that people say they would do or the way that they describe how they would behave is different from how they actually behave in practice. And so I guess I come from the school of thoughts where actual just telemetry in the real world matters so much more than asking people about what they would do. And so I'm curious what you think of that and how you think AI or ListenLabs can help bridge that gap? Yeah. And so we've done a lot of research on this. One of the things we've done with surveys, for example, is we went back to the same person and asked them a multiple choice survey again, and they were radically inconsistent. So even if you go back to the same person and ask them a survey question in a multiple choice fashion, they're much more inconsistent. But we did the same thing with listen when you actually have to think and you have to really reason through your answer and then you're much more consistent with at least how you answer the same question. And then we're constantly tracking, for example, with Chubbies, when we test their different charts, we, a couple of months later, look back and see how did that perform with the actual sales data. And I think it depends on different use cases. I agree that A-B test is kind of the holy [5:00] But in practice, it becomes really difficult to get right because you need a very large volume of users. And it's really useful to have some kind of input than no input at all. Does Listen do voice to text as in the actual customer who is answering the survey can speak their answer and then you guys transcribe it? Does it also do text to voice? It's a two-way conversation. what does listen start with and what does it finish with for the user experience yeah so it's essentially a zoom call that you have with the agent so you're on video and you can also detect their emotions so that's another way to bridge the gap between what they say and how they actually think and feel so it looks at your eyes the way you say it and and that's kind of much closer to how you actually behave in the real world. And have you seen Personja's point that actually having the person's face and their emotions and their voice and whatnot yields more engagement, truthfulness? Have we been able to have any studies or at least data to point in that direction Yeah specifically with advertising it a huge benefit because you might have people say on a liker scale which is like five questions that you click, are you extremely likely to buy this product? Versus when you might have very high scores on a survey question like that, but when someone also reacts very enthusiastically, it's going to be like perform much higher. And we've seen that those ads then perform better in performance marketing, for example, on Meta and LinkedIn. And can you, if you're the customer and you commission this and you get all this response, can you actually click in and if you ever wanted to watch the interview to get that level of granularity? Yeah, so we built the platform around traceability so that for every data point, you can always click and then look at the video or see the quote. So you know that AI is not just hallucinating kind of where it's coming from. That's awesome. Makes sense. How'd you come up with the idea to build this? So my co-founder and I actually built a consumer app. And- That did what? That went viral. It was called BeFake So you could create an AI avatar of yourself It was an early version of the ChatGPT images and you could fine stable diffusion and put yourself in that world. And that ended up going super viral, and overnight we had 20,000 users. And we were also kind of experimenting with different ways of using AI. So we built this AI interview for ourselves because we had a bunch of questions of how we had a ton of churn, So we wanted to understand why, how they thought about our positioning, different use cases. And it was really useful for us. And that's how we got started. Maybe just walk us through how the industry is changing before and after Listen Labs. Like historically, let's say you're somebody with an app with 20,000 users. You don't understand how users are using the app, what they want next, why they're turning. Historically, how did people go about doing that? Yeah. So what we discovered was that there are these survey tools that are pretty old school, like Qualtrics. But then there's also this very large services industry, because it becomes harder and harder, especially if you want to do market research, where you want to talk to your prospective customers, not your current customers. It becomes harder and harder to do that as you scale. So that a multi dollar industry And what they do is come up with questions to ask which is an academic subject in and of itself It's actually really hard to know, like, how do you ask questions to your point that get to how someone actually will behave? You can't just ask, like, how much are you willing to pay for this? There's different methodologies that work better than others. To finding the audience, how do you source the participants? to then analyze hundreds of these calls. And in the traditional industries like CPG, even in Microsoft, they spend tens of millions of dollars on focus groups to bring people in a room and interview them. And we can help speed that up much faster. Okay, so that's the old world of how this used to be done. Maybe describe the new world. And then it seems to me that there were obvious kind of first-order benefits. It's like, it's probably much more scalable, probably much more cost effective, but there's probably also less obvious benefits. Maybe just talk about, talk about some of those benefits of, you know, what is it like when you actually do AI first market or customer research? Yeah. So most decisions that gets made are not based on the customer input, right? And the reason for that is that, you know, [10:00] That is, it's just a lot of friction to even talk to customers. So when you can lower the barriers of talking to customers, you end up making much smarter decisions. So the speed advantage is actually huge. For us, you can get input within five minutes from real people. And it's a really magical experience when you see hundreds of people populate in your interview. And so that's one thing. And because it's asynchronous, it's also much more affordable. So you can pay people much less than if you would have to run like synchronous interviews. So actually, that's an interesting thing that people often ask us, like, do people even like being interviewed by an AI? And the objective answer is yes, because you can pay them less to talk to an AI than to talk to an actual interviewer. Why is that? I think it's mostly because it's asynchronous and people are very busy. But then also what we found, yeah, lower pressure, you can kind of go on and off. And we also found that people are more honest talking to an AI We had people really open up It a very therapeutic experience because it a non entity that really interested in you And we can also have sensitive conversations like interviewing kids how they react to different products. And so I think that's another advantage as well, that people can be brutally honest talking to the AI. Okay. So historically, for example, if I was doing research on the kids market, very, very hard to access that market. Is that a regulatory thing? Is that a scheduling thing? Yeah, it's you need parental consent. Kids are really busy. They go to school. They have extracurricular activities. How do you find time with them? And you need to find the right kind of kids. Like one of the things we realized is the audience is extremely important. And that's actually where we spend 80% of our engineering resources. Every company is driven by a power law in customer segmentation. So even a product like Sweetgreen, which you would think is for everyone, the right audience is typically urban, high household income, mostly female. And by the way, they need to know what seed oils are, which only like 1% of the population does. And then you find that some people go to Sweetgreen every single day and that 80 of their revenue So if you can find that segment the research is so much more actionable Yeah, there's probably a network effect to it as well, where when you get a certain scale and people use it, you can access the same kind of person that otherwise might be really difficult to access. Or maybe it's a scale economy, something along the lines of accessing those really, really specific people that are really valuable for the type of product that you're trying to introduce? Yeah. It's really about, you know, we have, our goal is to get to a billion people in our audience and then to be able to stratify and know what exactly is this person an expert on? And it might be, you know, even something like sneakers, you have some people who are influencers and kind of early adopters. And if you're able to find that audience and interview them first, the insights are much more valuable. And we can learn across all of the interviews that we do. So we build profiles of people as we do more interviews in the platform. And then we can search and find the right person. So someone might say in a totally unrelated interview, I'm a total sneakerhead. And you can keep that in the database on that person And then when Nike or what have you is launching the new product line you can offer that person up That right That amazing And that was not possible to do before because it was usually like separate entities and it would be a very manual process where you would have an email list and you would just spam email. I've been on the receiving end of those. Yeah, they're terrible. And one of the problems with that is that you then need to have an extensive screening process. So you have something called an incidence rate, which can be 10%, which means only one in 10 people gets qualified to even take the interview. And that causes significant churn on these databases. Yeah. Because it's really annoying to be screened out 10 times to even get paid the first time. Why do these brands even need you for access? If, let's take Sweetgreen. Sweetgreen knows who that 80% is. Can't they just reach them? Don't they have direct relationships with them already? Yeah. So they can, and we do that as well. We connected their CRM and they can send that out. But then the really interesting part is how do you talk to prospective customers? People who also may not be kind of current power law users and how do you compare those [15:00] And then also what we found is that the CRM is typically really unorganized. And sometimes there's also like regulatory issues. If you're at Google, you can't just send emails to people who use Gmail. And it gets much easier to use an external third party. And you run the risk of spam, which can get you totally blocked. I have seen that at some of our companies over the years where, you know, you do outbound and then eventually you're in the Google filter. and next thing you know, you're in Microsoft purgatory. I guess going through you guys, you don't have to deal with that. Yeah, exactly. It's cool. What does this mean for the McKinsey's or whoever's of the world that are, you know, the people that are building the hundred slide decks that, you know, reach 3,000 people to reach some set of insights? Didn't you do that, Sonia? Wasn't that a former wife? You know, no, Constantine, but I'm glad that that's what you think of me. Isn't that what banking does? We hired consultants. I used to hire these people. Got it, got it. So I was a layer on top of the layer on top of the layer. Okay, got it, good. I was even more redundant. But what does it mean for all these people? Do they still have a role to play in this new future Yeah I think AI is changing all roles very quickly And we work a lot with Bain for example So they use us to speed up their traditional processes. And I think they still have a role to play. I think traditional services and being able to then implement these changes is still extremely valuable. But a lot of margins are going to drop. and you have to make sure you kind of unbundle a lot of your services to maybe allow for AI agents to help solve some of the problems that you would go to traditional consulting firms before. Maybe I'm an optimist here, but why wouldn't it be more? Why wouldn't I, if I'm running a business, say, oh, great, I want to find five new areas to expand to now that I have AI and these tools, and I will pay you, Bain, or what have you, the same dollars you use listen and just explore those new areas and tell me where to where to build is that overly optimistic no i think it's one of those areas where the ceiling is very high you can kind of learn more about your customers and you can build more things and so i think i think you right i was thinking the chest hair shirt thing I glad that 20 minutes in you still thinking about chest hair There so many little things that I love to tell the companies that I am a consumer of Like the smallest little thing. Even the way they lace these shoes. I'd love to give that feedback. This is why you're a venture capitalist. Details. Details. We hope to live in a world that finally works the way people want. That would be great. Are you seeing any price and compression already hit the industry? I would imagine if I am Bain's customer, I'm thinking, well, you're able to do this survey a lot more efficiently now with AI than before AI. Who's getting the benefit of that economic surplus? So because you're able to do it faster, I would argue you should be able to charge more for it. And is that what's actually playing out? We have done some studies where we're able to charge hundreds of thousands of dollars to speak to 20 doctors across eight countries. So maybe over the long term, like the individual interview will become more affordable. But I think you be doing kind of two orders of magnitude more of research And I think what really exciting is also simulation which is something we building now where you able to unlock the 99 of use cases where you would never have time to talk to real people. I think that's so awesome in part because there are so many areas where they don't even listen to the customer, like medicine. There are a million little problems with the medical system. I hear about it all the time. And these are, you know, they're doctors, are busy, important people, but it feels like the companies haven't even invested the time in figuring out where all those paper cuts are. And the doctors are really busy. So they're not going to go schedule an appointment and have some long conversation and meet with some group. But if they could do it at any time, like in an app on their phone as part of the normal homepage app and give feedback on their EHR or something in the operating room or something along those lines, that seems like a life-saving use case for listen over time. Yeah. I think what I'm really excited about as well is taking all those small things and then telling another agent to go and solve that problem. And we're getting pulled in this direction by some of our customers where they will have a churn interview and then they will connect them if you [20:00] And a bug, for example, they'll connect that to another coding agent to go and solve the problem. That's cool. Let's talk about generative agent simulation. It seems like the entire industry has gone from market research 1.0, call 100 people one by one, collate them manually, to market research 2.0, AI native, where AI designs the question track, is able to talk to thousands of people simultaneously, synthesize the answers. it seems like we're maybe moving to market research 3.0 with generative agent simulation what do you make of that and you know i both see the dream of it um i see how synthetic data has changed for example self-driving cars and then i also am inherently skeptical of it like is a bunch of synthetic data just remixing what's already in the pre-training sets and are you actually learning anything useful or with alpha there and so i'd love to i'd love to hear your take on it and how you guys are taking on the 3.0. Yeah, and maybe what is it to start? Yeah, so the way we are building simulation is by interviewing a single person. So if I interview Konstantin for one hour, I can probably start to predict your preferences to some degree Fascinating insights about chest hair And it turns out that LMs are quite good at this as well So you can essentially try to feed in as much information as possible on a single individual And then in some cases, we're able to get 95% accuracy to predict how they will answer certain questions. What sorts of things are you finding that you can predict well versus can't? One of the most useful things is message testing. So that's the idea of like, what's the tagline on the billboard? Or I was actually using it this weekend. So I have created a panel of our customer base. And I had to come up with the title for a talk at a conference. And it's like a small thing, but it actually does matter because it will increase conversion if people show up. And I came up with 100 different titles for my talk and inputted that into our simulation. And then the top talk was like twice better than the next one. Wow, cool. And I don't know if it's correct, but it certainly felt correct. And it was really helpful to have guidance in making that decision. And I also think like even if it wrong it just nice to have some help in making a decision It also nice to outsource your decisions And how does it compare to just asking ChatGPT the same thing Yeah so then I inputted the same questions into ChatGPT And I actually had one. I had another talk I did that was not so successful. And I inputted a competitor or another talk that was more successful. And I showed both of them to ChatGPT and both of them to our simulation. and in ChattoBT, it picked the wrong one and in our simulation, it picked the right one. So, you know, it's early for us. We're going to release this in a couple of months but it seems like it's performing better than the general models and the models are trained on the average person and you want to build for a very specific niche and that's how we can kind of essentially train the models to follow that niche. And just to push on this because I think it's so fascinating. Like, can't you kind of force the models into a specific niche or personality? Like, hey, ChatGPT, you're a 35 year old, really grumpy software engineer that likes using your terminal. Like, and then it does then take on the preferences of that niche. It like it sort of my mental model at least And so I actually surprised that you know ChatGPT wasn able to arrive at the right answer and then bootstrapping off real user data was because ultimately it all is kind of a reflection of real user data, right? And so actually, what is the intuition for why kind of SIM only on pre-trained data isn't sufficient? Yeah, so we've tried many different inputs, and that certainly performs a little bit better than just vanilla ChatTBT. But what performs much better is we try credit card spend, kind of behavioral data, purchasing behavior. But what we found was the best data set is interviews because it's more kind of allows you to go off tangents. It understands you can ask behavioral questions. So also it can't just be any interview. like the way you design the questions is also really important. And the intuition I think is that the models don't have clean data on how a specific persona acts and how they think. It's anecdotal, but it makes perfect sense because if you want to understand someone, what better way to understand them than asking them a lot of questions? That's why we're all here. I guess kind of the purpose of this type of format. [25:00] And if you have enough people that follow a certain group as opposed to the average, that can tell you a lot about other things that they might not have explicitly said. All of AI is this generalization of some sort of compressed data of some sort. And so if you have this compression in a slightly different part of this hyperspace that you say, now complete this orbital of what everybody is thinking in this category of person, Listen can fill that out because it has enough interviews. Yeah. Do you think that you will offer that package as a product? As in, if I wanted to understand my customer and for me, for us, our customers, founders, they're very different, though. So extremely different people. So if I wanted to understand my customer, could you do active interviews, the normal ListenLab interviews, have a thousand or ten thousand cumulatively, and then offer a little special purpose ListenLab's bot that then I can use instantaneously for any ad hoc question? Yeah, that's exactly what we have. Okay. So that's what we call augmented responses. The cool part of this as well is that it can also live in your coding agents or your other agents So I think in the future you will want to have almost a human API where the agents are able to call the preferences of your users to be able to know what to build how to do it or who to invest in or how to help them the best Today, is it all rag? Is it fine-tuned? Is it something else? How do you take those conversations and then combine them with, you know, the models that you're doing the rest of the list and labs with? Yeah, we are doing post-training, typical RAG as well. There's like a bunch of different techniques. Some of them are proprietary, but yeah. All right. We'll do customer interviews and all your engineers report back. I'm curious what you think of multi-agent systems and their role in helping us kind of iteratively use, you know, at inference time, iterate to a better answer. Is that part of how you're doing simulation or not? Yeah, like the way we do simulation is essentially you have one person that you model really well and then you scale it up with a thousand people. So you have a representative sample and it essentially multi But you not having those thousand people debate each other That what I asking Oh yeah No we don have that yet but that a good question Do you think that would help Potentially but there are these other competitors that are doing that approach more. I think the worry is that, again, chaos theory tells us when things kind of compound, it becomes really hard to predict how the things are going to interact with each other. and it's something we definitely should explore more, but I'm a little bit skeptical of the approach. Maybe the analogy I'd make is like the AI council approach of send the same query out to three different LLMs and then have one LLM act as judge in synthesizing them. I do think on average you get a slightly better response. Yeah. Cool. So where else do you see yourself going from here then? You're going from market research 2.0 to market research 3.0 now with generative simulation. Do you expect that 3.0 takes over as the majority of queries over time? And then what else is ahead? Yeah, I think you'll still need human input, but I think there would be many more use cases that are now opened up where you can get customer input. So for the large decisions if you doing a Super Bowl ad or things like that you will still need to run real interviews But for the smaller things like what should be the tagline for your billboard if it a small billboard then you can use simulation to answer that And I still think that there's a lot of alpha on the core product as well to improve. I mean, when we started, the core idea was just making the interview less annoying to go through. Like we had an eval that looked at repetitive questions or looked at, is the AI even able to follow the instructions? And with GPT-4, sometimes we would ask the same question a hundred times. And in the beginning, that eval was like 20%. Now we've been able to climb that eval to be 85%. But now we created a new eval that's much more advanced. So it's able to understand what are you doing on your screen when you're screen recording, or can you skip questions that are not relevant anymore? And now we're back at like 20%, which I think is one of the values that vertical AI companies can have is that they have a lot of [30:00] this proprietary eval that they can use and essentially climb that eval. And that's your advantage as a vertical AI company. Keep pushing forward, better data, harder problems, better data, repeat. It seems to me like you're in the middle of a very interesting infinity loop, right? Because fundamentally a company is figure out what to build, build it, figure out what to build, build it. Write code and talk to users. Exactly. And the build it is coming up rapidly up in exponential. And the figure out what to build is the thing that you are pushing forward. And then it's not only even outside of product and engineering, the broader loop is actually strategy execution, strategy execution. And so much of what AI is enabling us to do is it's making execution faster, cheaper, better, all these things. And the thing that you guys fundamentally are positioning yourselves to do as a company is the strategy part from what to build to what to say. Is that a fair synthesis? Yeah. And I think when we have that one person billion dollar company, we'll be part of that loop. So I have a coding agent and listen and then run that in a loop. And we have these autonomous organizations And even the big companies though like back to this idea of you can implement things faster Let say you have an agent I mean, if you can be a big company and we're talking in software because software is native to us. But in software, you can talk to a customer, figure out a bug, create a PR, have a coding agent, close it, ship it. customers happy, that seems like a really important left-hand side of the equation, find the bug from an actual human. But I imagine it's the same thing in a big Adams company. Like if you're consumer packaged goods, if you're clothing, if you're any of those things, I imagine that's even more important because you have to figure it out because once you actually do the thing, it's done. Yeah, exactly. Like Procter & Gamble, when they're launching in a new market, that can be tens of millions of dollars, if not more. and you have to make sure that that is right when you launch. And that's one of the reasons why the customers have listened. Who has done this historically really well? Who are the companies that are admired in history, have done a great job of listening to their customers, either in the consumer space or in the software space I think Procter Gamble is kind of the archetype of best market research organization where they're essentially marketing companies that are trying to figure out what are niches that people really care about and then build specific brands to solve those problems. I mean, one example is the Tide washing machine. Yeah, pods. Pods, that they were able to figure out that it was really uncomfortable to use the washing liquid. And they discovered that people wanted something that was much more easy to use. And through customer interviews, they found this insight, made this new Tide pod and became really successful. another example which is in the acquired podcast when they talk about mars and they did the first one of the first market research studies in the 1950s where m&ms were originally designed for being in the used in the army because they were like a sweet treat they don't melt in pocket and they discovered through market research that another great segment was young kids and And they then decided to pivot the entire advertising strategy to focus on this because it doesn melt and ruins your furniture for example and things like that As we progress towards this, you know, Listen Labs, future vision of the world, what are things that you're confident will work? And what are the things that you're still not sure about? I'm confident that in the future, you'll still need to have human input. Because even if you have a perfect rational being, like AGI, humans are still irrational. And they will still kind of be chaotic in their nature where they all of a sudden get obsessed with a new product, you know, a new TikTok trend that shows up and you have to change your entire marketing strategy towards that. And so I think that will remain a really huge part of how we do things. I think I'm still uncertain of what level simulation will play. And I'm confident that it will work for certain questions, but we'll see how good the models get to predicting human behavior. I mean, I'd imagine that it's actually even more important the better AI gets to have the delta. [35:00] Because the competition, if companies are about serving people, which I think we can all agree on, like at the end of the day, every company is about serving humans. Constantine, our resident humanist. I'm a humanist, absolutely. But if companies are about serving people, because that's why we're all working, is to help someone else in some way, and intelligence gets better and better and better. and you kind of have what the human wants here and the intelligence is approaching that asymptote, then the delta in that asymptote, which is what is in a human's mind that isn't in the AI's mind, only becomes more important. Yeah. And one of the things that we've also realized is that there's a lot of talk around what's the mode of these AI, vertical AI companies. And... Yeah, what's your mode? What is our mode? We've got network effects and scale economies. Yes. Those are nice. I feel like we're on an episode of Acquired right now. Hey, I'm feeling it. I'm feeling it right now. It's a good book. Recommended. Seven Powers. Yeah, Seven Powers. Love Seven Powers. Yeah, on the modes, I mean, we have the clear modes, which are the network effects on the panel where you have supply and demand dynamic We also have the network effects the data mode that as we do more interviews you gets better simulation and then the product is very sticky because you have all these interviews in your platform and you don't want to lose that you want to track things over time but even like the simplest things i think in terms of product advantages like one of the first things that brian shire said one of our sequai partner was that founders want to build something that's complex, but customers want something that's stupid, simple, and it just works. They don't want to configure their own workflow. They don't want to sit and build a custom software. And just one example of this is creating the interview guide is really difficult. It's actually an academic subject. And it's one of the reasons why you have services firms, because they know what methodology to use if you want to understand pricing or brand perception, these kind of things. You don't want to lead the witness. You don't want to lead the witness. And it's really hard to get that right. In the beginning, we just used the vanilla LM models and the customers would create the interviews they would get the data back and then they come back to us really frustrated saying like what is this Like I can use this data for anything And we took the blame for that Now, we've trained it to follow the best practices so that you always get good data out of the interviews. And I think that's the advantage you have as a vertical AI company, that you can essentially train this agent to follow best practices and the work that you do. So I want to go back to the concept of tide ponds that you had mentioned earlier. I think it's really interesting. And so much of market research, as I understand it today, is almost more inviting people to pass judgment on ideas that you feed them. But it seems to me that one of the, you know, hallucinations can be a bug. They can also be a feature with generative AI. And, you know, do you think we're going to see user research actually evolve into live product ideation? I could almost imagine, you know, AI inventing solutions as customers are going about their interview process, even helping visualize those solutions. Are your customers doing that already? Or do you think we going to have a moment where AI can create a Tide Pods idea in a market interview anytime soon Yeah I think that really exciting Today they do that manually use AI to generate images of different concepts and feed that into the interviews But I think specifically also with simulation it becomes really powerful So we now have an MCP as well so that you can feed that into Claude and then you can tell Claude like, hey, run, listen in a loop and then come up with a bunch of ideas for how to market something or different concepts. And then you can have it run like that. I'm even thinking in the course of an interview as somebody's complaining about, you know, the tide is, I can't sound very portable for the AI to be, you know, live brainstorming with you solutions, not just listening. Yeah, this is what it could look like, an image generator too. That'd be cool, Sonia. Yeah, I think it's a good idea. You should be on our product team. Awesome. Well, Alfred, we really love what you're building. Thank you for taking the time to share insights both on the brother market, which I think is just so fascinating, and also what it takes to be building in the application layer right now. We really admire the business that you've built, and thank you for your continued partnership. Thank you so much. [40:00] Thank you.