CTGT, Eyepop Founders Build Model-Agnostic Moats to Beat AI Slop

CTGT与Eyepop创始人构建模型无关护城河，抵御AI劣质垃圾 早期AI初创公司必须用模型无关的运行时与深度的技术研发取代肤浅的套壳及未经检验的“氛围式编程”，才能在基础模型的飞速迭代中生存并赢得企业合同。 经验丰富的初创公司操盘手分享了在不断演进的企业环境中打造具备防御力的高增长AI公司的“从零到一”实战指南。讨论深入剖析了团队为何必须采用模型无关的运行时和严谨的工程标准，而非停留于浅层的“氛围式编程”，并利用机制性模型可解释性等深层技术护城河来解决实际的合规与对齐挑战。创始人们还强调了跨学科招聘、线下协作、快速客户实测竞比以及建立专属分发渠道的必要性，以避免在基础模型的迅猛演进中沦为平庸的同质化产品。 与 Web3 等早期技术周期不同，企业级 AI 落地能够带来立竿见影且可衡量的 ROI，但金融等高风险机构在全面部署前仍高度依赖严格的模型安全性与可解释性保障。 (2:27) 初创团队虽然能借助 AI 工具实现巨大的运营杠杆，但扎实的计算机科学基础依然必不可少，以防范未经审查的“vibe coding”引入安全漏洞和严重 Bug。 (4:17) 针对非确定性 LLM 引入的不可预测边缘情况，初创公司应维持工程严谨性，通过每日自动化回归测试集与严格的人工 PR 代码审查来保障系统稳定性。 (7:19) 初创企业应构建模型无关（model-agnostic）的应用运行时与架构，以便在更先进的基础模型出现时实现无缝切换，避免被单一供应商生态锁定。 (17:25) 在 AI 原型开发成本趋近于零的当下，建立持久护城河需要超越薄层 API wrapper 的非对称优势，例如稀缺的学术 IP、定制化的垂直工作流或专属分发受众。 (12:09)

Navigating the zero-to-one phase of an AI startup requires balancing rapid prototyping tools with rigorous engineering standards and sustainable moats. Founders Cyril Gorlla of CTGT and Andy Ballester of Eyepop.ai argue that enduring defensibility relies on deep technical IP, model-agnostic architectures, and in-person team execution rather than superficial wrapper applications. Defensibility Against Commoditization and AI Slop (12:09) As low-cost development tools make software prototypes virtually frictionless to build, Ballester and Gorlla warn that early-stage markets are flooded with thin API wrappers and low-value 'AI slop.' Ballester stresses that investors and enterprise buyers look past Total Addressable Market projections to scrutinize team-level unfair advantages, specifically execution velocity, proprietary data, and curated audiences that competitors cannot easily replicate. Gorlla argues that durable moats stem from specialized research that eliminates high-cost enterprise overhead. For example, when a Fortune 30 financial institution sought to use the open-weights DeepSeek model for historical research, CTGT used mechanistic interpretability to identify and excise internal censoring subnetworks—enabling queries on sensitive historical topics without resorting to brittle prompt engineering or expensive fine-tuning. Gorlla contends that solving such fundamental alignment and compliance hurdles replaces the need for enterprises to hire dozens of machine learning engineers. Engineering Rigor Versus Unvetted Vibe Coding (4:17) While developer assistants like Cursor, Claude Code, and GitHub Copilot allow lean teams to punch above their weight, Gorlla warns against uncritical 'vibe coding.' He points to incidents like the viral 'Tea' app, where rapid automated development left a production Firebase database completely exposed. Because foundational models still display benchmark hallucination rates between 30% and 40%, CTGT relies on in-person whiteboarding interviews to verify that incoming engineers understand foundational computer science abstractions before letting them leverage AI assistants. To manage the non-deterministic nature of large language models, Eyepop.ai runs automated daily regression suites, testing known prompt matrices against expected outputs while enforcing mandatory human code reviews on pull requests. Concurrently, CTGT experiments with an isolated monorepo where agentic tools author 85% of the codebase under 15% human oversight. Both founders emphasize routing developer tooling through private enterprise deployments, such as Google Cloud Vertex AI, to prevent customer data leakage. Model-Agnostic Runtimes and Practical Deployments (17:25) Neither founder advocates locking product architecture into a single foundational model vendor. Ballester explains that with the rapid churn of Vision-Language Models, Eyepop.ai built an AI application runtime that acts as an abstraction wrapper, enabling developers to swap underlying vision models without re-architecting their software. Similarly, Gorlla notes that CTGT remains strictly model-agnostic to protect enterprise clients from vendor lock-in across OpenAI or Anthropic ecosystems. To establish trust without large marketing budgets, both founders recommend bypassing defensive sales postures in favor of live customer bake-offs against incumbent enterprise tools. Ballester notes that enterprise buyers frequently do not realize current AI capabilities; demonstrating tailored vertical workflows that spin up in 10 to 60 minutes without requiring an in-house machine learning engineer proves far more persuasive than high-level pitch decks. Team Dynamics, 90/10 Solutions, and Co-Location (19:32) Ballester cautions founders against hoarding unreleased prototypes for months under the guise of refinement, urging early validation of customer willingness to pay. Gorlla encourages founders to seek '90/10 solutions'—scrappy approaches delivering 90% of the outcome for 10% of the effort—which incumbents routinely overlook as crude. This pragmatic philosophy also shapes hiring: CTGT recruits cross-disciplinary talent, including a NASA astrophysics researcher, valuing first-principles reasoning and adaptability over traditional computer science pedigrees. Both founders maintain that physical co-location is vital during the zero-to-one stage. CTGT operates fully in-person using an open 'Athenian forum' structure that allows junior hires and Berkeley interns to push production code directly to Fortune 500 infrastructure. Eyepop.ai uses a three-day in-office hybrid schedule paired with bi-weekly retrospectives, allowing immediate desk-side debugging while preventing burnout as individual employees balance multiple functional roles. The discussion underscores that while generative AI lowers the barrier to software generation, long-term commercial survival hinges on solving acute, verifiable pain points. Unresolved questions remain around how early-stage teams will sustain rigorous human code review as autonomous coding agents author increasingly large proportions of production codebases.

CTGT联合创始人Cyril Gorlla与Eyepop.ai联合创始人Andy Ballester指出，在生成式AI创业的“从零到一”阶段，初创团队必须摆脱浅层的API包装与脆弱的“氛围式编程”，转而依托深层技术研究、严格的模型评测与模型无关架构，以快速验证企业级真实痛点并建立持久的竞争壁垒。 核心壁垒：深层技术与抵御浅层包装 (12:09) Andy Ballester强调，在投资人和客户面前，仅仅证明庞大的潜在市场规模（TAM）并不足以构成壁垒，真正决定胜负的是团队的不对称优势，包括专有分发渠道、受众聚合与极高的执行速度。Cyril Gorlla进一步指出，随着行业充斥着过度包装和伪造数据的低价值“AI废料”（AI slop），薄弱的应用外壳极易在底层大模型升级时被瞬间颠覆。CTGT的竞争护城河建立在针对模型内部机理的学术研究上，这种稀缺的技术能力让企业无需雇用数十名机器学习工程师即可解决合规与对齐难题。 以CTGT在白宫展示的一家财富30强金融机构案例为例，该机构需要使用中国开源权重模型DeepSeek进行经济与历史分析，但面临天安门事件等敏感话题的模型审查限制。CTGT并未采用脆弱的提示词工程或成本高昂的全量模型微调，而是利用模型可解释性技术分析激活回路，精准外科手术式地移除审查子网络。Gorlla认为，这种机理层面的介入方法比粗暴微调更为稳健，是构建技术护城河的典型路径。 研发纪律：抵制盲目编程与严控系统安全 (4:17) Ballester观察到，在过去两年紧缩的风投环境下，初创企业倾向于利用AI工具赋能多面手以提升人效。然而Gorlla警示不可盲目依赖未经审查的“氛围式编程”（vibe coding），并举出“Tea”应用因快速生成代码而将生产数据库毫无防备地暴露在Firebase链接上、酿成严重用户数据泄露的教训。Gorlla强调，高效工程师必须掌握底层计算机科学原理以识别LLM的缺陷，毕竟基准报告中大模型仍存在30%至40%的幻觉率。为此，CTGT将面试流程全面转向现场白板考核，并通过学术圈引荐与直聘科研人员（如NASA天体物理背景学者）补充跨学科力量。 在技术栈与数据防护方面，CTGT允许工程师选用Cursor、Cline或Roo Code等IDE，但所有代码与数据均通过Google Cloud Vertex AI的私有模型端点流转，严格杜绝专有数据外流或被用于第三方模型训练。Eyepop.ai则以VS Code搭配GitHub Copilot为基准，禁止直接向主分支提交代码，所有变更必须经过拉取请求（PR）与严格的人工审查。 架构解耦：模型无关运行时与持续评测机制 (7:19) 针对非确定性LLM引入的不可预测边缘情况，初创团队不能仅沿用传统软件工程的负载测试。Ballester介绍，Eyepop.ai建立了每日自动化回归评测套件，通过已知提示词矩阵持续测试模型输出与行为退化，在生产部署前设置体验安全阀门。此外，面对视觉模型与多模态大模型（VLM）快速迭代的不确定性，Eyepop.ai开发了独立的AI应用运行时，将其作为模型能力的抽象封装层，确保客户可在更先进的底层架构出现时无缝切换底层模型。 CTGT同样采取了严格的模型无关策略，以避免客户被OpenAI或Anthropic等单一生态系统锁定。Gorlla透露，CTGT设立了隔离的Monorepo进行智能体编程实验，其中85%的代码由Claude Code等工具自主生成，人工干预上限为15%。这一实践促使工程团队将LLM视为主动的协同工程伙伴，而非简单的自动补全工具。 敏捷验证：规避伪需求陷阱与坚持现场协作 (19:32) 借助bolt.dev等工具，软件原型的边际开发成本已降至接近为零，但两位创始人警告这反而容易让初创团队陷入“拿着解决方案寻找问题”的误区。在Y Combinator等顶尖孵化器中，半年内便会出现解决真实痛点与展示技术噱头团队的两极分化。创始人往往借由“持续打磨”借口滞留早期版本，延误了验证用户付费意愿的最佳时机。Gorlla提倡推行能以10%的研发精力达成90%成效的“90/10解决方案”，并在金融等垂直场景直接与企业现存工具开展现场实测比拼。 在组织运作上，两位创始人均强调早期阶段面对面协作的不可替代性。Eyepop.ai采用每周3天现场办公的混合模式，并引入敏捷双周回顾机制以平衡超负荷工作、防止核心人才倦怠；Ballester指出，早期产品验证与排错无法容忍异步沟通的延迟，工位间的即时交流能迅速纠偏。CTGT则坚持100%全员现场办公，构建仿效“雅典论坛”的扁平化学术文化，让从伯克利实习生到资深专家的全员均能直接挑战假设，将顶尖科研成果转化为企业级生产部署的周期压缩至数天。 总结而言，两位创始人认为AI初创企业的长期生命力取决于能否解决唯有AI才能原生地解决的深层痛点，而非依附资本热度的表面化功能叠加；然而，如何在底层模型飞速通用化、基础API不断被商品化的产业格局下持续守护专有技术与垂直运行时的溢价空间，仍是留给所有AI构建者的长期追问。

## Background 该视频由频道 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台上。内容记录了一场关于 2025 年 AI 初创企业从 0 到 1 实战路径的深度圆桌对话，由 Glean 的 Rebecca Hinds 主持，邀请了 CTGT 创始人兼 CEO Cyril Gorlla 与 Eyepop.ai 联合创始人兼 CPO Andy Ballester，旨在为 AI 原生企业提供涵盖模型策略、技术架构、防御性壁垒及团队搭建的实战框架。 ## TL;DR 对谈围绕 AI 原生企业在起步期的核心决策展开，深入剖析了如何跨越盲目依赖 AI 编码（Vibe Coding）的安全陷阱，以及在 LLM 具备非确定性特征下建立每日回归测试体系的必要性。两位创始人分享了构建模型不可知（Model Agnostic）应用运行时以规避供应商锁定的架构哲学，并介绍了利用模型可解释性切入银行金融等高风险场景的深层壁垒打法。此外，嘉宾们还探讨了如何通过“90/10 方案”实现极速原型验证、通过现场白板考核筛选非典型技术人才，以及如何警惕“AI 垃圾（AI slop）”与浅层套壳伪需求。 ## Atmosphere & style 整场对话务实干练、干货密集，彻底褪去了风口期常见的浮躁炒作与空洞概念。年仅 23 岁的顶尖学术派创业新秀与前 GoFundMe 资深连续创业者在主持人的引导下形成互补，既有对现实安全暗礁与大模型幻觉的犀利警示，又有关于雅典论坛式辩论文化和极客精神的真诚流露，给观众带来一种置身于硬核工程研发与真实商业战场前线的沉浸感。 ## Key takeaways - [02:27](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=147s) 传统大型企业对 AI 的采购热情前所未有，在高风险业务落地中，能够提供透明度、可解释性与信任度的底层工具已成为核心刚需。 - [04:17](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=257s) 警惕盲目依赖 AI 编写代码（Vibe Coding）背后的安全暗礁；AI 能带来数十倍生产力杠杆的前提是工程师具备扎实的软件底层架构把控能力。 - [06:31](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=391s) 面对 LLM 固有的非确定性与幻觉率，初创公司需通过回归现场白板编码面试选拔人才，并建立每日自动化 Prompt 探针等基准测试体系作为安全防线。 - [09:30](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=570s) 在内部工具链管理上，建议给予开发者选用 Cursor、Cline 等 IDE 的自由，但必须通过私有云网关隔离专有数据，并坚持严格的代码 PR 人工审查。 - [12:09](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=729s) AI 大幅拉低了通用开发门槛，但在商业底层逻辑中，投资人与市场考核的核心依然是团队壁垒（Defensibility）以及“为什么是你们团队”。 - [13:34](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=814s) 高风险行业（如金融、银行）部署 AI 的痛点在于严苛合规与高昂调试成本，初创企业通过在真实业务环境中与既有方案进行正面盲测能最快建立客户信任。 - [18:17](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1097s) 架构设计应秉持“模型不可知（Model Agnostic）”哲学，打造可随时热插拔最新模型的应用运行时封装，避免被单一模型供应商深度锁定。 - [19:32](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1172s) Bolt.new 等工具让原型构建成本趋近于零，但早期团队切忌“拿着锤子找钉子”，必须在第一时间将产品交付给真实受众以验证付费意愿。 - [22:15](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1335s) AI 双向赋能了非技术人员与技术人员，招聘逻辑正在从名校计算机专业转向招募具备敏锐产品直觉、深厚逻辑底蕴的非典型跨界人才。 - [27:10](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1630s) 加速器（如 YC）的核心价值在于压缩指数级增长曲线前期的平缓摸索阶段，通过顶尖导师网络与声誉背书大幅提升融资与起跑速度。 - [29:04](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1744s) Agentic AI 能够以极低的边际成本充当叙事教练与情感解析伙伴，帮助用户克服向他人开口求助的心理恐惧，并从客服工单中提炼深层情绪洞察。 - [32:15](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1935s) 在模型对齐与可解释性领域，深入神经网络定位并编辑具体激活神经元，开辟了一条比传统全参数微调和脆弱 Prompt 工程更具扩展性的干预路径。 - [36:16](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2176s) 早期初创团队需要通过高频面对面协作、双周复盘机制来平衡敏捷节奏与成员心理状态，并建立崇尚真理的雅典论坛式扁平文化以加速科研落地。 - [38:21](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2301s) 0 到 1 阶段应尽早沉淀垂直受众壁垒，并善于寻找“90/10 方案”——用 10% 的敏捷成本实现成熟巨头 90% 的功能质量以实现弯道超车。 - [39:30](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2370s) 警惕仅套在基础模型之上的浅层包装（AI Slop），初创团队必须严谨求证业务场景是否存在非 AI 不可的真实痛点，构建具备长期防御性的业务架构。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=4s) 嘉宾介绍与创始背景：从斯坦福实验室到 GoFundMe 之后的下一站 圆桌讨论在座无虚席的热烈氛围中开场。主持人 Rebecca Hinds 率先介绍了自己的背景，她目前在 AI 工作平台 Glean 负责思想领导力（thought leadership）团队，此前曾在 Asana 领导其内部智库“工作创新中心”（Work Innovation Center），长期专注于 AI 与组织协作的研究。随后，她将话筒交给了台上的两位创业老兵与新秀。 CTGT 联合创始人兼 CEO Cyril Gorlla 讲述了自己从学术界转向工业界的故事。几个月前，年仅 23 岁的 Cyril 毅然离开斯坦福大学的研究工作创办了 CTGT。该公司的核心使命脱胎于他在斯坦福多年的底层机理研究：探索 AI 模型在根本层面上究竟是如何“思考”的。Cyril 指出，随着像 Glean 这样的平台加速推广 AI，以及银行业和金融业等高风险业务场景开始大规模部署 AI，理解模型行为并确保其按照人类意图可靠运行变得前所未有地紧迫。今年 2 月，CTGT 完成了由 Google 和 General Catalyst 领投的首轮融资，目前公司正处于快速扩张期。 Eyepop.ai 联合创始人兼首席产品官（CPO）Andy Ballester 则是一位资深连续创业者，他曾是知名众筹平台 GoFundMe 的联合创始人兼 CTO。2021 年走出疫情阴霾时，Andy 在圣地亚哥的一家风险工作室（venture studio）担任导师，当时他接触到一个孵化班，里面有四家截然不同的初创公司都在尝试利用计算机视觉技术。虽然各家的技术实现路径大相径庭，但他们试图解决的本质问题完全一致：“我能否在图像和视频中定位、测量或统计特定物体与状态？” 这种碎片化痛点促使 Andy 创立了 Eyepop.ai——一个旨在让任何产品都能“极简接入”AI 视觉能力的自助式计算机视觉平台。 - **Cyril Gorlla（CTGT 联合创始人兼 CEO）**：23 岁离开斯坦福大学创立 CTGT，致力于解决大模型黑盒机制与可信赖 AI 问题；2025 年 2 月获得 Google、General Catalyst 等机构领投的首轮融资。 - **Andy Ballester（Eyepop.ai 联合创始人兼 CPO）**：前 GoFundMe 联合创始人兼 CTO，针对企业在图像/视频中“寻找、测量、计数”的通用需求，打造低门槛的自助式计算机视觉平台。 - **Rebecca Hinds（Glean 思想领导力负责人）**：曾任 Asana 工作创新中心主管，主导本次关于 AI 创业从 0 到 1 实战策略的对话。 --- ### [02:27](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=147s) AI 创业新纪元：企业采纳飞轮与产品体验的质变 谈及为何当下是创立 AI 公司的绝佳时机，Cyril 与 Andy 从企业落地与产品体验两个维度给出了剖析。Cyril 强调，企业对 AI 的拥抱速度与此前的区块链或 Web 2.0 浪潮截然不同。哪怕是那些内部流程极其繁复、引入外部供应商门槛极高的大型传统机构，如今展现出的热情与投入也是前所未有的。尽管落地过程中依然存在重重官僚障碍——例如初创公司通常需要先打通客户内部的“卓越中心”（Center of Excellence），再在各个业务部门内部播撒用例——但由于 AI 带来的投资回报率（ROI）显而易见，能够提供透明度、可解释性与信任度的底层工具已成为企业采购中最紧迫的刚需。 Andy 则从自身多年专注于“产品与技术交汇点”的视角分享了他的兴奋之情。过去许多高度复杂的交互逻辑受限于硬编码成本而难以实现，如今大语言模型（LLM）能够精准理解用户意图并在系统中自主流转，从而为用户带来真正“魔法般”的智能化体验。更重要的是，创始团队的生产力边界被彻底重塑：借助先进工具，极其精简的微型团队现在能够爆发出以往庞大工程团队才能具备的交付能力。 ```mermaid flowchart LR accTitle: Enterprise AI Adoption and Tooling Need subgraph Enterprise["大型企业导入路径"] A["卓越中心 (CoE) 审核"] --> B["业务部门 PoC 用例"] end subgraph Bottleneck["企业核心痛点"] C["高风险场景 (金融/银行)"] --> D["透明度与模型信任缺口"] end subgraph Solution["初创企业机会"] E["CTGT 等可解释性平台"] end Enterprise --> Bottleneck Bottleneck --> Solution ``` --- ### [04:17](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=257s) 精益团队与“通才”兴起：“Vibe Coding”背后的安全暗礁 随着过去两年风险投资环境的收紧，投资人对初创公司的资本与人力效率提出了严苛要求，期望小团队能够“以少胜多”，以极快速度同时推进产品开发与市场营销。这种环境催生了大量依赖 AI 辅助的“通才”（Generalist）工作方式。然而，Cyril 对这种盲目依赖 AI 编写代码（即所谓的“Vibe Coding”）的趋势提出了尖锐预警。 Cyril 分享了前一晚与 Andy 晚餐时讨论到的真实案例：近期爆火的一款匿名社交应用 Tea App 几乎完全是通过“Vibe Coding”快速拼凑出来的，上线后迅速积累了大量用户。然而由于开发者缺乏对底层安全架构的理解，其后端数据库竟然直接挂在未加权限控制的 Firebase 存储链接上，导致任何公开访问者都能轻易抓取所有用户的敏感个人信息和上传记录，酿成了灾难性的数据泄露丑闻。 Cyril 认为，AI 辅助虽然能让工程师的生产力提升 10 倍、50 倍甚至 100 倍，但前提是工程师必须深刻理解软件工程的底层抽象原理。只有当开发者能够敏锐识别出 AI 在何处出错时，LLM 才能成为强大的杠杆；若是将架构和安全等关键决策全盘托管给 AI，后果将极其致命。 --- ### [06:31](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=391s) 重返现场白板面试与非确定性系统的工程严谨度 面对 AI 时代潜在的“水货”与安全陷阱，初创公司在人才选拔与系统测试上正在经历深刻的范式转变。Cyril 透露，CTGT 的招聘流程已经全面回归“全流程现场面试”，甚至专门将外地候选人飞到公司进行面对面的白板编码考核。这一硬性门槛旨在确保候选人具备真实的工程实力与底层掌控力。对于服务 Fortune 500 强企业的团队而言，数据安全与隐私合规不容有失，而当前最顶尖的语言模型仍存在极高的错误率——正如 OpenAI 发布的 GPT-4o 技术报告所披露的那样，模型的幻觉率依然维持在 30% 到 40% 左右，完全无法在无人把关的情况下直接接管核心业务。 ```mermaid flowchart TD accTitle: Dual Testing Pipeline for AI-Native Systems subgraph Traditional["传统工程质量保障"] T1["架构评审 (Architecture Review)"] T2["高并发负载测试 (Load Testing)"] end subgraph AINative["AI 原生质量守护"] A1["每日自动化 Prompt 探针"] A2["已知输入与预期输出回归测试"] A3["非确定性与幻觉指标监控"] end Traditional --> System["高可用、安全的企业级交付"] AINative --> System ``` Andy 进一步补充了从传统软件工程跨越到 AI 原生开发时的测试体系重塑。在初创公司“从 0 到 1”的极速推进中，团队必须时刻平衡交付速度与工程严谨性。由于 LLM 存在固有的“非确定性”（Non-deterministic）陷阱，团队除了需要保持架构评审、负载测试等传统工程规范外，还必须引入常态化的实验与质量监控流程。在 Eyepop.ai，所有接入系统的新模型每天都必须经历严格的自动化工具检测，通过注入已知 Prompt 并校验输出结果来设置工程“安全阀”，防止模型漂移或产品体验崩塌。 - **现场白板面试回归**：CTGT 坚持全员现场代码实操考核，防止被“AI 代写”伪造技能的候选人混入团队。 - **不可忽视的模型幻觉**：OpenAI 官方技术报告显示 GPT-4o 仍有 30%–40% 的幻觉率，高风险企业级应用必须依赖资深工程师严格把控。 - **每日自动化模型回归**：在传统架构测试之外，为所有接入的 AI 模型建立每日 Prompt 基准测试体系，将其作为对抗非确定性风险的核心安全防线。 ### [09:30](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=570s) LLM 协作实验与创业团队的内部 AI 工具栈 团队在内部开发中进行了激进的 AI 实验，设立了一个专门隔离的 Monorepo（单一代码仓库），直接让大语言模型主导开发，其中高达 85% 的代码由 LLM 自主编写，人类工程师的参与度被严格限制在最多 15%。观察这种纯 AI 生成的代码库与人类完全手动编写的代码库如何相互调用与交互，让团队重新审视了人机协作的本质——AI 不再只是一个“输入代码、吐出结果”的简单补全工具，而是一个真正意义上的协同研发伙伴。 ```mermaid flowchart TD accTitle: 现代 AI 研发与代码合规流程 subgraph IdeLayer["开发者自由工具层"] IDE1["VS Code + GitHub Copilot"] IDE2["Cursor / Windsurf"] IDE3["Cline / Roo Code / Claude Code"] end subgraph SecurityLayer["模型安全隔离层"] Gateway["私有云托管网关 (Google Vertex 等)"] SecCheck["数据防泄漏 / 禁止供应商模型训练"] end subgraph CentralLayer["中心化代码管理与评审"] Repo["GitHub 集中代码仓库"] PR["严苛的 Pull Request 手动审查"] Main["合并至生产代码库"] end IdeLayer --> Gateway Gateway --> SecCheck IdeLayer --> Repo Repo --> PR PR --> Main ``` 在具体技术栈与 IDE 选择上，团队给予工程师极大的自由度，不强制锁定单一工具，员工可以根据习惯选用 Cursor、Cline、Roo Code 或 Claude Code 等。但为了确保企业数据安全并防止代码被外部大模型厂商用于二次训练，所有编辑器插件都必须统一接入企业在私有云（如 Google Vertex AI）上自建的模型实例。不仅技术岗位被深度赋能，非技术人员的生产力也发生了质的飞跃。例如增长负责人（Head of Growth）深度依赖 Deep Research 这类深度研究工具生成详尽的分析报告，其产出质量完全能够媲美以往需要专门雇佣初级分析师（Junior Analyst）才能完成的工作。 对于 Eyepop.ai 而言，工程团队同样没有强制推行 Cursor 或 Windsurf，而是让所有开发者默认从“Visual Studio Code 搭配 GitHub Copilot”开始起步。团队收拢和把控质量的核心机制在于 GitHub 仓库的严格规范：任何代码都绝对不允许直接提交（Commit）到主分支，必须统一走 Pull Request 流程，并在合并入主代码库之前经过严格的人工审查。在非工程团队中，Perplexity 和 OpenAI 等智能体（Agentic）工具也已被全面应用于日常的研发调研与创意生成。 - **开发者自由与集中管控**：支持开发者自由切换 Cursor、Cline、Roo Code、Windsurf 或 VS Code，但所有代码必须经过 GitHub PR 人工审核。 - **数据合规机制**：通过 Google Vertex AI 等私有网关接入 LLM，避免企业专有代码流出并被基础模型训练抓取。 - **非技术人员杠杆**：利用 Deep Research、Perplexity 等 Agent 工具替代基础分析师职能，显著提升人效比。 --- ### [12:09](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=729s) 创业浪潮中恒定不变的底层拷问：“为什么是你们？” 尽管当前的技术范式、开发工具与五到十年前相比发生了翻天覆地的变化，但在创业投资与底层商业逻辑中，依然存在着恒定不变的法则。作为指导过众多初创企业的资深创业导师，Andy Ballester 指出，在审阅当下的商业计划书时，创始人通常很容易说服投资人市场上存在一个巨大的真实痛点、存在一个可规模化落地的解决方案，并且该赛道具备极具吸引力的总体有效市场（TAM）。然而，无论技术如何迭代，创始人始终要回答那个最核心的问题——“为什么是你们团队（Why your team）？” 在 AI 技术大幅抹平通用开发门槛的今天，“为什么是你们”这一答案反而变得越来越模糊与脆弱。投资人与市场最终考核的依然是团队的综合护城河（Defensibility）：你们的研发与迭代速度是否比所有人更快？你们是否拥有独家、专有的数据访问权限从而构筑起数据壁垒？或者你们是否掌握着竞品必须从零艰难搭建的独特用户群体与分发渠道？这些关于防御性的核心拷问，与以往历次技术浪潮中的标准并无二致。 --- ### [13:34](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=814s) 跨越“AI 垃圾”泡沫：在高风险企业场景中建立高壁垒 当前 AI 赛道正在出现明显的分化，制造“AI 垃圾（AI slop）”的短期投机者与真正为客户创造核心业务价值的公司之间，差距正被急剧拉大。市场上充斥着大量依靠融资造势或虚报客户数据而“昙花一现”的公司，因为 AI 带来的全民狂热在无形中拉低了讲故事的门槛。但对于真正做深层技术的企业而言，真正的商业机会在于解决高门槛、高风险行业的底层阻力。 ```mermaid flowchart LR accTitle: 高风险企业 AI 落地核心价值 subgraph Enterprise["高风险行业客户 (银行 / 金融 / 保险)"] Req["严苛合规要求 & 内部风控政策"] end subgraph Traditional["传统解决路径 (高成本 & 高摩擦)"] Pain["堆叠 50+ 算法工程师 (MLE) 进行微调与提示词调试"] end subgraph CTGT["CTGT 核心壁垒方案"] Core["前沿学术实验室底层研究成果"] Engine["低摩擦自动化合规引擎 (免微调/免大量提示调试)"] end Req --> Traditional Req --> CTGT CTGT --> Core Core --> Engine ``` 以 CTGT 为例，其核心业务聚焦于银行、金融、保险等对合规性与安全性要求极其严苛的高风险场景。在这些领域部署 AI 存在巨大的摩擦成本：模型必须百分之百遵守行业监管与内部策略，绝不能产生越轨输出。过去企业往往需要组建由 50 名机器学习工程师（MLE）构成的庞大团队，耗费数月进行繁琐的模型微调（Fine-tuning）和提示词工程（Prompting）。CTGT 将这种底层复杂性全面抽象封装，为客户提供立即可用的安全部署方案。这种技术护城河源自顶级学术实验室极其罕见的前沿研究，这种稀缺程度堪比“Meta 等科技巨头开出天价薪酬争抢的顶尖研究员”。 在面对资源有限、营销预算不足的挑战时，建立客户信任的最有效方法就是践行 Y Combinator 的经典格言——“做出人们真正想要的东西（Make something people want）”。CTGT 放弃花哨的说辞，直接在客户的真实业务环境里与金融、银行领域的既有方案进行正面基准测试（Head-to-head live deployment）。这种敢于直接“拼刺刀”的产品自信，成为了打消大客户疑虑并建立信任的最快途径。 - **AI Slop 现象**：警惕仅凭短期融资或虚假客户指标繁荣的包装型项目，聚焦真实价值交付。 - **高风险行业痛点**：银行、金融与保险机构部署 AI 的最大瓶颈在于合规风险与高昂的人力调试成本。 - **验证护城河的极简逻辑**：通过与行业现有竞品进行现场真实业务的 Head-to-head 盲测与对比，用数据与效果直观建立客户信任。 --- ### [16:20](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=980s) 场景化演示与破除大模型厂商锁定的架构哲学 在产品推广过程中，客户最常出现的反应往往是惊叹：“天哪，我以前居然不知道 AI 还能做到这种事！”由于 AI 技术进化极为迅速，客户对其当前能力的边界缺乏明确认知。因此，持续输出具有标杆示范效应的内容并进行具象化的场景演示，其重要性丝毫不亚于底层产品的研发。像 Eyepop.ai 和 CTGT 这类横向能力极强的平台，能够跨越多个行业提供通用能力，但也正因如此，它们必须深入垂直场景，为每个细分行业量身定制具体的痛点解决方案，向客户清晰展示“原本需要昂贵 MLE 团队开发的项目，在系统上只需 10 分钟到 1 小时就能快速跑通”。 在探讨“调用基础模型 vs 训练自有模型”这一核心架构权衡时，创立一家公司的本质是对“未来世界将如何运转”下注。正如 Gmail 之父 Paul Buchheit 在 2000 年代初将 Gmail 作为 Google 周末项目开发时秉持的理念——“用 10% 的精力做出竞品 90% 的功能”，CTGT 对未来的核心判断是：企业绝不希望被单一的模型供应商（无论是 OpenAI 还是 Anthropic 的企业方案）深度锁定，而是强烈渴望能够在不同的 AI 方案与供应商之间保持灵活调度与定制的自由。因此，为企业提供能够解耦供应商锁定、灵活适配和定制多种模型的中间层架构，成为了极具前瞻性的战略抉择。 ### [18:17](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1097s) 保持模型不可知论：构建可快速热插拔的 AI 应用运行时 在模型微调（Fine-tuning）和提示工程（Prompt Engineering）层面，初创公司绝不能被某一个特定的底层大模型深度绑定。发言嘉宾指出，其平台的核心设计哲学是**模型不可知（Model Agnostic）**，允许客户灵活对接不同供应商的模型与不同的业务场景。平台不会强行替客户做选择，关键任务是确保客户无论调用何种模型，都能真正符合其投资回报率（ROI）目标与内部业务指标。 Andy Ballester 补充道，AI 技术迭代演进极快，如果过早地把客户锁定在单一模型上会带来巨大风险。为此，Eyepop.ai 专门打造了一套 **AI 应用运行时（AI Application Runtime）**，将其与市面上针对特定底座模型的专用运行时区分开来。这套运行时实质上是对模型功能抽象的一层封装（Wrapper），使得客户可以在底层随时无缝替换与热插拔最新模型。尤其在当下，各类新型计算机视觉模型（Computer Vision Models）与视觉语言模型（VLMs）层出不穷，没人能预判谁是最终赢家，甚至可能根本不存在单一赢家。因此，技术架构本身必须将这一层完全解耦并封装起来。 ```mermaid flowchart TD accTitle: AI Application Runtime Architecture accDescription: Architecture diagram showing how the AI application runtime encapsulates and decouples rapidly evolving models from business logic. subgraph ClientLayer["客户端与业务场景"] App["AI 业务应用 / ROI 与指标驱动"] end subgraph RuntimeLayer["Eyepop.ai 抽象层"] Runtime["AI 应用运行时 (Application Runtime)<br/>统一功能封装与标准化接口"] end subgraph ModelPool["底层可插拔模型池 (Model Agnostic)"] direction LR M1["计算机视觉 (CV) 模型"] M2["视觉语言模型 (VLMs)"] M3["LLM / 第三方供应商模型"] end App --> Runtime Runtime -.->|"无缝热插拔 / 快速切换"| M1 Runtime -.->|"无缝热插拔 / 快速切换"| M2 Runtime -.->|"无缝热插拔 / 快速切换"| M3 ``` --- ### [19:32](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1172s) 从 0 到 1 的致命陷阱：拿着锤子找钉子与原型陷阱 当主持人将话题引向初创团队从 0 到 1 的非技术陷阱时，Cyril Gorlla 犀利指出，最普遍的错误依然是老生常谈的“拿着解决方案找问题（A solution in search of a problem）”。即便是进入了录取率低于 0.5% 的极度严苛的 Y Combinator（YC）孵化器，在为期六个月的项目周期里也能清晰分辨出两类团队：一类在切中要害、解决能带来实质痛苦的真问题；另一类则仅仅是在做一个看起来很酷、但脱离实际商业需求的技术玩具。 随着 Bolt.new 等各类 AI 辅助研发工具的爆发，如今构建一个产品原型的成本“基本已经缩减至零”。然而，原型门槛的急剧降低虽然催生了更多创业公司，却也导致创始人对核心难题的思考变得浮躁。相比过去 Andy Ballester 创办 GoFundMe 时需要手动部署服务器等大量物理与工程约束，如今这些复杂性全被抽象化了。这造成了一个反直觉的现象：尽管做出初始版本（v0）极度容易，许多团队却迟迟不敢将其交到真实用户手中，借口是“仍在继续打磨完善”。这种闭门造车往往让产品沿着完全偏离目标客户需求的方向演进，埋下巨大隐患。 - **原型开发成本趋零**：生成式代码工具让任何人都可以在极短时间内搭建出产品 Demo。 - **精益创业（Lean Startup）原则依然适用**：必须在第一时间把产品放到真实受众面前，验证市场存在付费意愿（Willingness to Pay）与实际价值创造。 --- ### [22:15](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1335s) 双向赋能与新型招聘：非典型工程人才的崛起 AI 不仅让非技术人员能写代码，更反向赋能了技术人员。如今开发者无需庞大的 GTM（Go-to-Market）团队支持，便能利用 AI 研究工具快速搭建着陆页（Landing Page）、启动排队名单（Waitlist），或集成 Stripe 验证早期付费信号。产品与市场营销的专业能力正以极低成本注入到纯技术背景的初创团队中。 这一变革彻底重塑了早期团队的人才招聘逻辑。Cyril 提到一位市值 120 亿美元公司 CTO 的见解：过去初创公司往往盲目争夺排名前 20 的名校计算机系毕业生，而现在更倾向于招募那些或许计算机并非主修专业、但对产品有敏锐直觉、懂得如何与客户深度沟通的人才。借助 Cursor、Claude 等先进 AI 编程工具，这类具备优秀产品嗅觉的工程师能够飞速实现原型并根据用户反馈敏捷迭代。Cyril 以 CTGT 最新招募的一位前 NASA 专家为例，该工程师虽然此前并非标准全栈程序员，但凭借在天体物理学中沉淀的深厚数学与逻辑底蕴，结合现代 AI 工具，能够极快地理解（Grok）代码并高效落地。 --- ### [24:23](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1463s) 早期技术团队的人才拼图与反思 在竞争激烈的 AI 招聘市场中，两家公司展现了截然不同但同样高效的招聘打法： 1. **CTGT 的“洛斯阿拉莫斯（Los Alamos）”模式**：完全不发布任何公开招聘启事，完全依托极高信任度的学术圈与私人关系网络（如斯坦福教授等权威内推）进行严格背调与选拔，从而在高速变化的市场中彻底过滤掉不确定性。 2. **Eyepop.ai 的“以产品为招聘漏斗”模式**：让产品本身成为第一招聘界面。团队中每一位新入职的成员，最初与公司的接触都是作为其产品的真实使用者。他们亲身体验到了产品的价值与使命感，才选择加入。 在团队特质上，Andy 强调学术界与初创企业在节奏上存在巨大鸿沟。虽然团队需要极其专业精深的机器学习（ML）科学家，但更看重其能否适应初创公司的快节奏——既拥有深厚的技术积累，又具备敢于快速试错、在行进中修补缺陷（"okay with making mistakes and fixing them on the way"）的极客行动力。团队目前采取每周 3 天在办公室协同的混合办公制，坚持通过线下线面对面交流来考察候选人的综合适配度。 在复盘创业历程时，Cyril 坦言如果一切重来，他会选择“更早启动、更早从学校辍学（dropped out earlier）”。在 AI 狂飙突进的竞赛中，时间是唯一的最大筹码，唯一的制胜法则就是不惜一切代价更早地启动、更早地发版（Shipping），并尽早将产品推向真实客户。 ### [27:10](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1630s) 加速器的价值与指数增长：两天敲定种子轮融资 讨论首先围绕初创企业如何验证脑海中的构想展开。嘉宾们一致认同，若心中对某个想法是否真实可行尚存疑虑，最有效的验证方式就是尽快去接触并面对那些未来真正会使用该产品的人，唯有把节奏全面提速、更早地推向受众，才能避免闭门造车。当主持人询问 Y Combinator（YC）这类加速器在当今时代是否依然如过去那般具有价值时，Cyril 给予了极度肯定的回应。他坦言在 YC 的历程是自己经历过最棒的体验之一，团队通过这一网络获得了令人惊叹的顶尖导师资源，包括 Sam Altman、Airbnb 联合创始人 Brian Chesky 以及 Anthropic 的联合创始人等。 对于创业公司而言，加速器的核心意义在于改变成长曲线的斜率。Cyril 强调，创始人投身创业的前提必然是相信自己的公司能够走上一条指数级增长曲线，而 YC 的真正作用，实质上就是大幅缩短这条指数曲线前期漫长且平缓的摸索阶段。这最终取决于创始人如何衡量自身时间的价值——得益于这样的加速度，CTGT 仅用了两天时间就完成了由 Google 领投的种子轮融资。正如 Andy 所补充的对比，在过去两年相对严峻的创投环境中，一般的早期项目若是能在 8 个月之内搞定一轮融资，就已经算是相当出色的表现了。 - **顶级导师网络**：在 YC 期间获得了 OpenAI 首席执行官 Sam Altman、Airbnb 创始人 Brian Chesky 以及 Anthropic 联合创始人等重量级人物的直接指导。 - **极端高效的融资速度**：在普遍需要历经长达 8 个月左右融资周期的市场大环境下，CTGT 创下了仅用 2 天便敲定 Google 投资的种子轮战绩。 ### [28:16](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1696s) MVP 后的功能权衡：深度挖掘与横向拓展的拉锯 在初创期，时间永远是决定生死的最关键变量。Andy 回顾了 Eyepop.ai 的早期实践，坦言团队虽然很早就将产品发布了出去，并且在当时的技术与资源条件下，几乎已经不可能更负责任地比那个时间点发布得更早，但在最小可行产品（MVP）面世之后，团队在产品功能策略上走了一段曲折的探索之路。他们当时选择针对极少数几个产品功能进行了深度的钻研与打磨，但回过头来看，或许采取更广泛、更具横向探索性的铺开策略会更加明智，从而能在早期捕捉到更多层面的市场感知。 构建平台型产品的挑战在于，用户天然会对平台的完备性抱有预期，希望平台在某些既定维度上表现得无可挑剔。在只有极少数初始用户参与反馈的冷启动阶段，判断某项功能做到什么程度算作“足够好”（good enough），进而果断抽身转向下一个环节，是一项极其痛苦且充满不确定性的优先级决策。面对这种权衡取舍，Andy 感慨道，如果能在一些细小的产品细节上更果断、更迅速地进行微调转向（pivot），整个迭代过程将会顺畅得多。 ```mermaid flowchart TD accTitle: 平台型产品 MVP 后的功能迭代路径 subgraph S1["早期分歧抉择"] A["MVP 早期发布"] --> B{"功能迭代策略"} B -->|"实际路径：纵向深挖"| C["少数功能极度深度打磨"] B -->|"反思建议：横向拓展"| D["广泛涉猎探索多维感知"] end subgraph S2["决策挑战与解法"] C --> E["受制于早期极少量用户<br/>难以判断'足够好'的边界"] D --> F["保持敏捷度<br/>针对细微功能小步快跑迭代"] end ``` ### [29:04](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1744s) 重构 GoFundMe：用 Agentic AI 克服求助恐惧与赋能叙事 谈及如果现在利用现代 AI 技术重新打造众筹平台 GoFundMe 会有何不同时，联合创始人 Andy 分享了一个关于人性的深刻洞察。公众通常认为“公开发言”和“直面死亡”是人生中最大的两大恐惧，但 Andy 认为事实并非如此——人一生中最深刻的恐惧和在公共场合最难启齿的事情，其实是“向他人寻求帮助”。即便发起的筹款并不是为了自己，而是为了身处困境的朋友，代他人开口求助依然会让人感到极其艰难。 大多数人天生并不擅长讲故事，然而一次成功的众筹呼吁，恰恰依赖于将真情实感转化为能够打动特定受众的精彩叙事。这就构成了平台早期最大的瓶颈： - **AI 作为协作写手（Co-writer）**：大语言模型能够充当一个平易近人的写作伙伴，协助发起人梳理细节并消除心理负担，让他们明白公开发起求助并不难堪，某些细腻的情感细节也理应坦然呈现。 - **破除人力边际成本限制**：在生成式 AI 出现之前的很多年，GoFundMe 团队就曾构想过为每一位用户配备专属辅导员，手把手帮助他们撰写故事，但在纯人工运营模式下，这种全天候一对一的服务成本高昂到完全无法落地。 - **Agentic AI 带来的普及化解法**：如今具备自主能力的 Agentic AI 能够随时驻守在每一位用户身边，成为不知疲倦的专属叙事教练，Andy 称这是对众筹形态一个“难以置信的重大解锁（incredible unlock）”。 ### [30:50](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1850s) 非典型 AI 落地：挖掘沉默数据与客服情绪深度分析 当主持人追问是否有其他非传统的 AI 应用场景时，Andy 结合 Eyepop.ai 的业务逻辑与内部运营分享了独到见解。外界产品每天都在产生海量的图像和视频——无论是一家零售店监控摄像头下累积的无数小时录像，还是用户上传的上百万张图片与视频帧。在缺乏有效分析工具的情况下，这些未被加工的素材静静躺在云端存储桶里，对企业而言无异于毫无意义的“黑色像素（black pixels）”。 Eyepop.ai 除了致力于将这些沉默的非结构化图像数据转化为商业洞察，还在内部利用 Agentic AI 开辟了反传统的用法： - **跨平台非典型数据挖掘**：运用智能体对原本散落、孤立的客户数据进行纵深分析，探查用户对产品的深层心理状态。 - **客服工单智能汇总与情感提炼**：现有的主流客户支持平台普遍缺乏深入的问题溯源和情绪分析机制。团队引入 Agentic AI 对所有客户支持工单进行自动归纳与聚类，不仅能直击用户报障的核心痛点，更能穿透整个底层支持数据库，提供多位人工客服在日常沟通中极难沉淀出的高颗粒度情感分析（detailed sentiment analysis）。 ### [32:15](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1935s) 模型可解释性突破：无需微调重构 DeepSeek 的价值对齐 在底层模型技术的探索上，Cyril 揭秘了团队在白宫展示的一项重磅技术突破——针对中国近期开源发布的 DeepSeek 大模型展开的模型级可解释性（model-level interpretability）研究。当时一家全美排名前三十（Fortune 30）的大型金融机构主动找上门来，希望探究如何使该模型更贴合美国的价值体系与分析标准。因为在原始状态下，如果无法就天安门事件等特定历史事件向模型提问，该模型在面向经济走势预测、历史回溯等金融严肃场景时，其实用价值就会大打折扣。 面对这一诉求，CTGT 并未采用常规的暴力微调或容易失效的提示词工程，而是从模型激活机制切入： | 维度 | 传统干预手段（微调 / Prompt Engineering） | CTGT 的模型可解释性激活层编辑 | |:--- |:--- |:--- | | **操作方式** | 全参数重新微调，或编写极其冗长、脆弱的 Prompt 包装 | 深入神经网络底层，定位触发审查机制的具体激活神经元 | | **副作用表现** | 强行提升审查力度容易导致模型吐出无逻辑的胡言乱语（gibberish） | 精准剥离（remove）导致特定内容受限的模型激活模块 | | **版本稳定性** | 一旦模型版本更新或参数漂移，编写的复杂 Prompt 极易崩溃 | 平台化模块化修改，具备更高的可持续性与可扩展性（extensible） | | **实际效果** | 成本高昂、周期漫长且无法保证完全对齐 | 接入平台后模型即可自然讨论原本受限的议题，无需繁复微调 | Cyril 指出，这种深入模型内部激活机制的干预手段，为定制智能模型开辟了一条远比陈旧微调方式更可持续且具弹性的技术路径。 ### [33:28](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2008s) 雅典论坛式文化与从尖端科研到企业级落地的极速转化 当谈及未来半年到一年面临的严峻挑战与左右摇摆的困局时，Cyril 坦率地将其形容为“扑灭所有火苗并尽力让头浮出水面以防溺水（putting out all the fires and trying to maintain our head above water）”。在极速狂飙的超高速增长期，业务的急剧膨胀极易导致许多细枝末节被无意遗漏。这种白热化的状态倒逼团队在人才招募上不拘一格：他们的员工跨度极大，既有经验极其丰富、按常理不该出现在早期初创公司的高级别专家，也有来自伯克利、年仅 19 岁的年轻天才。 为了驾驭这种复杂的人才结构，Cyril 致力于打造一种高度扁平、崇尚真理的内部文化： - **雅典式辩论论坛（Athenian forum）**：效仿历史上孕育了划时代技术创新的施乐帕克研究中心（Xerox PARC）与贝尔实验室（Bell Labs），营造纯粹开放的学术思辨氛围，不讲资历、只让“最佳的想法胜出”，坚信唯有最佳的思想交锋才能打磨出最顶尖的企业级产品。 - **深度链接学术研究生态**：办公室频繁举办各类 Meetup 与深度研讨，甚至吸引了著名机器学习库 Keras 的创造者、也是其投资人的 François Chollet 等学术界泰斗亲临办公区发表前沿演讲。 - **全扁平化协作与快速推向生产**：在这种毫无官僚层级的环境中，连实习生编写的高质量代码也能直接合并，并最终稳定运行在《财富》500 强企业的服务器核心中。团队将自身的“超能力”归纳为将尖端科研（cutting-edge research）转化为客户生产环境可用代码的时间差（delta）压缩到极致。 在访谈尾声，Andy 同样展望了 Eyepop.ai 现阶段的攻坚重点。在经历了“从 0 到 1”打磨出 MVP、开启早期市场进入（GTM）之后，团队已经高度明确了客户的真正诉求。当前他们正全力向 A 轮融资（Series A）的里程碑冲刺，眼下最核心的挑战就是确保小体量团队全力以赴高速运转，在行业自然演化的基础 AI 能力之外跑出差异化的超额价值，始终在竞争中保持身位领先。 ### [36:16](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2176s) 早期团队的敏捷节奏、人文关怀与线下办公 在初创团队中，每个人往往身兼数职，甚至一个人就承担了三个部门的职责。在全力冲刺首个营收里程碑的过程中，如何既保持团队高产，又不让成员在混乱与重压中迷失或流失，是管理上的核心考题。为此，团队在敏捷开发流程中融入了大量团队建设和复盘反思环节，通过每双周一次的里程碑会议，坦诚讨论哪些机制奏效、哪些遇到阻碍，从而及时识别出人性层面的现实问题——例如某个成员需要休假喘息，或是当前团队推进过猛需要后撤一步、重新调整规划，以实现“通过稍慢一点的节奏来走得更快（move faster by moving slightly slower）”。 关于办公协作模式，早期创业公司高度依赖线下现场沟通。在产品快速试错阶段，如果发现 Bug 或设计缺陷，发一条 Slack 消息并约之后的 Zoom 往往效率过低，直接端起笔记本电脑走过去指出问题，或者在工位旁无意中听到同事讨论测试功能并及时纠偏“我们对产品理解有误，必须立刻转向”，能省去巨大的沟通损耗。CTGT 则实行完全全员线下办公（100% In-Person），团队成员在激烈的现场讨论中即兴碰撞灵感，这种“当下（in the heat of the moment）”的即时反馈被认为是孕育突破性创新的关键土壤。 - **敏捷复盘（Sprint Rituals）**：通过双周同步机制监控团队心理状态，防止成员在多重角色超负荷运转下过早倦怠（burnout）。 - **混合与全线下办公**：Eyepop.ai 采用每周 3 天到岗制，CTGT 采用全员全天现场办公，优先保障面对面交流的敏捷度与高频自发碰撞。 --- ### [38:21](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2301s) 从 0 到 1 的创业建议：构建受众壁垒与追求“90/10 方案” 针对现场准备从零起步的创业者，嘉宾们给出了两条极具实操性的核心建议。首先，创业者必须尽早明确要为极其垂直细分的目标受众提供何种确切价值。比起埋头敲代码，尽早开始沉淀和运营这部分核心受众群体同样至关重要——在技术迭代日新月异的时代，“拥有一批黏性受众本身就是一道极具防御性的护城河（defensible moat）”。即便未来产品愿景发生转型（Pivot），这批受众资产依然具备长期价值。 ```mermaid flowchart TD accTitle: 0 到 1 阶段创始人核心破局策略 subgraph S1["受众与壁垒建设"] A["明确超垂直客群的核心价值"] --> B["早期启动受众沉淀与运营"] B --> C["形成对抗技术迭代的受众护城河"] end subgraph S2["执行与方案破局"] D["锁定 90/10 极简替代方案"] --> E["以 10% 成本实现 90% 核心质量"] E --> F["发挥初创期无畏/敏捷优势实现弯道超车"] end ``` 另一方面，初创企业应当积极寻找“90/10 方案”——即用 10% 的精力/成本实现同类方案 90% 的质量。大公司和成熟巨头往往由于嫌弃方案过于简陋（scrappy）或偏见而弃之不用，但正是初创团队所具备的这种“初生牛犊不怕虎的无畏（scrappy naivete）”，能让这些方案在极低试错成本下快速成型，甚至在打磨成熟后全面超越现存的竞品。 --- ### [39:30](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2370s) 规避 AI 泡沫：拒绝薄层包装与伪需求陷阱 在圆桌尾声，讨论回归到底层防御性与行业清醒度的思考。面对市面上充斥的“AI 垃圾（AI slop）”，每一位创始人必须时刻自问：自己的产品是否仅仅只是套在 ChatGPT 之上的一层薄薄外壳（thin layer wrapper）？在基础大模型每隔一年就出现跨越式进化的背景下，这项功能是否会在数月内被彻底商品化（commoditized）？团队必须找到产品在未来依然有理由存在的硬核防御性逻辑。 ```mermaid flowchart LR accTitle: AI 创业决策对比：真需求 vs 伪概念 subgraph TrueAI["AI 原生（AI-Native）模式"] A1["深层业务痛点"] --> A2["非 AI 技术不可解"] A2 --> A3["构建长期稳健业务架构"] end subgraph FakeAI["追逐风口（AI Slop）模式"] B1["追逐融资热度/概念炒作"] --> B2["浅层套壳 / 强行嵌入 AI"] B2 --> B3["短期决策导致架构脆弱/一年内被淘汰"] end ``` 这种审视不仅适用于采购 AI 的企业客户，更关乎早期初创公司的生死存亡： - **警惕追逐热钱带来的隐形毒性**：若只是为了迎合投资热点而硬塞 AI 概念，会导致团队做出大量短视的架构妥协与战术决策，造就极其脆弱、活不过一年的公司架构。 - **验证 AI 原生性**：必须严谨求证业务场景是否存在非 AI 不可的真实痛点，确保产品在底层是真正的 AI Native，而非仅在表面贴上 AI 标签。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=4s) 我能否在图像和视频中定位、测量或统计特定物体与状态？ — Andy Ballester > [13:34](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=814s) 做出人们真正想要的东西（Make something people want） — Cyril Gorlla > [16:20](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=980s) 用 10% 的精力做出竞品 90% 的功能 — Cyril Gorlla > [19:32](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1172s) 拿着解决方案找问题（A solution in search of a problem） — Cyril Gorlla > [24:23](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1463s) 更早启动、更早从学校辍学（dropped out earlier） — Cyril Gorlla > [29:04](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=1744s) 向他人寻求帮助 — Andy Ballester > [33:28](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2008s) 扑灭所有火苗并尽力让头浮出水面以防溺水（putting out all the fires and trying to maintain our head above water） — Cyril Gorlla > [36:16](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2176s) 通过稍慢一点的节奏来走得更快（move faster by moving slightly slower） — Andy Ballester > [38:21](https://www.youtube.com/watch?v=ZZ27XE6J5C8&t=2301s) 拥有一批黏性受众本身就是一道极具防御性的护城河（defensible moat） — Andy Ballester ## Mentioned Ai4, Airbnb, Andy Ballester, Anthropic, Asana, Bell Labs, Bolt.new, Brian Chesky, Claude, Claude Code, Cline, CTGT, Cursor, Deep Research, DeepSeek, Eyepop.ai, Firebase, Fortune 500, François Chollet, General Catalyst, GitHub Copilot, Glean, Gmail, GoFundMe, Google, Google Vertex AI, GPT-4o, Keras, Meta, NASA, OpenAI, Paul Buchheit, Perplexity, Rebecca Hinds, Roo Code, Sam Altman, Stanford University, Stripe, Tea App, Visual Studio Code, Windsurf, Xerox PARC, Y Combinator ## Video description Cyril Gorlla, Founder & CEO, CTGT Andy Ballester, Co-Founder, Eyepop.ai Rebecca Hinds, Head, Work Transformation Lab, Glean Abstracts Starting an AI company in 2025 demands more than a cool demo. From model strategy to infrastructure choices, regulatory landmines to defensibility, AI founders today face a maze of complex, high-stakes decisions. In this tactical session, proven startup operators break down what actually works: how to choose your technical architecture, when to use foundation models vs. train your own, how to access proprietary data, and what metrics matter at each funding stage. You’ll get crisp insights into go-to-market timing, team design, and which AI infra decisions to lock in early. No fluff—just concrete frameworks, battle-tested lessons, and scalable strategies for AI-native startups. [Source](https://www.youtube.com/watch?v=ZZ27XE6J5C8)

[0:04] Andy Ballester: 我能否在图像和视频中定位、测量或统计特定物体与状态？ [13:34] Cyril Gorlla: 做出人们真正想要的东西（Make something people want） [16:20] Cyril Gorlla: 用 10% 的精力做出竞品 90% 的功能 [19:32] Cyril Gorlla: 拿着解决方案找问题（A solution in search of a problem） [24:23] Cyril Gorlla: 更早启动、更早从学校辍学（dropped out earlier） [29:04] Andy Ballester: 向他人寻求帮助 [33:28] Cyril Gorlla: 扑灭所有火苗并尽力让头浮出水面以防溺水（putting out all the fires and trying to maintain our head above water） [36:16] Andy Ballester: 通过稍慢一点的节奏来走得更快（move faster by moving slightly slower） [38:21] Andy Ballester: 拥有一批黏性受众本身就是一道极具防御性的护城河（defensible moat）

[12:09] Cyril Gorlla: The moat there is that we're saving them a lot of pain on that side... it's based on our technical research that basically very few people in the world actually know how to do... it's very similar to what you see in sort of like the huge comp packages that Meta is paying these researchers because their knowledge is so, so rare in the world. [15:26] Andy Ballester: Oh, I didn't know you can do that. [29:04] Andy Ballester: asking for help is people's most number one fear and the hardest thing for them to do in public. [33:28] Cyril Gorlla: putting out all the fires and just trying to maintain like our head above water. [35:35] Andy Ballester: step back, reset the plan, and figure out how to move faster by moving slightly slower. [35:35] Andy Ballester: I can't Slack somebody to Zoom them later and show them the bug. I need to pick up my laptop and show, 'Hey, we got a problem here.' [38:21] Andy Ballester: Audiences are a defensible moat in this day and age.

[0:04] Thank you so much everyone for being [0:05] here. It definitely is a packed house [0:07] and I want to thank Sarill and Andy for [0:10] joining us. I think it's going to be a [0:12] fun conversation and it's also a really [0:14] important impactful one. So I'm Rebecca. [0:16] I lead the um thought leadership at [0:20] Glean. I recently joined Glean the AI [0:22] work platform. Prior to Glean, I was at [0:25] ASA where I led their work innovation [0:27] center, essentially our internal think [0:29] tank, and did a lot of research on AI. [0:32] I'm going to pass it on to the the real [0:34] stars of the show and talk to us a [0:36] little bit about your journey as startup [0:39] founders. [0:40] >> Yeah, thanks Rebecca. Um, my name is [0:42] Sirill Gorla. I'm the co-founder and CEO [0:44] of CTGT. I left my research at Stanford [0:47] when I was 23 um just a some months ago [0:50] uh to start the company based on my [0:52] research that had been going on for a [0:53] few years trying to understand how AI [0:55] models really think on a fundamental [0:57] level um and this is a problem that's [0:59] become only more preient because of [1:01] companies like Glean um and folks that [1:03] are trying to deploy AI in high-risisk [1:04] scenarios right so like banking finance [1:06] etc we really have to understand what [1:07] these models do and be able to trust [1:09] that they're actually doing what we [1:10] intend them to do um so that's [1:12] essentially what the company does um in [1:14] February We raised a uh first round of [1:16] funding led by Google, General Catalyst, [1:18] a couple other folks. Um we're sort of [1:20] scaling rapidly. So yeah. [1:22] >> Awesome. Hey everybody, my name is Andy [1:24] Ballister. I have been a startup person [1:27] my entire career. I'm a serial founder. [1:30] I I was the CTO and co-founder of uh [1:34] GoFundMe and now the chief product [1:37] officer and co-founder of ipop.ai. [1:40] In 2021, coming out of COVID, I was [1:44] mentoring at a venture studio in uh San [1:47] Diego where I live and we saw a cohort [1:51] of four different companies that were [1:54] trying to utilize computer vision at the [1:56] time and just starting from fundamental [1:59] technologies and really divergent in the [2:02] way that they implemented computer [2:04] vision. Uh even though that it they were [2:07] all fundamentally solving the same [2:08] problem. Can I find, measure, or count [2:11] objects or conditions in images and [2:13] video? And so we saw that as a call to [2:16] action to put together a platform that [2:18] makes it just dead simple to add AI [2:21] vision to your product and and that's [2:23] IPOP. It's a self-service computer [2:25] vision platform. [2:27] >> Wonderful. And you've both spent many [2:29] years in the startup realm, at least [2:31] tangentially. talk to us about why it's [2:34] so exciting right now to be a startup [2:36] founder specifically around AI. [2:39] >> You know, I think the rate at which [2:40] enterprise are actually adopting AI is [2:42] different from other waves like [2:44] blockchain uh web 2 earlier. I think the [2:46] buyin that we're getting and the amount [2:48] of excitement we're seeing around even [2:51] like very large institutional [2:52] enterprises um with typically a lot of [2:55] red tape to onboard new vendors. Um I [2:58] think the problem is sort of so real and [3:00] the ROI from AOI is so apparent. Um [3:02] there's so many roadblocks to get there, [3:04] right? You have to there's usually like [3:05] a center of excellence that you have to [3:07] go through and then there like seeds use [3:08] cases all around um the organization. [3:11] And so there's a lot of red tape to [3:12] basically cut through and I think [3:14] anything that can make um AI more [3:16] transparent, more understandable um and [3:19] more trustworthy fundamentally that's [3:21] like a really hot pain point. And I [3:22] think that contributes to a lot of the [3:24] excitement that we see um even for an [3:27] early stage company like us. [3:29] >> Yeah. And I think what I'm I'm most [3:31] excited about is my entire career has [3:32] been around where product meets [3:34] technology and delivering those magical [3:36] experiences to people. And with AI, you [3:40] can deliver just incredible experiences [3:43] that would be prohibitively hard to to [3:46] code like use cases around when you have [3:50] LLMs. uh being able to interpret intent [3:53] and work within a system, you can [3:55] deliver these things that are that are [3:57] just really intelligent. And so I think [3:59] that that's a really exciting reason to [4:01] be a founder right now in in any space. [4:04] I also think the ability um of founders [4:07] is just so much more now. Um there is an [4:10] expectation that you're going to do [4:11] more, but it it gives us the ability to [4:14] do a lot more things with a much smaller [4:16] team. [4:17] >> I'd love to talk about that expectation. [4:19] How do you see it show up in terms of [4:21] how founders go about building? [4:24] >> Yeah, we've seen it over the last couple [4:27] of years, um, in a in a myriad of [4:29] different ways. One, we just see [4:31] companies starting out of nowhere and [4:33] and just having a lot of stuff, a lot of [4:35] product, a lot of marketing. And so, um, [4:37] I think that contributes to the speed at [4:40] which we have to produce as a small [4:41] team. But we also see it as an [4:43] expectation from investors. They want to [4:45] see um that you can do more with with [4:49] less. Right now, especially in the last [4:50] two years, funding's been fairly tight. [4:52] And so being efficient in terms of human [4:55] capital is really important. And so I [4:58] think that that expectation drives a lot [5:01] of generalist behavior. Um expertise is [5:04] is definitely needed in certain areas, [5:06] but in a lot of other ones, it's it's [5:08] okay to just be good enough at it. and [5:11] using AI adds a level of expertise that [5:13] you may not have in the company. [5:15] >> Yeah, we we've been doing a lot of [5:16] research on this shift around [5:19] generalist, specialists, how do you go [5:21] about, you know, leveraging the [5:22] tradeoffs? Is there a risk and what is [5:25] it in terms of leveraging a more [5:27] generalist mindset specifically as [5:29] you're building? Yeah, I think a great [5:30] example of this is like the advent of [5:32] vibe coding and you know we've seen a [5:34] lot of Andy and I were just talking over [5:36] dinner last night about the T app which [5:37] is um an app that was basically vibe [5:39] coded gained a lot of popularity and [5:41] then its database was uploaded to an [5:43] unsecured Firebase storage link [5:45] essentially and so anybody could see the [5:46] ids that were uploading all the personal [5:48] information uh that was disastrous. I [5:51] think the most effective people in our [5:52] organization, you know, we have a really [5:54] high bar for talent and the level of [5:55] talent we're getting is also like [5:56] unprecedented um that are joining early [5:58] stage companies. And I think what really [5:59] makes that possible is, you know, folks [6:02] that understand the underlying [6:04] fundamentals of software engineering, [6:05] right? They can actually tell when the [6:06] AI is going wrong. And so if they can [6:08] understand that abstraction level, then [6:10] they're able to operate, you know, with [6:11] the LLM uh as a, you know, as a useful [6:14] assist rather than like just trusting it [6:16] to make really critical stuff. And that [6:17] means they can effective like 10x, 50x, [6:19] even 100x their productivity. Um, which [6:22] has allowed us to keep a much leaner [6:23] team than we otherwise would have [6:24] needed. So I think um, making sure that [6:27] they actually understand the [6:28] fundamentals of what they're using AI [6:29] tools for, that's still very important. [6:31] >> And is there a way you've learned in [6:32] terms of how to test for that throughout [6:34] the interview process or [6:36] >> Yeah, it's kind of interesting because [6:38] um, we sort of seen a return to like [6:39] very in person. So our entire interview [6:41] person is now our entire interview [6:43] process is now all in person. Uh, we do [6:45] whiteboarding in person. and we fly [6:47] people out. Um, and that's to really [6:48] make sure there's not only culture fit, [6:49] but that they actually have the skills [6:51] they claim to, right? Um, it's really [6:53] important, especially when you know you [6:55] are dealing with Fortune 500 clients [6:56] that have really high standards for data [6:58] security, privacy, etc. You can't really [7:01] afford to relegate that um completely to [7:03] like LLMs, which still uh make a lot of [7:06] mistakes. Um, GPT04, if you look at the [7:08] technical report that Open released, um, [7:11] they report a hallucination rate of [7:12] around 30 to 40%. Which is still sort of [7:14] an issue. So um we just have to be very [7:17] mindful of that. [7:17] >> Definitely. [7:19] >> Yeah. I think to to add on to that um in [7:23] startups there's always there's always [7:25] been this push to move really quickly [7:27] and the balance between rigor and [7:30] process and moving fast is always [7:33] something that's top of mind for for me [7:35] and any startups especially when you're [7:36] in the zero to one phase because there's [7:39] just not enough time to to move slowly [7:41] about things. [7:42] that rigor is is even more important now [7:46] and putting putting good processes [7:48] around how do we know this is coded well [7:50] or how do we know this product is going [7:52] to work for our users um there's so much [7:55] that when you're using LLMs the [7:58] non-deterministic trap of of either you [8:02] know vibe coding your way into something [8:03] where you don't know what you don't know [8:06] um on architecture can lead to to [8:09] disasters but same thing on product if [8:11] you're if you're not 100% sure that the [8:13] person at the on the other end of your [8:16] product is going to have a good [8:17] experience there, then probably more [8:19] process is needed there to to figure out [8:21] how to get to that shity. [8:24] >> And let's let's get just a little bit [8:26] more specific in terms of that process. [8:28] Is there something concrete that you've [8:30] done where you're more confident now [8:32] that that process will be in in place? [8:35] Yeah, I think you know uh separate from [8:38] traditional software development. [8:40] They're um you know there's you know we [8:44] have things like load testing and and [8:46] architecture reviews and all those [8:47] things are kind of beaten into the [8:50] software development process, but now [8:51] you're adding an extra layer of [8:53] experimentation and some sort of quality [8:56] metrics that's being taken on an ongoing [8:58] basis. And so all of our models that we [9:01] bring into the system undergo like [9:04] rigorous checks every day. We have [9:06] automated tools that that go and put [9:08] known prompts in there and expect [9:10] certain outputs. [9:11] >> Um, and that's a level of of software [9:14] development that like is is newer to me [9:16] coming in from more traditional software [9:19] development in the past, but that's the [9:20] types of things that we need to put in [9:22] there as a as a as a safety valve. [9:25] >> Yeah, it's a really interesting point [9:26] actually. Um, with the advent of agentic [9:28] coding tools like cloud code, uh, it's [9:30] been very, it's been just cool to see [9:32] these tools. We have a sort of a monor [9:34] repo that we've isolated that we [9:36] basically just let only LLMs work on, [9:37] right? It's like 85% LMS and only 15% [9:40] maximum is allowed to be humans. And [9:42] it's just interesting to see what these [9:43] tools sort of go and do. And in fact, [9:46] seeing how those code bases interact [9:48] with the code bases that we create like [9:50] totally manually. um the interplay [9:52] between those things and the way that we [9:53] can actually think about these tools as [9:56] more of partners than just something we [9:57] put code in like a response comes out. [9:59] Um that's also been very interesting to [10:01] see. [10:02] >> So what's what's in your tech stack as [10:04] you're building? What are those critical [10:05] technologies? You've mentioned some of [10:07] them that you're anchoring around. [10:10] >> Yeah. So we use quite a lot. Um you know [10:12] we basically give employees the [10:14] flexibility in terms of actual IDs to [10:16] use whatever sort of fits them best. Um [10:18] some people use cursor some people use [10:20] client some people use rue and the key [10:22] there is we basically hook it into our [10:24] own instance of those particular models [10:27] whether it's on like google vertex or [10:29] something uh just to ensure that you [10:30] know data security and stuff is is sort [10:33] of uh protected because we can't you [10:35] can't let the companies train on that um [10:37] in terms of like you mentioned coding [10:38] tunes I mentioned cloud code but another [10:40] one that uh James our head of growth [10:42] who's here really likes is deep research [10:44] um because it essentially allows you to [10:46] get like very deta detailed reports on [10:48] something that would have required you [10:49] to basically hire in the past like [10:51] probably like a junior analyst um and [10:53] basically does the level of work at that [10:54] quality. Um so it's been really [10:56] interesting to see across like not only [10:57] technical fields but also uh folks that [10:59] are doing nontechnical stuff how AI has [11:01] really empowered them to streamline [11:03] their workflows um and you know make [11:05] their head count more efficient than [11:07] they would usually be. Mhm. Yeah. [11:08] >> Yeah. I think similar we're we're not at [11:10] the place where we've mandated you have [11:13] to use cursor, you have to use Windsurf [11:16] or or any of those tools, but we do [11:18] start all of our developers off in uh [11:20] Visual Studio Code with um GitHub [11:23] Copilot on top of it. If they want to [11:25] veer away from that, that's fine. Where [11:27] we come back together is, you know, we [11:29] centralize on GitHub repos. Nothing's [11:32] committed directly. It's all through [11:34] pull requests. And so there's a there's [11:36] a lot of manual check before code gets [11:38] merged into any of our code bases. Um [11:41] and so I think that's interesting. I [11:43] think also like serial is mentioning the [11:45] uh the the ability for our nontechnical [11:49] folks to use agentic um sources as well [11:52] for research and development and [11:55] creative. It's it's awesome. It's just [11:57] amazing what you can do. Uh and so some [11:59] of the things we use there is uh we're [12:01] heavy users of perplexity of of course [12:04] open AI um and uh just a myriad of [12:07] different tools that help us do our [12:09] work. [12:09] >> Definitely. So we've talked a lot about [12:11] what's different now, how things may be [12:14] different than 5 10 years ago. What's [12:16] the same? And Andy, I'm especially [12:18] interested and you've been through this [12:20] rodeo many times. What is something that [12:22] you've learned 5 10 years ago that you [12:24] think is still consistent in terms of [12:26] being important for startup founders? [12:29] >> Yeah, the thing that uh I I'm still a [12:31] mentor of of a lot of different [12:32] startups. So I have a decent purview [12:34] over across a lot of different [12:36] industries that are starting up right [12:37] now. And I think what remains the same [12:39] is the why you question. And so in pitch [12:43] decks that I see, you can often convince [12:45] somebody there is a problem here and [12:47] there is a solution that if implemented [12:50] will solve that problem on a massive [12:53] scale and there's great TAM there and [12:54] all that good stuff. But the the [12:56] questions founders always have to answer [12:57] in those interviews are okay so why your [13:00] team and that why your team is becoming [13:04] blurriier and blurriier. And so the [13:05] things we look at is okay well do you [13:07] have are you is your team faster for [13:10] some reason? Do you have access to data [13:13] that no one else has and that [13:14] proprietary data is going to give you an [13:16] advantage? Do you have access to an [13:18] audience that no one has access to or um [13:21] would have to build from scratch if they [13:23] were to to come chase you? And so those [13:25] are the types of like defensibility sort [13:28] of questions we we talk to founders and [13:30] those largely stay the same as like what [13:32] it was before. [13:34] So there is this narrative around [13:36] defensibility and how it is more [13:38] challenging than ever to build this moat [13:40] that is enduring. Do you agree with that [13:42] sentiment or do you think it's just as [13:44] possible now as it was in the past? [13:47] >> I think the gap between folks that are [13:49] building you know what we effectively [13:51] term AI slop and people that are [13:53] building you know solutions that really [13:55] drive um value to customers. I think [13:58] that gap is widening right cuz we see we [14:00] were just talking about this too. we see [14:01] a lot of companies that burn very [14:02] brightly whether it's through funding or [14:04] supposed traction but then it turns out [14:06] like they were lying about their [14:07] customers or something like that so that [14:09] that particularly has become more um [14:12] more prevalent because the bar has sort [14:14] of been lowered by AI of like there's [14:16] just so much fervor and excitement [14:17] around these things right [14:18] >> but I think how we've approached it and [14:20] what's really resonated with our [14:21] customers is you know thinking [14:23] fundamentally about the value that we're [14:25] driving right we know our customers want [14:27] to deploy AI in like very high-risisk [14:28] use spaces um like banking like finance [14:32] like insurance right and we know there's [14:33] like fundamental friction in deploying [14:35] AI in those use cases and so the value [14:38] there is providing a streamlined way to [14:41] you know essentially ensure that your AI [14:43] is doing what you want it to do that [14:45] it's not going against like regulations [14:46] in that space that it's actually [14:48] adherent to your policies um without all [14:50] the friction that you need to do today [14:51] like fine tuning and prompting right so [14:53] there it's like very clear that we're [14:54] abstracting away the complexity that [14:56] they would typically task like 50 MLE to [14:58] do right and so there like the moat [15:00] there is that we're saving them a lot of [15:02] pain on that side, right? And you know, [15:04] like Andy said, there's a lot of [15:06] different ways you can approach it, but [15:07] for us, it's based on our technical [15:09] research that basically very few people [15:10] in the world actually know how to do and [15:13] um you know, basically only my lab, [15:15] right? And so that means that it's very [15:17] similar to what you see um in sort of [15:19] like the huge comp packages that Meta is [15:21] paying these researchers because their [15:23] knowledge is so so rare in the world. Um [15:26] so that can be a pretty strong signal. [15:28] And how do you demonstrate that moat to [15:30] customers as you're, you know, have [15:32] limited resources, limited marketing [15:34] budget? How do you develop that trust? [15:36] >> I mean, we find that just uh build [15:38] building something that people want uh [15:40] to use the old YC tagline is sort of [15:42] what really resonates with people, [15:44] right? And so fundamentally getting [15:47] people to understand, you know, what [15:49] your product does, the best way is just [15:51] to show them. And so our demos and our [15:53] deployments, our test deployments are [15:55] like very effective in terms of we're [15:57] like go headto-head with any other [15:58] solution that you're putting us against [15:59] um that you typically use for like any [16:01] deployment in these use cases and [16:02] there's like a couple each for like [16:03] banking, finance, etc. But we just tell [16:06] them like go headto-head and uh see [16:08] which one works and we are confident [16:09] that ours works. So um the the issues [16:12] come when you're very like catty around [16:14] your deployment and you're like no, you [16:15] can't put a head down on this because [16:16] it's like a different solution, etc., [16:18] etc. Um, but [16:20] >> Andy, how about you in terms of demoing [16:22] any? [16:23] >> Yeah, I I think more than ever, the [16:25] question of, "Oh, I didn't know you can [16:28] do that," comes up a lot with customers. [16:30] And so, um, putting exemplary content [16:34] out there and showing how to use these [16:36] technologies is just as important as as [16:38] having the product that can do these [16:40] things because, [16:42] um, there's so much power there and it's [16:44] changing so rapidly that customers [16:46] really don't have a good sense of what [16:48] can I do with some of these technologies [16:50] right now. And so, especially our [16:52] company, um, iPop.ai, AI, CTGT, both [16:56] fairly horizontal sort of companies. We [16:59] can serve a lot of different industries. [17:01] So, we have to be really specific in [17:03] each industry we're targeting of showing [17:05] them examples from their industry of [17:07] here's a way you can solve a problem [17:08] that is a painoint and um the solution [17:12] to which would give you a lot of value [17:14] and here it is on our system and hey, [17:16] you can spin this up in like 10 minutes, [17:17] right? Or it's going to take an hour to [17:19] do this, but it's all there and you [17:21] don't need a machine learning engineer [17:23] to to get there. [17:25] So, how do you think about building in [17:27] terms of using foundational models [17:29] versus building your own? How have you [17:31] navigated that trade-off? [17:33] >> Um, you know, fundamentally when you [17:35] make a company, you're basically making [17:36] a statement about the future of how you [17:38] think some world would work. Um, Paul [17:40] Bukat, he's a creator of Gmail. [17:42] Basically, for him, when he made Gmail [17:43] as a weekend project at Google in the [17:45] early 2000s, he basically said, you [17:48] know, I want to make an email client [17:49] that's like basically does 90% of what [17:51] the other ones do for 10% of the effort. [17:52] and it worked pretty well, right? And so [17:54] for us, our bet on the vision of the [17:56] future is that companies want the [17:57] flexibility to use different AI [18:00] solutions, different AI vendors, and [18:01] they don't really want to be locked into [18:04] um just like Sam Alman's enterprise [18:05] solution or like Anthropic's enterprise [18:07] solution, right? That's sort of what [18:08] we've seen. And I think the natural [18:11] extension of that is building a [18:12] framework that allows them flexibility [18:14] to, you know, customize their models, [18:16] like do the things they typically would [18:17] do like fine-tuning, prompt engineering, [18:19] but not have that beholden to a specific [18:21] model, right? And so our platform very [18:23] much enables the use of different use [18:26] cases, different vendors. We we are [18:27] actually model agnostic, right? We don't [18:28] care what what models they use. Um we [18:31] think that that choice should be left to [18:32] them and that the most important thing [18:34] we can do is essentially ensure that [18:36] whatever they're using that it's [18:38] actually aligned with their their ROI [18:40] goals um and their internal metrics. [18:42] >> Yeah. [18:43] >> Yeah. We find the space is moving too [18:45] quickly to to prematurely lock our [18:48] customers into any one particular model. [18:50] And so we um developed an AI application [18:54] runtime and we're trying to [18:56] differentiate that from like the model [18:57] runtimes out there that are model p uh [19:00] specific but it's a wrapper around the [19:02] concept of what these models do so that [19:05] we can easily swap them in and out for [19:07] our customers. And um we're seeing that, [19:10] you know, uh we're seeing that in a lot [19:12] of different ways, but but especially [19:13] right now there's new um you know, [19:16] there's new computer vision models [19:17] coming out, there's new VLMs. We don't [19:19] we don't know who the winner is going to [19:21] be there eventually or if there even [19:23] will be one. Um and so we need to be [19:26] agnostic to that layer. And so that has [19:28] to be encapsulated in our technology. So [19:30] it's very easy to switch out. And when [19:32] we think beyond the technical aspect, [19:35] what do you see as the most common [19:37] pitfalls startup founders are making [19:40] right now in terms of building going [19:41] from zero to one? A lot of it is a [19:44] solution in search of a problem and [19:46] that's an old age problem as Andy can [19:48] attest attest to a lot of companies [19:50] right now um and I you know I even I saw [19:53] this when I went through the Y cominator [19:54] program which is like extremely [19:55] selective right less than like 0.5% [19:56] accept substance rate but even then like [19:58] the sort of like six-month outlook of [20:00] that program we can sort of see really [20:01] well who was actually tackling a problem [20:05] that was causing a real pain to people [20:07] and who was sort of just doing something [20:08] cool that's like great but then it [20:10] doesn't actually solve a real business [20:11] need And I think that's fundamentally [20:13] what you have to come back to, right? Is [20:16] what is the real problem that you're [20:18] solving. And I think especially with AI, [20:20] the sort of cost to build a prototype is [20:23] basically scaling to zero, right? With [20:24] like bolt.dev, blah blah blah, etc. And [20:27] so since the bar is so low for making a [20:29] prototype, um there's naturally like a [20:31] lot more companies coming up. And that's [20:32] like a great thing that also means that [20:34] people are perhaps thinking less about [20:36] the hard problems that they may have had [20:38] to think more about. um such as when [20:40] like Andy started GoFundMe, there was [20:41] just like like a lot of physical [20:42] constraints like putting something up on [20:44] a server, right? Um and like all that [20:46] like physical work that you would have [20:47] to do that's now abstracted away. Um so [20:49] I think you know ensuring that this is [20:51] you're building something that's like [20:53] people actually want I think that's just [20:54] like fundamentally a thing that I still [20:56] see um people leaving until it's too [20:58] late. [20:59] >> And any concrete recommendations as to [21:01] how you go about doing that other than [21:03] >> Yeah, I mean getting it out there is [21:05] important. And I see people that work on [21:07] uh sort of counterintuitive to the [21:09] easiness of deploying a a prototype. Um [21:12] people wait a lot until they actually [21:14] put it in the hands of customers, right? [21:15] Like they'll say they have this company, [21:16] they built like this initial vzero, but [21:18] then before they actually deploy it to [21:20] their ideal customers, um they'll wait [21:21] like quite a long time. They'll say [21:22] they're refining it. And that means that [21:24] you're doing all this refinement in a [21:26] direction that might not even be [21:27] relevant to who you want to use the [21:29] product, right? And so that basically is [21:32] uh is a good setup for trouble down the [21:34] line. Um, [21:35] >> so that's that's one thing that I'd say. [21:36] >> Yeah, validating your your solution out [21:39] there in the market is create key. Uh, [21:42] like Sel was saying, you know, the I [21:44] think all those great common startup [21:47] practices that we have before the advent [21:50] of I like still still uh are great here. [21:53] You need to prove out that there is an [21:55] audience here willing to pay for this [21:57] solution that there is um value creation [22:00] actually being created by your product [22:02] uh for these customers. And so getting [22:04] it out there, getting in the hands of [22:05] the right audience early will give you [22:08] good directionality and a willingness to [22:10] pay indicator [22:12] >> um which then allows you to move in the [22:14] right direction. [22:15] >> So it's very lean startup. Yes, [22:17] absolutely. What's different? What's [22:18] different in terms of specifically [22:21] identifying the pain, making sure [22:22] there's a need for it? Any specifically [22:25] are using AI to go about that process? [22:27] >> Yeah, I think there's there's um better [22:29] research tools that we have now. I think [22:31] that just the ability for [22:34] um you know developers who maybe hadn't [22:37] focused on go to market strategies [22:39] before to spin up a uh an early landing [22:42] page and get weight list um going or to [22:45] integrate with Stripe to get a signal of [22:47] payment early. That's all so much easier [22:50] now. And so it goes both ways. It's not [22:52] just non-technical folks using AI to be [22:55] more technical. It's it's the other way [22:56] too. How do you add product expertise [22:59] and marketing expertise onto technical [23:01] teams that wouldn't have otherwise had [23:02] that without additional hire? [23:04] >> Yeah, one really interesting thing that [23:05] I got from the CTO of a $12 billion [23:07] company that I was talking to a few [23:09] months ago is the dynamic of early stage [23:12] hiring has sort of changed a bit. [23:13] Whereas before it was very common to [23:15] sort of just hire from the top, you [23:17] know, 20 computer science schools. Now [23:20] there's much more of an impetus for you [23:23] to hire people that are maybe not their [23:25] primary um expertise may not be in [23:28] computer science or technology but they [23:30] actually have a really good sense for [23:31] how to talk to customers and incorporate [23:33] feedback on the product right and so [23:35] that means you can actually hire people [23:37] um as engineers like very early who are [23:39] you know adept at you know using tools [23:41] like cursor like claw etc but they have [23:44] a really good knack for product and [23:46] understanding what really drives you [23:47] know customer fervor in momentum, right? [23:50] And so you can now hire those people at [23:51] a much early stage than you ever can [23:53] before. So the actual dynamics of early [23:55] stage hiring have sort of um shifted in [23:57] that sense. One thing that's interesting [23:59] for us is we've been able to hire like [24:00] we you know we just hired this guy from [24:02] NASA um that we typically I don't think [24:04] we would have hired um but he's just [24:07] able to like gro things like very [24:09] quickly on the coding side to where his [24:11] expertise and like his raw expertise in [24:12] like astrophysics and stuff makes it [24:14] really useful um to be able to have like [24:16] those mathematical foundations [24:19] >> um and he can get up to speed on the [24:21] other stuff like pretty quick. So that's [24:22] that's one example. [24:23] >> And hiring, how do you how do you hire [24:26] right now? How do you stand out in a [24:27] very crowded AI market, especially for [24:30] technical folks? [24:31] >> Yes, so we have a bit of a copout. We [24:32] actually don't have any job postings. [24:34] Um, I like to say it's a little bit like [24:35] Los Alamos cuz we think the problem is [24:37] just as important. Um, it's all through [24:39] our networks. It's all through [24:40] referrals. It's like who you know. And I [24:42] think that that's really useful because [24:43] again, as Andy said, the rate at which [24:45] the market is moving, um, it's really [24:48] when you have like sort of a vetting [24:50] process done by like a professor I [24:51] really trust at Stanford or something [24:53] like this, um, it removes a lot of the [24:54] uncertainty and we've found that that [24:55] works like really well. Um, so that's [24:58] that's pretty much like the way to get [24:59] into CCGT. [25:01] >> Yeah, we're we're a small team so we're [25:04] able to leverage uh our co-founders [25:06] networks to to hire, but you're not [25:08] going to find everyone that way. And so [25:10] I would say two things. one our we let [25:13] our product also speak for us on the [25:15] hiring front too. And we've found that [25:18] every single person we've hired in the [25:19] company um their first interaction with [25:22] us has been through the product. And so [25:24] that that product has to work for them. [25:27] They have to see the value that it [25:28] creates themselves before they're going [25:30] to want to join us as a as a mission. [25:32] And so we think of the product as from [25:34] that angle as well. Um and we also think [25:37] about who we're bringing into the team. [25:39] We're a small startup. We need to move [25:41] quickly. The we need very specially [25:45] trained machine learning engineers to be [25:48] able to provide that intelligence to our [25:50] users. But where and where we see them [25:53] come from often is academia. And [25:56] academia has a different pace of life [25:58] than than startups. And so finding [26:00] people who can fit both roles of I have [26:02] deep experience in using this [26:04] technology, but I also have the just [26:08] drive to move quickly and to be okay [26:11] with making mistakes and fixing them on [26:13] the way. Those are the types of [26:15] qualities we look for in in the folks we [26:16] bring in. [26:17] >> Any advice and guidance on how do you [26:20] find those people who have that unique [26:22] balance? [26:23] >> I Yeah, I think it's just talking to a [26:25] lot of people. And so we're we're a [26:27] hybrid workforce. We work three days [26:29] together in office. Our interviews are [26:31] in person and so just working through [26:34] down the list and general, you know, [26:36] regular hiring pipeline to to figure out [26:38] who fits the molds there. [26:40] >> Yeah. So, you've both had quite a bit of [26:42] traction most recently in in your [26:44] startup. What would you have done [26:46] differently? What's one thing you would [26:47] have done differently? [26:49] Um, you know, in the in the vein of [26:53] common advice they give folks in Y [26:54] Common era, I would have just started [26:56] earlier and I probably dropped out [26:57] earlier. Um, I think, you know, time is [27:00] sort of the most valuable thing in in [27:02] this game. And so, in light of that, I [27:04] think just doing everything earlier and [27:06] shipping earlier and like like I said, [27:07] getting the product out in front of [27:08] customers earlier um is super important. [27:10] So, you know, I would say if there's [27:12] something that's sort of like rattling [27:13] around in your mind, you're not sure [27:14] whether it's whether it's real or not, I [27:16] think like the best way like Andy said [27:18] earlier is just get in front of the [27:19] people who would would use it, right? [27:20] And so, I think that's what I'd have [27:21] done differently. Just uh started [27:23] everything like quicker. [27:24] >> Okay. And is is Y Cominator accelerators [27:28] still are they still as helpful now as [27:30] they were in the past? [27:31] >> Oh, yeah. Absolutely. Um it's YC was [27:34] like some of the best experiences that [27:36] I've had. you know, we had like amazing [27:38] folks become our advisers through that [27:39] like Sam Alman, Brian Chesky from [27:40] Airbnb, um, co-founder of Anthropics. [27:43] Um, yeah, that just I mean for us like [27:47] we believe the company is an exponential [27:48] growth curve and obviously we wouldn't [27:50] be doing it if we didn't believe that [27:52] and for us YC was basically a way to [27:54] shorten the flatten part of that [27:55] exponential curve. So it's basically [27:56] what do you value? I mean we finished [27:58] our raise uh from Google in two days [28:00] which is like pretty atypical for a seed [28:02] raise. I think it usually takes I mean [28:03] Andy can comment on how long it [28:05] typically takes but [28:06] >> Yeah. Yeah. I mean the last two years if [28:08] you get it done like sub eight months [28:11] it's pretty good. [28:12] >> So um yeah it depends on how much how [28:13] much you value your time and different [28:15] aspects like that. [28:16] >> Yeah. And for us uh you know I think [28:18] time is always the the biggest factor [28:20] early on. And so for us I think we we [28:23] shipped early. I don't I don't think we [28:26] could have responsibly shipped earlier [28:27] than we did, but um after MVP, we went [28:32] we went deep on a couple of product [28:35] features where potentially we should [28:37] have gone more wide with them and uh get [28:40] a little bit more sense there. Um it [28:42] it's tough as a platform. You're [28:44] expected to do certain things really [28:46] well. And so figuring out the balance of [28:48] what's good enough for a particular [28:50] feature before you can move on from it [28:52] is is a hard hard prioritization [28:54] process, especially when you have [28:56] relatively few users early on to to make [28:58] that determination. [29:00] >> But yeah, maybe pivoting on some small [29:02] things quicker would have been the way [29:04] to go. [29:04] >> And I'm curious, Andy, specifically [29:06] thinking about GoFundMe. Yeah. [29:08] >> What would you have done differently now [29:09] with AI versus how you built it in the [29:12] past? [29:12] >> Yeah. you know, not dissimilar from [29:14] where the current team at GoFundMe is [29:17] going now. Um, one of the things that, [29:20] uh, maybe isn't immediately obvious in [29:23] using the GoFundMe platform is people [29:26] often talk about public speaking and [29:28] dying as the biggest fears in life. And [29:30] I would actually call on both [29:31] of those things. I think asking for help [29:33] is people's most number one fear and the [29:35] hardest thing for them to do in public. [29:37] And so one of the things that AI is [29:41] helpful for is allowing people to tell [29:43] authentic stories in a way where they [29:46] can have a writing partner and someone [29:49] who uh or someone an AI that is helping [29:52] them write these things because if [29:54] you're in a time even if it's not [29:56] yourself that you're raising money for [29:58] but you're raising money for your friend [29:59] it's still hard to ask on behalf of [30:01] somebody else and people are not in [30:04] general natural storytellers but you do [30:06] need to tell tell a story to make it [30:08] relevant to the audience that you're [30:10] addressing for this appeal. And so [30:13] having a coowwriter helping you tell [30:16] that story, helping you understand that [30:19] it's okay to ask for help and it's okay [30:21] to put certain details into the story. [30:23] That's really the number one hurdle that [30:25] we have. And so I would I would have [30:28] loved to have that ability. And early on [30:31] we talked about this is way before uh [30:33] generative AI. We would have loved to [30:35] have coaches that could sit with every [30:37] single customer and help them tell their [30:39] story. It's just prohibitively hard to [30:42] do in a humanbased kind of uh operation [30:45] when you have that agentic AI that can [30:47] sit with every single customer. That's [30:49] an incredible unlock. [30:50] >> Yeah, that's a that's a great example. [30:52] Is there another unconventional way that [30:55] you're using AI that perhaps most [30:58] founders aren't that you found helpful? [31:02] >> Um, so very akin to how how we put our [31:06] product out into the world. Our our [31:07] product is there to get insights out of [31:10] images and videos that your product is [31:12] already creating. Your retail store is [31:14] already creating however many crazy [31:16] amounts of hours of of video in the [31:18] space. your users are sending you [31:21] millions of photos or millions of of [31:23] frames of video that we're all just [31:26] putting in a bucket. They might as well [31:28] be black pixels to us. And so I think in [31:30] internally we think about how do we mine [31:34] atypical sources of data that provide us [31:37] with more insight about how our [31:38] customers are thinking and feeling about [31:41] our product. And so using Aentic AI to [31:44] help roll up and summarize customer [31:47] support inquiries to get to um to get to [31:50] the heart of what the problem is there [31:53] is is one way that is um maybe not [31:56] obvious and and not certainly not built [31:59] into a lot of the platforms that we're [32:00] using for for customer service yet, but [32:03] I think is really important because that [32:06] would otherwise be a really hard thing [32:08] to do across a lot of support agents. um [32:11] or just the entire database is getting [32:14] that detailed sentiment analysis. [32:15] >> Yeah, I mean we just because of the work [32:17] we do, we have to do a lot of model [32:19] level interpretability stuff. And so one [32:21] thing that's pretty interesting that we [32:22] were recently um at the White House [32:24] talking about was our work on the [32:26] deepseek model that was released by [32:28] China, right? And so we actually were [32:30] approached by a fortune 30 financial [32:32] institution um to try to understand like [32:35] how can we actually get this model more [32:36] aligned with American values, right? [32:37] Because if you can't ask it for example [32:39] about uh Chinese historical events like [32:41] tan men square the model becomes like [32:43] pretty not not super useful for like [32:45] economic analysis historical analysis [32:47] things like that right and so what we [32:48] actually had to do was look at the parts [32:50] of the model that activated when it did [32:52] that right and so we found very [32:54] interesting things like we could [32:55] actually um you know make it more [32:57] censored and the model would basically [32:58] spew gibberish but the interesting thing [33:00] is we could basically remove the parts [33:01] of it that are censored and then after [33:03] you apply our platform it actually [33:04] started talking about um TMN square and [33:07] things like that, right? So you don't [33:08] actually have to do a full fine-tune the [33:10] model or prompt engineer and like have a [33:12] crazy prompt um that you know breaks [33:14] with the next version or something, [33:15] right? So there's actually more [33:17] sustainable and extensible ways to [33:19] interact with these models and customize [33:21] them to use cases that don't really have [33:23] to be in like the archaic ways that I [33:25] think you know we typically interact [33:26] with intelligence. [33:28] >> Interesting. So as you think about the [33:30] next 6 months, the next year, what's [33:33] something you're struggling with right [33:34] now? What's a tension where you're not [33:36] sure whether to go to the left path or [33:38] the right path? [33:40] >> I think for us right now it's like [33:41] putting out all the fires and just [33:42] trying to trying to maintain like our [33:44] head above water, right? Um when you're [33:47] growing like super fast, there's a lot [33:50] of things that can sort of fall by the [33:51] wayside. And so that's what sort of [33:54] influences our decision to hire folks [33:57] that are, you know, that really [33:58] shouldn't be at early stage startups. Um [34:00] like whether by their experience or like [34:01] they're a 19-year-old kid from like [34:03] Berkeley or whatever. um we have like a [34:04] wide range of experience and I think one [34:07] of the things that I'm really trying to [34:08] do in the in the near- term and long [34:09] term is create that environment that's [34:12] really almost like um you know the the [34:15] Athenian forum where it's like an open [34:17] intellectual uh exercise where basically [34:20] the best ideas win and I think that [34:23] translates very directly to the best [34:24] product and I think the best products in [34:26] the past um if you look at something [34:27] like Xerox Park um Bell Labs uh like [34:30] things that have had really incredible [34:32] technological innovations they've [34:33] they've had that open academic [34:35] environment and I think that's what we [34:36] want to bring to the enterprise. We want [34:38] to like raise that standard there. [34:39] >> Interesting. And is there anything you [34:40] do with you know the folks at Stanford [34:43] these research entities to [34:44] >> Yeah, we are very uh active and [34:46] connected with the research communities. [34:47] We do a lot of meetups. Our office um [34:50] we've had folks like France who's like [34:52] very well known. Um he created this [34:54] machine learning library caris. He's one [34:55] of our investors. We have people like [34:56] that come and give talks um you know in [34:59] our [35:00] >> in our office and that really creates [35:01] like a collegial environment. Um it [35:04] makes it one where you know we're very [35:06] in organization we're not very [35:08] hierarchical. Um we're very flat and I [35:10] think that allows people to like come to [35:11] me with their ideas. um there's like [35:13] interns who have pushed code um that's [35:15] really high quality that is now running [35:16] in like Fortune 500 servers right and I [35:19] think very few companies actually [35:20] embrace openness in that way and I think [35:22] that's really our superpower like moving [35:24] fast incorporating new ideas um we [35:26] basically make the delta from like [35:27] cutting edge research to what we [35:29] actually deploy on customers servers uh [35:31] very short and that's those are some of [35:33] the ways that we do that [35:35] >> yeah I think that um in the in the [35:38] timeline of of zero to one you have to [35:41] have an MVP out there, you have to start [35:43] your initial go to market. We're we're [35:45] pretty far along the way of of knowing [35:47] what our customers want now. And I think [35:50] for us the key before we, you know, hit [35:53] a our series A milestone, which is kind [35:55] of the next thing we're we're focused on [35:57] is just making sure that we're moving [36:00] fast enough and we're staying ahead of [36:03] um what uh you would just be able to do [36:06] without IPOP out there and um providing [36:09] that value and and it's going to require [36:11] a lot of effort from the team um and [36:14] pushing our our small team hard. And so [36:16] some of the things we think about there [36:18] are how do we how do we get good [36:20] productivity out of everybody on the [36:21] team, but also how do we keep the team [36:24] together when we're all wearing multiple [36:26] hats and you know one person might be [36:28] three departments. You don't want to [36:30] lose that person in the scramble for um [36:34] for that that first you know revenue [36:37] milestone. And so that's some of the [36:38] things we think about there. [36:39] >> And is there something you've done to [36:41] try to foster that that culture? Yeah, [36:43] we we started this year um uh just a a [36:47] true sprint um agile method and we have [36:51] peppered into our sprint rituals a lot [36:53] of team building and team reflection [36:55] time. And so having those uh bi-weekly [36:59] milestones to come together as a team [37:01] and discuss what's working and not [37:02] working has really helped us to identify [37:04] those human level problems. And [37:06] sometimes it's like, "Hey, you know, uh, [37:09] Bob needs a break." Or sometimes it's [37:12] like, "We're doing something really [37:14] wrong here. We're pushing the team too [37:15] hard. Let's step back, reset the plan, [37:18] and figure out how to move faster by [37:20] moving slightly slower." [37:21] >> And is that in-person environment as [37:24] important for you as well? Do you think [37:26] about, you know, 3 days versus 4 days [37:27] versus 5 days? [37:29] >> Um, I think for us it's important to be [37:31] together most of the time. And so 3 days [37:34] makes sense to us. the um we're at the [37:38] stage where if there's a bug, I can't [37:41] slack somebody to Zoom Zoom them later [37:43] and show them the bug. I need to pick up [37:45] my laptop and show, hey, hey, we got a [37:47] problem here. We need to fix this. Or [37:49] you want to be able to overhear a [37:52] discussion around your desk of part of a [37:55] product feature they're testing or [37:56] something like that and say, "Hey, like [37:58] we need to change course here. We have a [38:00] misunderstanding of what that product is [38:02] supposed to be." [38:02] >> Yeah. I mean, as for us, we're totally [38:04] in person. We think that's very [38:05] important to an early stage company and [38:07] just the again similar to what Andy [38:09] said, being able to bounce ideas off of [38:11] each other in the heat of the moment. I [38:13] think that leads to a lot of uh a lot of [38:16] great innovations and a lot of good [38:17] ideas and so yeah, everybody just comes [38:19] into the comes in the office in person. [38:21] >> Yeah. So we have a lot of founders in [38:24] the audience. I know. What's your [38:25] biggest piece of advice right now for [38:27] the founders in the audience in terms of [38:29] going from zero to one? [38:32] I think if you haven't started yet, um, [38:35] start and start by figuring out what is [38:38] the exact value that you're going to [38:40] provide to a very specific audience. Um, [38:43] and one of the ways you can start is not [38:46] just by building, but by also start [38:48] curating that audience. Audiences are a [38:50] defensible moat in this day and age. And [38:52] so having that audience, even if [38:55] ultimately the product vision changes, [38:57] it'll still be a relevant uh asset to [39:00] have. [39:01] Yeah, concurring with everything Andy [39:03] said and also adding on, you know, [39:04] really trying to find those 9010 [39:06] solutions, right? The solutions that are [39:08] 90% of the quality of something else but [39:09] take 10% of the effort because that's [39:11] really what allows you to like reap [39:12] leaprog other people's and incumbents in [39:14] the space, right? Finding those [39:15] solutions that um other people won't do [39:17] because they're too scrappy or they [39:19] think they won't work. Um bringing like [39:21] that naive that you have just as being a [39:23] startup often that's what makes [39:24] solutions actually, you know, in their [39:26] full form become superior to um all the [39:28] other competing ones. So [39:30] >> great. Well, I think we're we're almost [39:31] at time. Just a couple minutes. Anything [39:33] I didn't ask that I should have? [39:35] Anything we didn't talk about that you [39:37] think is important? [39:42] >> I think I think um you know just going [39:45] back to to the AI slop and like am I [39:50] differentiated from just a thin layer on [39:53] on chat GBT? I think that's a question [39:55] we all need to be asking ourselves in [39:57] the the meta view of like is this is [40:00] this functionality going to be [40:01] commoditized soon and that's a really [40:04] hard question to answer because the [40:05] space is moving really quickly and so [40:07] figuring out your own answer to why this [40:11] is defensible in a in a in a realm where [40:14] next year we could have models that are [40:16] doing something even more ingenious and [40:19] uh you know so why why why do should you [40:23] exist in that sort of world. [40:24] >> Yeah. And we see this a lot with the [40:26] customers that we work with, but I think [40:28] um interestingly it's also applicable [40:29] for the early stage companies. Are you [40:32] actually deploying AI because there's a [40:34] real pain point that only AI can solve [40:36] or are you deploying AI because it's [40:38] like a buzz word and it's the it's the [40:39] thing that's getting most funding right [40:40] now, right? I think the latter um is [40:43] very insidious in that it'll cause you [40:46] to make short-term decisions and build [40:48] the company in such a way that it [40:50] probably won't be stable um even like a [40:52] year on out. So, I think you need to [40:54] really ask yourself and question whether [40:56] this is something that really needs to [40:58] be AI um and if it's like AI native at [41:01] its core or if it's just something [41:03] that's superficial um and has AI slapped [41:05] on top of it. [41:06] >> Yeah. [41:06] >> Well, I think that's a great note to end [41:08] on. Thank you so much for joining us [41:10] everyone. Please join me in thanking [41:12] Sarah and Mandy.