Ex-Meta Worker Launches ChatBlackGPT After Algorithmic Bias Layoff

前Meta员工因算法偏见被裁后推出ChatBlackGPT 标准AI算法打着政治中立的旗号压制历史事实，使得定制调优模型对于保障用户安全和文化准确性而言必不可少。 Erin Reddick分享了她从大型科技巨头离职并创立ChatBlackGPT™的转型经历，展示了默认AI模型往往因存在缺陷的搜索与检索指令，从而压制文化背景或曲解特定文化提示词。通过诸如定义“黑人工作”或检索“日落镇”地方历史等对比案例，讨论强调了在整个AI生命周期中融入多元视角的必要性。Reddick还介绍了底特律的313 GPT等本地化AI模型，并阐述了低代码定制GPT工具如何赋能边缘化社群构建具有文化适应性的技术，以减轻系统性偏见。 Erin Reddick 在经历微软、AWS 和 Meta 的十年科技大厂生涯后于 2023 年创立 ChatBlackGPT™，旨在通过调整算法指令与检索规则而非仅靠基础数据来纠正文化偏差与历史失真。 (0:04) 对比测试显示默认 AI 模型常以规避政治争议为由忽略敏感历史（如内华达州明登市曾作为日落镇的历史），而文化自适应模型则将其视为保障用户出行安全的重要背景事实。 (4:50) 依赖员工资源小组（ERG）在上线前夕进行把关审查会给少数群体员工带来巨大的职业压力，多元化视角必须贯穿从需求定义、数据收集到后期对齐的整个软件生命周期。 (12:23) 底特律的“313 GPT”项目通过定性研究与口述历史保存非裔长者的社区记忆，结合动态情感感知与推荐权威书单，有效防止算法回音室效应并降低幻觉依赖。 (14:55)

ChatBlackGPT™ founder Erin Reddick argues that cultural bias in artificial intelligence stems from retrieval instructions and selective historical suppression rather than an absence of training data. She contends that equitable technology requires integrating community-grounded perspectives throughout the entire engineering lifecycle rather than delegating risk reviews to late-stage employee checks. Algorithmic Instructions and Cultural Context Gaps (2:36) Reddick asserts that existing large language models do not fail due to a lack of Black user data online, but because default algorithmic instructions fail to search and prioritize culturally informed contexts. In comparative tests, standard ChatGPT defined the colloquial phrase 'a black job' by citing illicit black-market activities and drug trafficking, whereas ChatBlackGPT™ contextualized the query within historic segregated labor markets. Similarly, an inquiry regarding Minden, Nevada showed standard ChatGPT omitting the town's history as a sundown town and a 2020 law enforcement controversy regarding Black Lives Matter literature. When questioned, the default system cited tuning against 'politically inflammatory topics'—an omission Reddick characterizes as suppressing critical physical safety context for travelers under the pretext of neutrality. Shifting Diversity from Late Reviews to Lifecycle Design (12:23) Reflecting on her career at Microsoft, AWS, and Meta, Reddick criticizes tech corporations for limiting minority staff participation to late-stage approvals. Enterprise workflows frequently involve project leaders presenting finalized products or marketing initiatives to underrepresented employees immediately prior to deployment, asking if the release will inflict community harm. Reddick notes that this dynamic places unfair professional risk on individual contributors. Meaningful risk mitigation instead requires diverse practitioners to be embedded across every phase of software development—from initial data collection and early design to post-training and Reinforcement Learning from Human Feedback (RLHF)—which shapes tone, nuance, and contextual safety. Preserving Oral Histories Through Hyper-Local Models (14:55) To demonstrate alternatives to generalized foundation models, Reddick developed 313 GPT in Detroit, employing qualitative neighborhood research and oral histories from aging community figures, such as surviving members of the Last Poets and Black Panther Party. Using sentiment analysis, local data is embedded directly into retrieval logic to keep pace with evolving community dynamics, avoiding outdated cultural stereotypes. Reddick highlights that custom tools should point users back to external learning rather than creating closed ecosystems; her platform delivers curated reading lists of academic works alongside AI outputs to encourage verified reading and lower compute expenditures wasted on repetitive prompt corrections. Democratization, Compute Constraints, and Global Governance (28:30) Reddick disputes the notion that micro-communities lack the technical capability to build bespoke tools, observing that low-code prompt architectures allow non-technical builders to configure custom algorithms on existing foundational data for twenty dollars a month. However, audience feedback illustrated deep structural imbalances outside the United States: African developers face severe regional data center and compute deficits, frequently requiring reliance on Small Language Models (SLMs) rather than large commercial models. Addressing broader governance, moderator Nathan Walker pointed to upcoming United Nations dialogues in Cape Town designed to incorporate non-Western frameworks, such as Ubuntu ethics, into international human rights standards for AI. While low-code platforms allow marginalized communities to build representative proof-of-concept models, questions remain regarding how hyper-local architectures can safeguard intimate community data and prevent polarized echo chambers amid widespread international disparities in compute infrastructure.

ChatBlackGPT™创始人Erin Reddick指出，主流人工智能系统对非裔等少数群体的偏见与误读，其根本症结不在于训练数据的匮乏，而在于检索与优先级算法缺乏文化上下文规则，且大型科技公司在软件生命周期初期缺乏多元视角。她主张通过定制文化指令集、保存社区口述历史并借助开源工具，使边缘化群体能够构建兼顾文化共鸣与安全边界的包容性AI系统。 算法指令而非数据缺失：ChatBlackGPT的诞生 (2:36) Reddick结合其在Microsoft、AWS和Meta长达十年的用户体验（UX）研究经历指出，当主流模型在文化敏感节点（如历史人物错误生成）频发失误时，大型科技企业往往缺乏主动纠偏动力。她提出一项核心假设：在通用互联网规模的训练池中，非裔用户的数字足迹早已充分积累，基础训练数据并不欠缺，问题在于指导模型搜索、检索和分配资源优先级的底层算法指令。基于这一判断，Reddick在OpenAI开放自定义GPT功能后，制定了融入文化上下文的算法指令规则。该概念验证工具迅速吸引了超过2.5万名活跃用户，并帮助其募集资本以开发独立的专属平台，证明为厌倦反复提示标准模型的用户提供定制指令具有高度可行性。 规避敏感议题的局限：提示词对比与现实安全考量 (4:50) 标准大模型与文化语境AI的差距在具体测试中表现显著。以定义“黑人工作”（black job）为例，通用ChatGPT因缺失现代文化语境，将其与“黑市”（black market）混淆，给出了毒品走私和非法用工等负面回答；而ChatBlackGPT™则能精准锚定历史背景，主动询问是否指种族隔离时期的“黑人岗位”（Negro jobs）并详述当时劳动力市场史实。在针对内华达州明登镇（Minden）的联合实测中，标准模型仅输出人口统计与德国传统等维基百科式信息，并将该镇直至2023年仍鸣放“日落镇”（sundown town）警报器以及2020年警长威胁拒绝向支持BLM的图书馆提供警力等事件，视为“政治敏感话题”予以刻意过滤。Reddick对此强调，此类历史在少数族裔群体出行中关乎切实的人身安全，通用模型以“中立”为由压制敏感史实，实际上剥夺了特定用户获取关键安全背景的权利。 研发周期早期介入与摆脱“代币式”末端审查 (12:23) Reddick回顾大厂工作经历时揭示了科技行业在包容性治理上的结构性缺陷：少数族裔员工极少能参与数据采集与预训练等早期系统架构设计，往往只在产品或营销方案临近发布的最终审核环节，被副总裁或总监拉入会议，被动面对“该产品是否会伤害你的社群”的二元是非判断。Reddick认为，这种末端代币化（tokenized）审查不仅迫使基层员工冒着职业风险当面质疑管理层，而且无法从根本上消除算法偏差。真正的风险缓解要求在数据预训练、后训练以及人类反馈强化学习（RLHF）的全生命周期中及早且持续地吸纳多元视角，通过精细化调整人机反馈，优化模型在表述语气和复杂生存经验上的细微辨析力。 超本地化口述历史与动态社区知识建模 (14:55) 为避免模型退化为静态刻板印象，Reddick在底特律启动了名为“313 GPT”的超本地化项目，采用质性研究与口述历史记录当地长者及历史组织成员（如最后的诗人、黑豹党幸存成员）的第一手叙事。该系统通过情感分析将街区知识编入检索逻辑，探索由社区共同构建的AI范式。Reddick指出，“十年前的底特律并非今天的底特律”，AI必须动态保持对社区情绪脉搏的感知，否则便会加剧数字隔阂。此外，该平台还在AI答复后提供由权威学者撰写的推荐书单，主动引导用户离开封闭的对话界面，通过经核实的外部专业文本文献进行深度学习，从而兼顾心理安全感、避免因反复纠偏造成的算力浪费，并降低模型幻觉带来的潜在风险。 创业韧性、微型社群赋能与全球治理框架 (28:30) 在经历了包括遭遇裁员、失去亲人和宠物意外在内的人生低谷后，Reddick借助Hugging Face等开源平台重构技术理念，验证了非技术人员自建工具的可行性。针对资源匮乏或小众群体的构建门槛，她驳斥了微型社群无法自建工具的论调，指出用户每月仅需支付20美元即可利用基础大模型框架配置专属行为规则与最小可行产品（MVP），其初始产品正是依此迅速扩展至5万名用户。与此同时，全球开发者在算力与数据中心匮乏的现实下，正在转向小语言模型（SLM）以适配本土需求。在学术与政策层面，Reddick已在哈佛大学（3次）、乔治城大学和华盛顿大学等高校开展工作坊，推动学生进行红队测试；主持人Walker则透露将赴开普敦参与联合国人工智能与人权对话，旨在将非西方的“乌班图”（Ubuntu）伦理框架引入全球治理。 论坛总结强调，构建包容性AI并非边缘修补，而是关乎技术公平与公信力的基础工程。然而，现场讨论也留下了关于深层技术挑战的审视：当针对特定社群的定制AI深度汲取私密口述数据时，如何在尊重本土文化的同时有效防范数据隐私泄露、算法信息茧房以及社交媒体时代曾出现的群体极化与确认偏误，仍是去中心化AI建设亟待解决的课题。

## Background 该视频由频道“Ai4”于 2025 年 8 月 28 日发布在 YouTube 平台上。内容由罗格斯大学（Rutgers University）AI 伦理实验室主任 Nathan Walker 与 ChatBlackGPT™ 创始人兼 CEO Erin Reddick 共同呈现，围绕促进人工智能的多样性展开，探讨如何构建包容性的技术与团队以实现更公平、更具伦理创新的 AI 成果。 ## TL;DR 对谈探讨了主流大语言模型在文化语境与少数族裔真实经历上的认知盲区，指出算法偏见往往源于微调指令而非底层数据本身。Erin Reddick 结合自身从科技巨头离职到创立 ChatBlackGPT 及本土化 313 GPT 的经历，阐述了在 AI 全生命周期中引入多元视角的重要性，以及如何通过敏捷创业与开源生态赋能边缘化社群。同时，双方结合内华达州日落镇实测与非洲等全球本土化实践，强调了将文化感知、口述历史以及乌班图（Ubuntu）等非西方哲学融入 AI 伦理治理的现实紧迫性。 ## Atmosphere & style 整场对话兼具思想深度与人文温度，从个人职场阵痛与韧性重建自然延展至宏大的全球技术伦理，洋溢着务实而坚定的思辨力量。主持人循序渐进地抛出尖锐议题，嘉宾则以坦诚、敏锐且极具感染力的叙事拆解科技寡头话语下的表面包容。现场互动真诚而热烈，既有面对系统性不公的清醒批判，又传递出普通人利用开源工具重构技术话语权的振奋与希望。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=4s) Erin Reddick 在亲历科技大厂严重缺乏少数族裔代表性的环境后选择独立创业，利用遣散费启动 ChatBlackGPT，旨在填补文化感知与多元代表性 AI 的空白。 - [02:36](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=156s) AI 出现偏见或盲区往往不是因为底层训练数据匮乏，而在于检索与调用模型时所采用的算法指令（Instructions）和微调立场缺乏文化语境。 - [06:32](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=392s) 在内华达州明登镇的实测对比中，通用模型因规避敏感争议而过滤了日落镇等历史，但这些信息对少数族裔出行而言实际上是关乎切身人身安全的重要参考。 - [10:50](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=650s) 调优团队的认知边界直接决定了 AI 系统的盲区；若大模型将少数族裔的安全信息当作“政治争议”一味过滤，将严重削弱 AI 作为普惠基础设施的价值。 - [12:15](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=735s) 多元化视角必须深度融入 AI 构思、数据收集、RLHF 微调等全生命周期，而非在部署前夕才把少数族裔员工当作承担职场压力的被动审查“橡皮图章”。 - [14:56](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=896s) 底特律本土项目“313 GPT”通过采集长者口述历史、社区定性调研与情绪分析，探索了将未数字化的本土经验与邻里共创机制融入 AI 的新路径。 - [17:49](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1069s) 防止文化对齐 AI 陷入信息茧房与确认偏误的核心在于保持数据和反馈机制的“动态性”（Dynamic），时刻吸收社群的发展与变化，杜绝死板硬编码。 - [21:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1298s) 文化响应式 AI 让边缘群体无需繁琐编写 Prompt 进行文化解释与纠偏，不仅大幅降低算力与沟通损耗，还构筑了心理安全区并通过推荐真实文献打破单一答案迷信。 - [28:06](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1686s) 全球多元群体（如非洲本土生态）在面临文化认知鸿沟与算力基础设施瓶颈时，正通过微调小语言模型（SLMs）和深入基层的社区对话探索破局方案。 - [31:32](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1892s) 小微社群创业者无需拥有昂贵数据中心，借助每月低成本的平台 API 与自定义指令即可快速构建并验证 MVP，反哺特定社群需求。 - [37:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2258s) 将非洲“乌班图”（Ubuntu）等非西方本土伦理哲学融入《联合国人工智能与人权公约》等国际治理框架，是构建包容性全球 AI 伦理体系的关键探索。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=4s) 从科技巨头到独立创业：ChatBlackGPT 的诞生背景 在本场关于“促进人工智能多样性”（Fostering Diversity in AI）的对谈中，主持人罗格斯大学 AI 伦理实验室主任 Nathan Walker 开场致意未能到场的两位同仁，随后将焦点转向 ChatBlackGPT 的创始人兼首席执行官 Erin Reddick，探讨其职业轨迹与创业初衷。Erin Reddick 曾在微软（Microsoft）、亚马逊网络服务（AWS）和 Meta 等全球顶尖科技公司工作多年。回顾在西雅图及华盛顿州贝尔维尤（Bellevue）约十年的科技从业经历，她坦言在五年的线下办公期间，在园区见到的有色人种员工大概总共也就十个人左右，这种代表性严重匮乏的环境让她深切感受到科技行业的单一性。 2023年遭遇裁员后，Erin Reddick 没有急于重返科技大厂，而是选择将遣散费作为启动资金，创立自己的独立初创项目。她注意到当时围绕人工智能存在两种截然对立的叙事：一种将其誉为“继互联网以来最伟大的发明”，另一种则指责 AI 正在抹杀多元文化并扭曲黑人社区的历史事实。为了“按自己的方式掌控作为科技女性的话语权”，她决定通过独立创业去填补这一空白，探索具有文化感知力（culturally informed）的人工智能。 - **核心经历与节点**：在西雅图大厂从业约10年，涉及 Microsoft、AWS 与 Meta；2023年被裁后用遣散费启动 ChatBlackGPT。 - **行业痛点**：大型科技公司内部少数族裔代表性极度匮乏，且普通员工对产品的根本性变革缺乏足够的决策影响力。 --- ### [02:36](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=156s) 算法与视角的破局：构建文化感知的 AI 原型 Erin Reddick 指出，她在科技公司的背景主要是用户体验研究（UX Research），曾长期与高管团队紧密合作，研究多元化客群需求。当全行业都在目睹诸如谷歌历史人物图像生成偏差等 AI 大事件时，行业往往缺乏前瞻性的解决方案。当 OpenAI 推出自定义 GPT 功能（Custom GPTs）后，Erin 提出了一个核心假设：AI 出现偏见或盲区，问题本质上往往不在于底层训练数据本身——毕竟黑人社区在互联网上产生的数据量已足够庞大，真正的问题出在检索与调用的**算法指令（the set of instructions）**上。 ```mermaid flowchart TD accTitle: 数据利用与算法指令分流示意图 subgraph DataLayer["海量互联网公开数据 (Common Data)"] D["包含全网多元历史与当代文本数据"] end subgraph Pathways["不同算法指令与微调逻辑"] A["标准/主流默认算法指令"] --> R1["单一叙事 / 过滤敏感事实 / 缺乏文化语境"] B["ChatBlackGPT 公平算法指令"] --> R2["多元视角 / 历史真实语境 / 实用安全洞察"] end D --> A D --> B ``` 基于这一理念，她构建了 ChatBlackGPT 的概念验证版本（PoC）。该原型迅速吸引了超过25,000名活跃用户，其市场热度与实用价值帮助她成功筹集资金并开发了独立版本。对于产品的愿景，Erin 强调 ChatBlackGPT 并非要强行灌输单一结论，而是为了服务所有不满足于“单一主导叙事”的用户，让被边缘化的群体不再需要经过反复繁琐的提示工程（prompting）才能获得具有文化关联性的回答。 - **经典案例对比（"Black Job"）**： - **通用 ChatGPT 回答**：将 "black job" 机械关联为“毒品交易、非法活动、未注册黑市生意”，混淆了当代黑人生活语境与“黑市（black market）”。 - **ChatBlackGPT 回答**：主动提供历史与语境关联——“你是否指的是种族隔离时期历史上所称的‘黑人工作’（negro jobs）及相关背景？”，并详细列出当时语境下所指的具体职业类别。 --- ### [06:32](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=392s) 内华达小镇实测：日落镇历史与现实安全考量 在对谈前的一小时，Nathan Walker 与 Erin Reddick 进行了一场现场对比测试，向两个模型提出了同一个简单问题：“告诉我关于内华达州明登（Minden, Nevada）需要了解的一切”。明登是 Nathan Walker 长大的地方。标准版 ChatGPT 给出的答案极其类似维基百科词条，罗列了当地的人口统计、建镇背景与德国裔传承，却完全忽略了极其关键的历史与现实争议；而 ChatBlackGPT 则给出了更加全面且符合历史事实的回答，明确指出明登在历史上是一座著名的**“日落镇”（Sundown Town）**。 Erin Reddick 解释，明登直到近期仍保留着日落镇的象征机制（例如每晚5点鸣响的警钟，历史上用于警告少数族裔在日落前离开）。ChatBlackGPT 的提示不仅澄清了历史，还给出了实用的出行建议：“并非叫你不要去，而是要知情前往（go informed）”。对于黑人旅行者而言，缺乏对这类遗留观念与现实冲突的了解，可能会直接面临现实的人身安全风险，而这绝非一份平淡的“德裔传统”简介所能涵盖的。 - **明登镇（Minden, NV）近年现实争议与历史修正**： - **2020年警长事件**：当地警长公开在社交媒体表示，若当地图书馆敢宣传“黑人的命也是命”（Black Lives Matter），警方将拒绝向图书馆提供治安支持，并扬言逮捕支持者。 - **鸣钟叙事替换**：当地曾试图将每晚5点的鸣钟仪式粉饰为“向消防员、警察等公共服务人员致敬”，试图掩盖其针对原住民及黑人的日落驱逐历史。 - **立法废除**：历经数十年抗议（包括 Nathan 姑母在90年代发起的请愿），最终在内华达州总检察长调查与州议会立法干预下，才在近年正式废除相关日落规约。 --- ### [10:50](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=650s) 安全审查与调优盲区：多样性为何至关重要 当 Nathan 追问标准 ChatGPT 为什么在初次提问时隐瞒2020年事件与日落镇历史时，模型直言其调优机制倾向于**规避政治敏感或具有煽动性的争议话题**。Erin Reddick 指出，科技大厂将此类历史事实与现实安全信息归类为“政治争议”从而加以过滤，但对于少数族裔而言，这根本不是政治辩论，而是关乎**切身安全（safety concern）**的基本信息。如果大模型被推行为全社会的基础设施，少数族裔却无法安全、可靠地使用它，这种工具的普惠性便荡然无存。 Erin 澄清自己绝非“反 AI 主义者”，也不认为 AI 算法天生带有主观恶意。问题在于**“谁在负责调优”（who is tuning）**——调优人员的认知边界决定了系统的盲区；没有少数族裔的切身体验，开发团队就永远无法感知到这些隐蔽的安全隐患与文化偏见。这也正是技术开发中必须坚持“多样性（Diversity）”的根本原因。 ### [12:15](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=735s) AI 全生命周期中的多元化与在企业中的真实阻力 多元化在技术领域不仅是为了形式或乐趣，更是直接关乎安全性的核心议题。Nathan Walker 指出，多元化不应仅停留在数据集层面，而必须贯穿 AI 整个生命周期——从构思、设计、开发直到部署。他向 Erin Reddick 提问：在技术正式上线前，让多元化团队深度参与各个环节究竟有何重要性，能为企业带来什么实质益处？ Erin 回忆了自己在科技大厂工作时的亲身经历。当时她常被拉进专门的评审小组或大型宣发活动的咨询会议中。然而，流程往往是项目团队已经完成了所有设计并准备上线时，才在会议最后一刻把方案推到她面前，直接甩出一个问题：“这会对你的社群造成伤害吗？请回答是还是否。”Erin 直言，作为一个普通员工（IC, Individual Contributor），当面面对副总裁（VP）和总监们直说“我觉得这有点种族歧视或性别歧视，我认为这不行”，需要承受巨大的职场压力。这种在末端充当“橡皮图章”式的征询不仅不负责任，甚至在职业发展上都是一种冒险。她强调，多元化视角的融入绝不能仅仅是上线前的最后一道打勾审核，而必须在设计初期及全流程中频繁、深入地嵌入，只有这样才能在源头有效降低产品对特定社群造成伤害的风险。 ```mermaid flowchart TD accTitle: AI生命周期中的多元化介入模式对比 subgraph Bad["传统被动模式（高风险与高压力）"] B1["产品设计与开发"] --> B2["准备上线部署"] B2 --> B3["末期单次评审<br/>（要求员工回答'是否造成伤害'）"] end subgraph Good["全程共创模式（深度降低风险）"] G1["早期需求与概念设计<br/>（融入多元化社群声音）"] --> G2["数据收集与预训练"] G2 --> G3["RLHF 对齐与精细微调<br/>（优化回答立场与表达方式）"] G3 --> G4["安全部署与持续动态反馈"] end ``` - 传统模式将风险审核推迟至部署前夕，将巨大的责任和压力单方面压在少数族裔基层员工身上。 - 深度参与机制需要覆盖数据收集、预训练以及基于人类反馈的强化学习（RLHF）全过程，不仅决定模型“说什么”，更规制模型“怎么说”。 --- ### [14:56](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=896s) 从 ChatBlackGPT 到 313 GPT：口述历史与本土化 AI 共创 谈到如何将特定社群的生活经验深度融入 AI，Erin 分享了她的最新实践与生活转折。她最近搬离了华盛顿特区前往底特律定居，一方面是为了靠近家人，另一方面也是因为特区当时局势紧张——国民警卫队已部署上街，而她原本的住处距离白宫仅 8 分钟路程。搬到底特律后，她迅速融入了当地蓬勃发展的科技生态，并着手将 ChatBlackGPT 的方法论微缩复制为本土化项目——“313 GPT”（313 为底特律的电话区号）。 313 GPT 的核心在于与社区邻里共同创造生成式 AI。Erin 观察到，像纽约的“最后的诗人”（The Last Poets）或老一代黑豹党（Black Panthers）成员正在逐渐老去离世，但他们的第一手历史与生活经历尚未完成数字化留存。313 GPT 深入社区，向居民提出极具生活温度的问题：在这里生活感觉如何？你住了多久？谁是社区中德高望重的长者或最具影响力的人？学校系统为你的社区带来了什么改变？通过收集这些口述历史与定性研究数据，并结合精细的情绪分析（Sentiment Analysis），系统能够将真实的社区脉搏映射到 AI 的交互与回答中。 Nathan 赞许这种结合“口述历史”与“社区学术”的方法，认为 AI 宛如一面折射人类全貌的镜子，既映照出潜能，也暴露出歧视，但大众过去缺乏平等的工具去修饰与纠正这面镜子中的自身形象。面对 Nathan 关于未来是否会出现“穆斯林版 GPT”、“犹太版 GPT”或“摩门教版 GPT”等特定文化 AI 的提问，Erin 表示，她之所以创立 ChatBlackGPT，是因为她能以专家级标准定义并验证自己的亲身体验；她希望将这套开源蓝图提供给所有社群，让每个人都有权利使用开源技术构建能够代表自身文化的 AI 工具。 - **313 GPT 社区定性调研核心维度**： - 居住年限与本土社区变迁认知 - 社区长者（Elders）与领袖人物的口述历史 - 当地教育与学校系统的实际影响评估 - 居民情感与生活经验的情绪分析嵌入 --- ### [17:49](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1069s) 零营销增长、规避算法信息茧房与动态知识更新 ChatBlackGPT 在没有任何付费广告、仅靠 Erin 个人在 LinkedIn 上发帖与口碑传播的情况下，便迅速获得了超过 25,000 名真实用户。Erin 指出，这种完全依靠有机传播（Organic growth）的爆发式增长，充分证明了文化专属型 AI 对于社群的巨大价值——“人们不会去使用那些无法在其中看到自己身影的工具。” 针对深度定制社区体验可能带来的隐私风险以及如同社交媒体算法般加剧政治/文化“信息茧房”（Echo Chambers）的担忧，Nathan 询问该系统如何避免陷入狭隘的确认偏误，并利用多视角切换来拓宽同理心圈子。Erin 回应称，防止固化和信息茧房的关键在于保持模型的“动态性”（Dynamic），坚决杜绝死板或硬编码（Hard-coded）的数据设定： > “构建这类模型时，保持动态性是极其重要的……底特律十年前的样貌绝不是今天的样貌。人们的感受会发生改变，你必须时刻保持对这种脉搏的感知，确保那些动态变化的声音能够不断被整合进来。” 通过建立持续吸收社会与社群新信息的反馈机制，模型不仅能避免自我封闭的循环回响，还能在涉及历史反思与和解的议题时（如内华达州历史案例），成为社区抚平创伤、避免重蹈歧视覆辙的载体。 --- ### [21:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1298s) 在技术迷局中构建心理安全区与引导深度阅读 谈及对未来解决 AI 偏见与歧视问题的展望，Erin 表达了一种务实且清醒的乐观态度。她深受 Timnit Gebru、Joy Buolamwini 博士、Ruha Benjamin 博士以及 Safiya Noble 博士等杰出女性学者的启发，深刻理解 AI 背后蕴含的巨大机遇与致命风险。Erin 将自己的使命定义为“在这片混乱的局势中尽最大努力（make the best of this mess）”，全力减缓数字鸿沟的蔓延。 Erin 阐明了文化对齐 AI 对少数族裔的现实意义：当用户使用生成式 AI 时，如果不必费尽心机去写提示词（Prompt）解释自己的文化背景与生活语境，不必反复纠正“不，我指的是这个意思”、“不，我来自哪里”、“不，我的真实感受是这样”，这不仅极大地节省了算力成本与时间损耗，更是在一个原本缺乏文化同理心的通用技术工具中，为特定社群构筑了一个宝贵的“心理安全区”（pocket of psychological safety）。 ```mermaid flowchart LR accTitle: 传统AI与ChatBlackGPT在交互效率与心理体验上的对比 subgraph Traditional["通用大模型交互痛点"] T1["用户提出诉求"] --> T2["缺乏文化语境理解"] T2 --> T3["反复纠偏 Prompt<br/>（浪费时间与算力）"] T3 --> T4["产生疏离感与挫败感"] end subgraph ChatBlackGPT["文化响应式交互优势"] C1["用户提出诉求"] --> C2["内嵌文化生活经验"] C2 --> C3["精准理解语境一次到位"] C3 --> C4["建立心理安全区与推荐延伸阅读"] end ``` 针对 Nathan 提到的产品特色——在给出首轮回答后主动为用户推荐一系列延伸阅读书单，Erin 揭示了背后的深层产品逻辑：主流商业 AI 往往希望将用户死死锁在其封闭的系统生态内，试图让 AI 成为包办一切问题的唯一终极答案；而 ChatBlackGPT 采用这种兼具深度与文学性的设计，旨在打破单一模型的全知神话，鼓励用户跳出对话框，走向更广阔的真实知识与历史文本。 ### [24:05](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1445s) ChatBlackGPT 的设计初衷与引导真实世界学习 Erin Reddick 阐述了她在设计 ChatBlackGPT 时的核心考量。她指出，现有 AI 模型不可避免地会出现事实性错误或幻觉，因此盲目依赖模型生成的内容存在风险。为此，她构思了一个机制：让系统敏锐捕捉用户的真实意图，并将用户直接引导至真实世界中由该领域专家撰写的权威资源，帮助用户进行独立、深入的学习。这种将用户重新带回现实世界、提供可靠文献来源的做法，恰恰弥补了现有主流 AI 普遍欠缺的关键能力。 ```mermaid flowchart LR accTitle: ChatBlackGPT Resource Routing subgraph Input["用户交互"] A["捕获用户意图"] end subgraph Process["系统逻辑"] B["过滤潜在幻觉"] --> C["匹配领域专家资源"] end subgraph Output["现实世界连接"] D["指引独立深入学习"] end A --> B C --> D ``` - 核心机制：通过精准理解用户探索的主题，分发由真实专家撰写的学术与实践资料。 - 设计哲学：承认 AI 的局限与幻觉风险，主动构建连接数字系统与现实权威知识的桥梁。 --- ### [24:45](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1485s) 面对系统性困境的坚韧与从零重构自我 主持人 Nathan Walker 对 Erin 展现出的从容自信与好奇心表示由衷赞叹。他提到，面对颠覆性技术带来的种种不公与系统性阻力，许多投身此领域的人常常感到被掏空或陷入绝望，而 Erin 却展现出强大的自我价值感与复原力。针对 Nathan 关于领导者自我价值与韧性来源的提问，Erin 敞开心扉，分享了自己跌入人生谷底后的重建历程。 Erin 回忆，在遭遇科技公司裁员的同一个星期里，她的爱犬遭遇车祸，祖母也离世了。此前，她的个人身份与大厂科技生活深度绑定——那种提供 24 小时预涂牙膏、洗漱用品和免费一日三餐自助的园区环境，让人彻底融入其中，难以抽离。失去工作后，她经历了巨大的心理冲击，不得不彻底从零开始审视自己的核心追求：继续留在科技行业究竟是为了虚荣的标签，还是源于对技术本身的真正热爱。在经历心理咨询支持并偶然接触到 Hugging Face 和开源生态后，她意识到自己完全有能力利用现有工具，构建出能为他人发声、赋能社群的产品。这种从情感和职业的彻底归零中重生并重获生活重心的过程，铸就了她当下的内在坚韧。 - 沉浸式科技职场的心理落差：企业提供的全包式生活福利容易让从业者将个人价值与雇主绑定，突然被裁会带来巨大的身份认同崩塌。 - 重建支点：通过 Hugging Face 等开源社区发现无需依赖大厂亦可亲手创造有价值的工具，并在心理咨询辅助下完成了心理重塑。 --- ### [28:06](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1686s) 全球视角下的 AI 代表性困境与非洲实践 论坛进入现场问答环节。一位来自尼日利亚、穿梭于美非两地的 AI 科技从业者首先发言，分享了非洲社群在使用主流 AI 工具时面临的代表性挑战。他表示自己是 ChatGPT Plus 的重度日常用户，但深深感受到模型并不理解非洲本土的文化视角与认知语境，这种隔阂比非裔美国人群体所经历的还要更深一层。 该提问者指出，即便在图像生成领域，主流模型也是直到近期才略微改善了对非洲本土形象与场景的刻画。背后的核心瓶颈在于：负责微调大模型的团队往往对未曾接触过的地域文化缺乏理解与共情；与此同时，非洲本土开发者虽然渴望开发自己的模型，却受制于算力和数据中心的匮乏。为此，本土生态正在积极探索采用小语言模型（SLM）等替代方案来应对资源限制，并呼吁更多处于主流文化边缘的全球群体（包括欧洲、亚洲等地的少数群体）积极发声。Erin 对此深表赞同，并强调走到具体社区中与真实的人对话虽然属于“无法规模化”（unscalable）的苦功，却是推动 AI 普惠必不可少的基石。 - 认知鸿沟：主流大模型的微调人员缺乏对跨文化、特定地域事实与背景的理解。 - 算力瓶颈与应对：在缺乏大型数据中心支持的地区，探索微调小语言模型（SLMs）成为突破算力制约的重要方向。 - 社群连接：Erin 强调真正深入社区的基层对话虽耗时耗力，却是实现 AI 公平不可替代的途径。 --- ### [31:32](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1892s) 降低创造门槛：小微社群如何利用自定义 GPT 创业 一位与会观众提问，对于穆斯林社群、老年女性等没有深厚技术背景的数千个细分小微社群（micro-communities），是否可以通过优化 Prompt（例如在提示词中明确申明文化身份、要求模型不设过度政治信息保护）作为折中过渡方案。对此，Erin 明确鼓励大家不要等待或止步于折中方案，而是应当立刻行动起来亲手构建属于自己的 AI。 ```mermaid flowchart TD accTitle: Low-Barrier AI Entrepreneurship Pipeline subgraph Iteration["概念验证与原型"] A["使用 OpenAI API / 自定义 GPT 平台 ($20/月)"] --> B["编写自定义指令算法"] B --> C["公开发布收集真实反馈"] end subgraph Scale["规模化拓展"] C --> D["验证 MVP (如获得5万用户)"] D --> E["迭代构建独立完整产品"] end ``` Erin 详细拆解了构建自定义工具的低门槛路径： - 算力借力：无需拥有专属数据中心，每月只需花费 20 美元使用 OpenAI 的平台功能，就能直接调用顶级模型与底层数据。 - 算法本质：所谓“算法”在落地层面本质上就是清晰的提示指令（Instructions），指导模型按特定群体期望的立场和视角进行交互。 - 敏捷创业：从最小可行产品（MVP）起步，像她当初一样将自定义 GPT 公开并迅速累积 5 万名用户，以此验证市场需求，完成从概念到独立产品的闭环。 - 赋能与反哺：小微社群主动搭建工具，不仅能在合法合规框架下突破大厂通用模型的研究盲区，还能利用巨头算力反向服务自身社群。 在解答完毕后，高校管理者 BJ Kitchen 接着发言，对 Erin 提出的“人人皆可利用 AI 创业、为社群赋能”的理念表示高度赞赏，并探讨如何将此类思维方式在教育与社会领域更广泛地推广传播。 ### [35:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2138s) 挑战科技特权与走进高校分享 现场观众在互动环节向嘉宾提问，谈及在关注世界各地的学生群体（涵盖 K-12 基础教育及高等教育等）时的感受。观众指出，正如观察 Sam Altman 等科技领袖所反映出的现象，正如学者所形容的“科技寡头”或“兄弟会寡头”（bro oligarchy），特权阶层往往“不会主动邀请多元群体来挑战自身特权”，而是“在一开始就寻求服从”。观众以 Erin Reddick 此前提到的“仅在产品开发末期才被邀请去提供反馈”为例，印证了这种表面化的包容，并随即询问 Erin 是否愿意接受付费邀请，前往大学社区向学生们分享这些前沿且深刻的观点。 Erin Reddick 当即给出了绝对肯定的答复。她表示自己与专门负责对接大学演讲的经纪机构签约，日常频繁穿梭于各大高校。她曾在哈佛大学（Harvard University）演讲过三次，两周前刚在乔治城大学（Georgetown University）发表演说，并且定期在华盛顿大学（University of Washington）担任客座讲师，还去过诺福克州立大学（Norfolk State University）等诸多院校。她非常适应且热爱高校的学生群体，认为学生们带来的启发是无与伦比的。 每次面对学生，总会有人站出来向她的理念和足迹发起挑战，而她对此的态度则是热情回应：“保持你的好奇与锐气”（"stay curious and furious"）。看到年轻一代意识到他们不仅拥有构建技术的机会，更能彼此携手共同创造，这令她倍感振奋。她分享了一个具体的教学案例：她曾同时为 70 名学生提供培训，每个小组在现场各创建一个 Custom GPT 模型，由她在前端统一配置指令，随后全场共同进行压力测试与红队对抗（red teaming），整个过程极具启发性与建设性。最后，她向提问者指引了屏幕上的 LinkedIn 账号及经纪人的预约联系方式。 - **高校演讲足迹**：哈佛大学（3次）、乔治城大学、华盛顿大学（定期客座）、诺福克州立大学等。 - **互动实训案例**：曾单场指导 70 名学生分组构建 Custom GPT，并通过全员红队测试（red-teaming）验证模型表现。 --- ### [37:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2258s) 全球视角下的 AI 伦理与乌班图（Ubuntu）哲学 在进入尾声之前，主持人 Nathan Walker 分享了一项即将展开的全球对话计划。他提到下周将启程前往南非开普敦（Cape Town），参与一场关于拟议中的《联合国人工智能与人权公约》（UN convention on AI and human rights）的南北对话（North-South conversation）。该对话的核心目标是将广泛的非西方本土视角、原住民世界观以及伦理体系纳入主流的 AI 伦理与人权框架中。 在开普敦期间，团队将重点探讨非洲传统的“乌班图”（Ubuntu）伦理哲学，研究如何将这种强调人际连接、互惠与共同体精神的实践融入人工智能系统，并在与用户的交互中予以体现。Nathan 强调，这项探索绝非主营业务的边角料或附带项目，而是深植于他们核心基因（DNA）与人生使命的终极追求，并呼吁在场同仁持续保持关注与交流。 ```mermaid flowchart TD accTitle: AI伦理框架拓展路径 accDescr: 展示将非西方本土哲学（如乌班图哲学）融入联合国AI与人权公约框架的路径流程。 subgraph Local_Philosophies["多元本土智慧"] A["非西方/原住民视角"] B["乌班图 (Ubuntu) 伦理哲学"] end subgraph Framework_Integration["国际治理与伦理构建"] C["联合国 AI 与人权公约对话<br/>(UN Convention on AI & Human Rights)"] D["包容性 AI 伦理体系"] end subgraph System_Deployment["技术与用户交互"] E["AI 模型与产品架构"] F["面向全球多元用户的价值反馈"] end A --> C B --> C C --> D D --> E E --> F ``` --- ### [38:44](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2324s) 拉斯维加斯峰会的结语与致谢 Nathan Walker 将话题引回位于拉斯维加斯米高梅大酒店（MGM Grand）的峰会现场，邀请 Erin Reddick 为全场与会同行发表闭幕寄语。 Erin Reddick 首先对所有到场的参会者表达了由衷的感谢。她坦言，在当前的政治气候背景下，自己最初曾有些许担忧，害怕本次研讨会关于多元化的议题名称可能会让部分人望而却步。然而，看到现场座无虚席且大家始终保持开放与求知欲，她感到非常欣慰与敬佩。她呼吁在座的所有同行为自己鼓掌，感谢大家在当下的环境中依然选择亲临现场、保持好奇心、渴望探索未知，并积极投身于关于人工智能多元化与包容性（Diversity and Inclusion）的深度对话之中。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=4s) 按自己的方式掌控作为科技女性的话语权 — Erin Reddick > [06:32](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=392s) 并非叫你不要去，而是要知情前往（go informed） — Erin Reddick > [12:15](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=735s) 这会对你的社群造成伤害吗？请回答是还是否。 — Erin Reddick > [17:49](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1069s) 人们不会去使用那些无法在其中看到自己身影的工具。 — Erin Reddick > [17:49](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1069s) 构建这类模型时，保持动态性是极其重要的……底特律十年前的样貌绝不是今天的样貌。人们的感受会发生改变，你必须时刻保持对这种脉搏的感知，确保那些动态变化的声音能够不断被整合进来。 — Erin Reddick > [21:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=1298s) 在这片混乱的局势中尽最大努力（make the best of this mess） — Erin Reddick > [35:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2138s) 不会主动邀请多元群体来挑战自身特权 — 现场观众 > [35:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2138s) 在一开始就寻求服从 — 现场观众 > [35:38](https://www.youtube.com/watch?v=uQjALlZ3dyM&t=2138s) 保持你的好奇与锐气（stay curious and furious） — Erin Reddick ## Mentioned Nathan Walker, Erin Reddick, ChatBlackGPT, 313 GPT, Rutgers University, Microsoft, AWS, Meta, OpenAI, ChatGPT, Hugging Face, Timnit Gebru, Joy Buolamwini, Ruha Benjamin, Safiya Noble, Sam Altman, BJ Kitchen, The Last Poets, Black Panthers, Harvard University, Georgetown University, University of Washington, Norfolk State University, Minden, Nevada, Detroit ## Video description Nathan Walker, Principal Investigator, AI Ethics Lab, Rutgers University Erin Reddick, CEO, Founder, ChatBlackGPT™ Abstracts This session will explore the critical importance of diversity in AI, focusing on how building inclusive teams and technologies can lead to more equitable and innovative outcomes. Panelists will discuss the challenges and opportunities in promoting diversity within AI development, from mitigating bias in AI models to ensuring representation in the teams that create these technologies. Learn how diverse perspectives can help solve complex AI challenges, drive ethical innovation, and create systems that better serve a wide range of users. The conversation will also cover strategies for improving diversity in the AI workforce and the long-term benefits of inclusive AI development. [Source](https://www.youtube.com/watch?v=uQjALlZ3dyM)

[0:04] Erin Reddick: 按自己的方式掌控作为科技女性的话语权 [6:32] Erin Reddick: 并非叫你不要去，而是要知情前往（go informed） [12:15] Erin Reddick: 这会对你的社群造成伤害吗？请回答是还是否。 [17:49] Erin Reddick: 人们不会去使用那些无法在其中看到自己身影的工具。 [17:49] Erin Reddick: 构建这类模型时，保持动态性是极其重要的……底特律十年前的样貌绝不是今天的样貌。人们的感受会发生改变，你必须时刻保持对这种脉搏的感知，确保那些动态变化的声音能够不断被整合进来。 [21:38] Erin Reddick: 在这片混乱的局势中尽最大努力（make the best of this mess） [35:38] 现场观众: 不会主动邀请多元群体来挑战自身特权 [35:38] 现场观众: 在一开始就寻求服从 [35:38] Erin Reddick: 保持你的好奇与锐气（stay curious and furious）

[12:23] Erin Reddick: Having the opportunity to integrate voices in the design lifecycle early on and often is a really good way to mitigate risks of harm in communities and what you're deploying. [4:50] Erin Reddick: For me, that's not a political inflammation. That's a safety concern. So when you're saying you have the default model of all generative AI and I can't freely use it safely and considerably as everyone else, then why would I go to that tool? [14:55] Erin Reddick: People aren't going to use things that they don't see themselves in. And there's a huge chunk of people that can't see themselves in the day-to-day AI right now.

[0:04] We are going to be discussing fostering [0:06] diversity in AI, but we are doing so in [0:10] the absence of two of our colleagues who [0:11] are going to join us for this. So, we [0:13] send them our love and and all our [0:16] support. But this is a really special [0:19] time to dive deep into the life and the [0:22] work and the legacy I should say of [0:25] Aaron Reddic, CEO and founder of Chat [0:28] Black GBT. So let's begin with your [0:32] amazing career trajectory. You have [0:34] worked for the world's biggest tech [0:36] companies, Microsoft, AWS, Meta. And so [0:41] looking back at how you entered into [0:44] this world of technology, what were [0:47] pivotal moments that prepared you for [0:49] this moment? [0:51] >> Yes. Um so very glad to be here. Happy [0:55] is it midafter afternoon after it's way [0:57] after midafter afternoon but I'm very [0:59] excited to have this conversation. Thank [1:01] you so much. Um yes so being in tech uh [1:05] I come from Seattle. I was there for [1:06] about 10 years. So for me, when it comes [1:10] to the diversity conversation, being in [1:12] Belleview, Washington, [1:14] I feel like I was one of maybe 10 people [1:17] of color I've saw across my entire time, [1:20] 5 years between like being on site for [1:23] sure. And it was like coming into tech, [1:28] I had that experience, but coming out of [1:31] it when I got laid off in 2023, I took [1:34] my severance and launched my startup. [1:37] And I had kind of the same feeling [1:39] exploring technology independently as I [1:41] did inside the company. There's things [1:44] that I can do to affect change that are [1:46] helpful to my team, but I didn't really [1:48] feel like I had so much power and [1:50] influence that I wanted to make impact. [1:54] So coming up with my endeavor into [1:59] culturally informed AI kind of happened [2:01] when I started doing research and [2:03] noticed there were two different [2:04] narratives to the conversation about [2:07] artificial intelligence. There was one [2:09] saying it's the best thing since the [2:10] internet and then there was another [2:12] saying that it's erasing culture and [2:15] like misrepresenting historical facts [2:18] for the black community. And I decided [2:22] that I was going to take my power back, [2:24] remain a woman in tech on my own terms, [2:27] and instead of running right back to one [2:30] of those companies, I stayed [2:32] independent. And I've been operating for [2:34] a couple years now [2:36] >> and doing really well with that. And one [2:38] of your moral compasses seemed to be to [2:43] focus on cultural responsiveness and [2:45] representation. So was there a moment [2:47] when you realized that cultural [2:49] representation in AI in particular was [2:52] not just important but it was absolutely [2:54] urgent. [2:55] >> Yeah. So I was watching a lot of [2:59] information about AI. My background in [3:01] the tech companies was UX research, not [3:04] necessarily AI specific, but I'm working [3:07] alongside senior leadership helping them [3:11] understand like customer bases like [3:13] recruiting for different um people who [3:16] are going to test products UX research [3:19] and I felt that in my research of AI [3:22] there was nothing really being done [3:25] about the big issues happening. So you [3:28] know we all saw what happened with [3:30] Google and the misrepresentation of like [3:34] historical figures and that was a big [3:37] problem but there was really nothing [3:38] being done proactively. So when OpenAI [3:42] came out with their customizables I [3:44] decided to test a theory. I thought, [3:48] okay, I don't believe that it is a issue [3:51] with the data per se because if it's [3:54] using the data of all the internet and [3:56] it's good enough for everyone else, I'm [3:57] pretty sure black people have been [3:59] online long enough for enough data to be [4:01] there. And I started focusing on the [4:04] algorithm, the set of instructions on [4:06] how it pulls resources. And that really [4:09] was helpful in understanding how a [4:12] different perspective and designing for [4:14] different perspective could be useful. [4:16] And so I created the customizable as a [4:19] proof of concept. It has over 25,000 [4:22] people using it currently. And because [4:25] of its popularity and usefulness, I was [4:27] able to raise enough to build my own [4:30] version as well. So it's still up and [4:33] running and free for anyone to use and [4:36] it's been pivotal for people who want a [4:40] different perspective when they're using [4:41] generative AI. So tell us about that [4:44] mission and that vision. Who are you [4:47] hoping will use chat black GBT? [4:50] >> Yeah. Anyone who is interested in just [4:53] not just having one single narrative. I [4:56] feel like a lot of people have [4:57] conversations about this is wrong and [4:59] this doesn't feel like me and you have [5:02] to prompt over and over and over again [5:04] to get something relatable. And that [5:06] could be just because you're not of the [5:08] majority or a male or somebody who isn't [5:12] in a marginalized community. And I feel [5:15] like what I've built answers in a way [5:19] that speaks to that, but without saying [5:22] this is exactly what this is. It's very [5:24] much like here's a perspective that you [5:27] could consider and why it may be [5:30] attributing to the type of answers [5:31] you're getting. A example that I use in [5:34] my keynotes is I asked Chad GPT what is [5:37] a black job and it answered that a black [5:41] job was drug trafficking um illegal [5:45] activity unregistered businesses and [5:48] because like I'm building with AI [5:50] separately I understand it's probably [5:52] thinking black market but as it is not [5:55] able to pull from contemporary context [5:58] when it comes to the black lived [6:00] experience I knew that that was a huge [6:02] gap app. And so chat black GPT answers, [6:06] do you mean what was historically known [6:08] as negro jobs in the times of [6:10] segregation and these other factors? And [6:12] then it listed the jobs that were in [6:15] fact being um actually talked about in [6:18] the context of that comment made. So I [6:21] built something that could help us [6:23] relate to the technology, but just prove [6:26] that it's possible and different [6:28] perspectives can be powerful when you're [6:30] using AI. [6:32] So over lunch we had a little experiment [6:35] of our own and comparing how my chatbt [6:39] was responding to a very simple question [6:42] which was tell me everything um about [6:46] tell me what I need to know about mind [6:48] Nevada. [6:50] Minden Nevada is about 8 hours north of [6:52] here where I was raised in northern [6:54] Nevada and um and it had a very [6:59] Wikipediaike entry. It gave [7:01] demographics. It gave some history and [7:03] some background. But it was missing some [7:05] really significant information. When we [7:08] asked Chat Black GBT the same question, [7:13] we got a lot a very different and [7:15] historically accurate response. Do you [7:18] want to talk about that experience and [7:20] what that was like doing that test live [7:22] just an hour ago? [7:23] >> Yeah. So, uh, in case you are not aware, [7:27] that town is historically known as a [7:29] sundown town. And as recent as 2023, [7:34] efforts have finally come to get rid of [7:36] some of the mechanisms that were very [7:38] symbolic of those times such as a bell [7:42] that was ringing um that was supposed to [7:44] signify, you know, this time to [7:46] separate, segregate, leave, and so [7:48] forth. So my tool is able to present an [7:52] answer that is going to inform me how I [7:55] can operate in that area which it said [7:59] it's not that you don't go but go [8:00] informed. Uh versus just a GPT saying [8:04] hey it's got German heritage and it's [8:06] got this heritage. That's not helpful [8:09] for me as a traveler if I were to go [8:11] there and come up against people who are [8:14] holding on to some uh you know legacy uh [8:19] activities or mindsets that would be [8:21] harmful for me to even exist there. So [8:25] yeah, that's what it's really designed [8:26] to do just to open up that perspective. [8:30] And as someone who was raised in that [8:32] area, I could I could really relate to [8:35] the significance of this sundown town [8:39] history. And it's not just its history, [8:42] it's it's contemporary conflicts. So, [8:44] you may have heard on the news that in [8:46] 2020, the sheriff of the town [8:50] told in in the news on um on social [8:53] media that he would refuse to provide [8:56] any police um utility presence to the [9:00] library because they were considering to [9:04] advertise Black Lives Matter. He later [9:07] talked about that if any of those BLM [9:10] supporters come to mind, Nevada, he [9:13] would arrest them immediately. This [9:16] created a whole series of controversies [9:19] and brought up this history of the bell [9:22] ringing at 5:00 p.m. each each night. My [9:26] aunt in particular was very active in [9:29] the '9s to petition to get that removed. [9:33] And there's a very interesting narrative [9:35] about trying to say this is not about um [9:40] trying to get rid of indigenous people [9:43] at sundown anymore. This is about [9:46] honoring our service providers um and [9:50] the and the firemen and the police in [9:52] the area. So there's this [9:55] incredible kind of um replacing of [9:59] narratives and this ahistorical language [10:02] that takes out the very fact of this [10:05] racialized set of laws that were [10:08] prevalent up until just a two years ago [10:11] like you mentioned when then it took an [10:14] act of the Nevada state legislature and [10:17] an investigation by the Nevada Attorney [10:19] General to actually remove this sundown [10:23] practice in our hometown. So the very [10:26] fact that our models were responding in [10:30] different ways led us to ask my model of [10:34] chatbt why didn't you tell me about what [10:38] happened in 2020? Why didn't you tell me [10:41] about this history? And it gave a series [10:43] of answers for why that was the case. [10:47] What was your take on that whole [10:48] exchange? [10:50] I found it um pretty much what I would [10:53] expect it to say, but it was interesting [10:55] that it was so forthcoming. And um to [10:58] put it in simple words, it kind of said [11:00] that the AI is tuned to not talk about [11:05] politically inflammatory topics or [11:09] anything like that. But for me, that's [11:12] not a political inflammation. That's a [11:15] safety concern. So when you're saying [11:17] you have the default model of all [11:20] generative AI and I can't freely use it [11:24] safely and considerably as everyone [11:26] else, then why why would I go to that [11:30] tool? The same tool that you're saying [11:31] should be in every single part of my [11:33] life. And so the fact that you know I [11:37] was able to create something with the [11:39] same exact data there's no proprietary [11:41] information on the customizable uh the [11:44] same exact data just an equitable [11:46] algorithm like they are the problem but [11:49] also the solution which is why I'm not [11:52] anti- AI. I don't feel that AI is [11:55] specifically racist or purposefully [11:58] derogatory. I think it's how it's [12:00] designed and how it's tuned. And so you [12:02] look at who's tuning and what kind of [12:04] issues do they see and what kind of [12:07] issues will they never be able to see [12:09] because it's not their lived experience. [12:11] And that's why the big dirty D word [12:15] diversity matters, right? It's not just [12:19] for fun. It's literally, you know, [12:21] sometimes for safety. [12:23] >> So when we talk about diversity in the [12:25] technology, we're not just talking about [12:27] the the data set itself, right? We're [12:29] talking about diversity at every stage [12:32] of the AI life cycle from from its [12:35] development from its design. Why is it [12:38] important to have diverse teams at every [12:42] stage of the of that before it even gets [12:45] deployed and what benefit does that have [12:49] for the company? [12:51] >> Yeah. So when I was working in um tech, [12:54] there were often opportunities for me to [12:57] sit on specialized groups and panels and [12:59] be consulted for like large campaigns. [13:02] And basically what I would do is I would [13:05] jump into a large meeting and they would [13:09] say, "Here's our campaign. Here's what [13:10] we've designed. We want to deploy it. [13:12] Will it create harm for your community? [13:15] Yes or no?" That's a lot of pressure for [13:18] an individual contributor to be talking [13:20] to VPs and directors and to say, "Hey, I [13:24] think this is a little racist or a [13:26] little sexist. I I don't think it's [13:29] good." That is so much pressure. And [13:33] it's like not being embedded in the [13:35] process and invited to share these [13:37] concerns, but just a yes or no at the [13:40] very end. That's that's not okay. It's [13:43] not even like career-wise responsible [13:45] for me to be openly honest. I was I was [13:48] like you cannot release this. This is [13:50] terrible uh for many reasons and I [13:54] didn't care. But again, I was a little [13:56] bit naive. Um but having the opportunity [14:00] to integrate voices in the design life [14:04] cycle early on and often is a really [14:07] good way to mitigate risks of harm in [14:10] communities and what you're deploying. [14:12] And it seems like that is done not only [14:15] in the data collection stage like you [14:16] mentioned throughout the life cycle but [14:18] also in in the the pre-ervice training [14:21] but also when um you can do some [14:24] post-training work where you with [14:26] reinforcement learning with with human [14:28] feedback you can get some really strong [14:31] refinement in how AI speaks not only [14:35] what it speaks right and so it's in this [14:38] context you're providing this incredible [14:40] nuanced way to go deep into a [14:44] community's lived experience with chat [14:46] black GPT. Are tell me tell me more [14:49] about how you think that model can be [14:51] replicated for other communities. I know [14:54] you've been working in Detroit. [14:55] >> Yeah. [14:56] >> Yes. I recently just left DC which was [15:00] turns out to be a good move because now [15:02] the National Guard is deployed in the [15:03] streets and I was like 8 minutes from [15:06] the White House. So, I moved to Detroit [15:09] to be closer to family, but I got [15:12] involved in the tech endeavors that [15:17] they're doing. And so, I decided that I [15:20] would replicate what I'm doing, but on a [15:24] smaller scale and call it 313 GPT. So, [15:28] essentially what I'm doing is it's a [15:30] co-creation of generative AI with the [15:33] communities and neighborhoods. So right [15:36] now there's a group called like the last [15:39] poets in New York or you've got Black [15:41] Panthers that are all dying. Like people [15:43] are literally dying and their stories [15:45] aren't being told or preserved [15:47] digitally. And so a project like mine [15:50] will go into these communities, ask the [15:53] questions like what's it like to live [15:55] here? How long have you been here? Who's [15:57] the elder in your community? Who's like [15:59] most influential? What are the school [16:01] systems doing for your community? all of [16:04] those questions can be answered and then [16:07] with the right sentiment analysis it can [16:10] be built into how the answers are [16:12] reflected and so that's how I've [16:14] designed for chat black GPT it's all [16:17] open source like when it comes to the [16:20] qualitative research I invite people in [16:23] to share your experience and once I do [16:26] that on a city level I I think I can [16:29] just lead as a replicable [16:32] um system. You know, this is how you [16:35] integrate community voices into your [16:39] models. And because they want us all to [16:42] use it, like people aren't going to use [16:44] things that they don't see themselves [16:45] in. And there's a huge chunk of people [16:48] that can't see themselves in the [16:50] day-to-day AI right now. [16:52] It's almost as if AI is a mirror to [16:56] humanity for all of its horrors and its [17:00] potential in that we're seeing the [17:03] discriminatory nature of of humans in [17:07] the data sets being reflected back to us [17:09] and yet there isn't equal opportunity to [17:13] refine that image in return. And I [17:16] really like how you're using kind of an [17:18] oral history approach and [17:20] community-based scholarship in order to [17:22] do that refinement and engagement. Are [17:25] you imagining that this like culturally [17:27] responsive type of AI systems can be [17:30] replicated in other cultures? Are you [17:32] imagining like a chat Muslim GBT and [17:36] Jewish GBT and a Mormon GBT and I'm [17:39] giving religions as an example, but [17:40] there's many others to include. What is [17:43] your big vision for uh replicating this [17:46] this plan you this blueprint you have? [17:49] >> Yeah, I mean firstly just proving that [17:52] the concept is valuable to communities. [17:54] I don't think you know 25,000 people [17:57] would be using it without any paid [18:00] advertising, without any real marketing [18:03] other than just me on LinkedIn. Please [18:05] connect with me. That's just my [18:06] LinkedIn. Um that's all that I've done [18:09] to get this tool out there. And so it's [18:12] really word of mouth and that organic [18:15] um conversation propelling the [18:19] usefulness of it. And so like when you [18:21] want to learn something, you go to an [18:24] expert. The reason I do chat black GBT [18:26] is because I can expertly define my [18:30] lived experience and validate a lot of [18:33] what is, you know, coming out. So that's [18:36] pretty much why I chose that. If I was [18:39] Muslim, I would try to build something [18:41] like that as well. Um or any anything [18:44] else, but I can only build I think with [18:47] how I can validate. Um and so as long as [18:51] long as I can provide a a blueprint, I [18:55] think people should you know have the [18:57] right and should exercise their right to [18:59] use open source technology to build [19:02] systems that represent them and do that [19:04] with community. uh because they want you [19:08] know everything to go back to AI but it [19:11] cannot be generalized [19:14] uh without bias and it's almost as if [19:17] human data human experience the our [19:20] lived experiences is the next frontier [19:23] in data. Now, this could result in a [19:25] whole series of privacy concerns because [19:28] if these models know certain things [19:30] about me and my community in these [19:32] really intimate ways, can it result in a [19:35] privacy risk? or as we saw in the whole [19:38] social media uh landscape. Um these [19:42] algorithms started to filter us into [19:44] these echo chambers politically and [19:47] culturally until we were engaging in our [19:50] own confirmation bias and like basically [19:53] reinforcing our own views to the to the [19:56] point of creating huge disruptions in [20:00] cultural and political systems all [20:02] around the world. How might this type of [20:05] approach, this blueprint for customizing [20:08] GBTs to the lived experience of a [20:11] particular community pro not make sure [20:14] that we're not replicating those kinds [20:17] of echo chambers, but instead trying [20:20] something else? like you've you've [20:21] talked about this switching on of [20:23] toggles between different points of view [20:26] and how the multiple perspectives could [20:29] be a way to widen our circles of [20:32] compassion. [20:33] >> Yeah, I think keeping building these [20:36] type of models dynamic is extremely [20:38] important. not having anything fixed or [20:41] hard-coded, just a way that you can [20:45] continually get information from that [20:49] society or that group like at a constant [20:52] basis. That's really helpful for making [20:54] sure you're not just echoing a certain [20:56] same sentiment because Detroit 10 years [20:59] ago isn't Detroit today. People are [21:01] going to feel different, but you have to [21:02] keep a pulse on that and make sure that [21:05] those voices are integrated. And I think [21:07] that's how you can avoid such something [21:09] like that that's not really going to be [21:11] that helpful [21:13] >> and and maybe it could be a part of [21:15] community healing. So if people for [21:17] instance in mind Nevada as our kind of [21:19] case study for today um could help [21:23] understand its history. it could engage [21:25] in that reckoning and engage in the in [21:28] the community voices, the oral history [21:30] making to ensure that it doesn't [21:32] replicate the same systems of [21:34] discrimination that it did in the past. [21:36] So [21:38] here's here's a basic question. Are you [21:41] hopeful given given all of your work in [21:44] tech with all of these companies and now [21:47] diving deep into this work, are you [21:49] hopeful that we can at least engage if [21:53] not solve these issues of of bias and [21:56] discrimination? [21:59] >> Yeah, I mean I think that like I pay [22:01] attention to a lot of different people [22:03] who work in this field. Uh I don't just [22:06] listen to people who champion it. I also [22:10] admire Timnit Jabru. I also admire Dr. [22:13] Joy Bloomini, Dr. Ruha Benjamin, Dr. [22:16] Safia Noble. I've met these uh [22:18] incredible women who understand AI like [22:22] thoroughly back and forth and they talk [22:25] about the promise and the peril of it [22:28] and understanding both sides. What I'm [22:31] here to do is try to make the best of [22:34] this mess is what I call it and try to [22:37] slow down the digital divide of at least [22:41] the usage of it. If you're going to be [22:43] using generative AI, you should be able [22:45] to use it just as quickly and [22:47] efficiently as the majority of people. [22:50] And so just having something that you [22:51] can relate to, you don't have to think [22:53] twice about prompting the context of [22:55] your lived experience. That's also [22:57] saving on like compute. Like that's [22:59] saving money. that's saving wasted time [23:02] saying no, this is what I mean. No, this [23:04] is where I'm from. No, this is how I [23:05] feel. So, I think in certain ways, in [23:10] even small ways, I'm creating a pocket [23:13] of psychological safety in a tool that [23:15] wasn't designed to be, you know, [23:18] understanding of cultural experiences. [23:21] Um, while also giving it a chance to [23:24] like just like be dynamic and [23:26] reflective. Regarding reflectiveness, [23:29] you do a really interesting thing with [23:31] your app, which is give people a set of [23:35] readings to consider after they see [23:38] initial response. Why did you take such [23:41] a sophisticated literary approach to [23:44] that kind of work? [23:45] >> Thank you. Yeah, because I mean, if you [23:48] think about most of the motivation when [23:52] it comes to building these systems, [23:53] companies want you to stay within the [23:55] AI. They want you to stay in it. They [23:57] want it to be the answer to everything [24:00] that you have, all your questions, but [24:02] at the same time, they're also saying [24:05] double check, you know, like it can be [24:07] wrong and then there's it can [24:09] hallucinate and all of these things. And [24:12] so I thought, why not design something [24:14] that's going to pick up on what you're [24:16] trying to understand and direct you [24:18] directly to resources that'll help you [24:20] further your learning independently that [24:23] are from authors that are going to be [24:26] experts on the topic. So it's like [24:29] taking you back into the real world. I [24:31] think that's something that no AI really [24:32] does, but it's it's important for it to [24:36] spread those resources as well because [24:38] that's what they're telling us to do [24:39] anyway. But how do we know where to look [24:41] and who to read? So I designed it that [24:44] way. [24:45] >> I'm curious about your self-worth and [24:47] and what I mean by that is you have you [24:50] bring a confidence and a curiosity to [24:54] this work where others feel despair. I [24:57] think all of us who have been engaged in [25:00] these questions can easily just throw up [25:02] our fist and scream into the wind and be [25:05] depleted by the number of injustices [25:08] that that we see these disruptive [25:11] technologies create. And yet you seem to [25:15] have this self-worth. And forgive me if [25:18] I'm overspeaking. I I'm just getting to [25:20] know you today. But I I deeply admire [25:23] how you seem to be so resilient [25:27] in this face of these systemic acts of [25:31] resistance. Like in a way you seem to be [25:34] finding your voice in this moment and [25:38] and in doing so you are empowering [25:40] others. What does that say about your [25:43] selfworth as a leader? [25:45] >> Thank you very much. That was very kind [25:47] of you. I hope we can be friends after [25:49] this. I [25:51] He's really amazing. Um, but uh, yeah, [25:55] so I've been through a lot, you know, [25:57] like when I got laid off, it was the [26:00] same week that my dog got hit by a car, [26:02] my grandma died. So, I was knocked down [26:06] to a point where I just didn't have any [26:10] clear vision of what my future was, who [26:13] I am as a person. My identity was so [26:16] tied up in working in tech, you know, [26:18] because it's a 24-hour facility where [26:20] they've got prepasted toothbrushes and [26:22] deodorant and brushes and combs and and [26:25] you don't need to leave. There's [26:26] breakfast, lunch, and dinner free buffet [26:28] style. I was really in it to win it. Uh, [26:31] and so when that suddenly goes away, [26:34] there should probably be a study done on [26:35] the psychological effects of that, but [26:37] gone in an instant, I really had to [26:39] rebuild from the ground up and taking [26:43] ownership over like why do I keep [26:45] working at these companies? Is it [26:46] because I think it's cool or whatever or [26:48] do I really love technology? And so [26:52] finding out that I could build in it, a [26:54] discovering hugging face, the [26:55] understanding like open source like oh I [26:58] have access, I can put these things [27:01] together and make something that I think [27:03] is helpful not just for me but for other [27:06] people like that was so inspiring. and [27:09] seeing other people realize that they [27:11] can take advantage of those same [27:14] resources and build to their voice and [27:17] their dreams, it just feels really [27:21] really good, you know, and I think I'm [27:22] I'm living through kind of like [27:26] reggrounding my life after so many like [27:29] tragedies in uh a way that was [27:32] productive. And I and I did go to [27:34] therapy, by the way. But um so I feel [27:37] like those things have helped keep me [27:39] grounded. Um and just like I think [27:42] starting from truly zero emotionally [27:45] like [27:46] in so many different ways. I think that [27:48] helped me get kind of where I am. [27:51] >> I think that deserves a round of [27:52] applause. Congratulations for your [27:56] resilience. [27:58] And you forgot to mention that was all [28:00] kind of in the pandemic, too. There's [28:02] all kinds of other stuff going on in all [28:04] these different seasons. We now want to [28:06] open it up for conversation with you all [28:08] because we know we can't have a [28:11] conversation about communities, lived [28:13] experiences without turning to our our [28:17] collegial community that has gathered [28:18] today. What questions do you have? What [28:21] is at the forefront of your mind? And [28:23] you can just come on up and I'll I'll [28:25] give you the mic to pose a question if [28:27] you have one or I'm happy to come to you [28:29] as well. [28:30] >> Hi everyone. [28:32] Um, I actually connected with Erin on [28:34] the app when I saw the topic cuz I I'm [28:38] kind of in between the US here and um, [28:42] Africa. Um, that's Nigeria. That's where [28:44] I'm originally from. And I'm an AI and [28:48] tech professional. And we are also on [28:51] the receiving end of a lot of the [28:54] models. Chad G. I use Chad GPT plus [28:58] every day. like if I open it right, it [29:00] crashes. That's how much I use it. But [29:02] um it doesn't quite get our point of [29:06] view and that's like a step further from [29:09] what you have been through as a black [29:12] American. [29:13] >> I'm talking African African now. [29:16] So we are also trying to see how we can [29:19] I mean just until recently [29:22] um the generative AI didn't quite get [29:25] our type of images until recently when [29:29] Google kind of upgraded and [29:33] yeah it's it's a real challenge and we [29:35] also back in Africa know that there's a [29:38] lot of work to be done especially [29:40] because like was mentioned the people [29:42] fine-tuning these models aren't really [29:46] able to relate to anything that they [29:48] don't know. So yes, there is the need [29:51] for a lot more inclusivity. [29:54] Um a lot more of us are trying to get in [29:57] the game. Um also put together some of [30:01] our own models. Um not just ride off of [30:04] what OpenAI and all these other [30:06] companies have done, but we don't really [30:08] have the power, the data centers, the [30:11] compute power. So it's a unique approach [30:14] that it's it's a whole lot of other [30:17] stuff going on where we're trying to [30:18] also figure out how to solve problems [30:20] also maybe using smaller models SLMs not [30:24] just the LLMs and um see how and that's [30:28] just from my point of view there's so [30:30] many other regions that I'm I'm sure [30:32] going through the same thing probably [30:33] even some parts of Europe, Asia um and [30:38] um I'm just here to say thanks for the [30:40] work. I'm going to try out the um app [30:44] and I just want to encourage anyone else [30:46] who's not in the mainstream culture of [30:49] what the um AI tools are used to to do [30:54] more work and see how we can, you know, [30:56] bring more inclusivity into the space. [30:59] Thank you. [31:00] >> Thank you so much. [31:01] >> Why were you not on our panel? You [31:04] should have been on our panel. [31:06] >> That's amazing. Uh and I appreciate your [31:09] acknowledgement. It's it's very much [31:11] unscalable work, you know, actually [31:14] talking to people, sitting down, going [31:16] to their communities, meeting people [31:18] where they are. It's difficult, but it [31:20] is so worthwhile because AI can really [31:24] reach a lot of people. And uh everything [31:27] you said I definitely agree with. Like [31:30] it's difficult, but so worthwhile. [31:32] >> Thank you. I am wondering I mean I'm [31:36] blown away by the little test you did [31:39] over lunch right and um so considering [31:45] that the Muslim community [31:47] older women community you know other [31:50] communities can't just jump right in and [31:52] build an AI the way you have is there a [31:55] halfway step where I mean is there a [31:57] have you tested um like prompting saying [32:02] like I think you said it um specifically [32:07] aims at protecting you from certain [32:09] political information. You think that if [32:11] we start out our prompts or introduce [32:14] ourselves and say don't protect me from [32:16] anything and by the way I'm of this [32:19] particular population [32:21] and then move forward. Have you? Because [32:24] what you've built is amazing, but [32:27] there's thousands of micro communities, [32:31] demographically micro communities out [32:32] there that can't do that. [32:34] >> Why? Why? Why do you feel that way? [32:38] >> Um [32:39] because I want them to do it today. [32:41] >> You can do it today. [32:43] >> It by [32:45] >> so elaborate [32:46] >> my customizable [32:48] I specifically created it within [32:51] OpenAI's [32:53] uh API, their forum because anybody can [32:56] do it. For $20 a month, you have access [32:59] to build your own GPTs. And when people [33:02] say, "Oh, you know, algorithm," all that [33:04] really means is instructions. And you're [33:07] using the exact same data that they have [33:09] for what they call the default model. [33:12] You get to use that data and you get to [33:14] write an algorithm of your own and say, [33:16] "This is how I want you to answer when [33:19] you use this." And then you can set it [33:20] to public and do what I did and 50,000 [33:23] people get on board and then you build [33:25] the real thing because you've proven [33:27] your concept with a minimum viable [33:30] product and that's entrepreneurship. So [33:33] you're an entrepreneur if you didn't [33:35] already identify that way but you can. [33:38] You really can. And they say like people [33:42] will say, "Oh, well, you know, I don't [33:44] know how to code and I don't." And then [33:46] vibe coding came out and vibe coding is [33:48] acceptable, but writing an algorithm [33:50] that customizable isn't. Like I've done [33:52] it the hard way and the easy way, which [33:56] still isn't easy because you're still [33:57] designing and writing and it takes a lot [33:59] of iteration. But you can create [34:01] something that's going to answer to what [34:04] you desire with this exact same [34:08] information. And when he's talking [34:09] about, oh, we don't have the compute [34:10] power. I really don't. I don't have a [34:13] data centers or comput power, but they [34:15] do. And they have a right to learn from [34:18] what you're doing. And that's why they [34:20] let you do it for $20 a month. A lot of [34:23] the people using my tool, that's [34:24] informing them on the back end of stuff [34:27] that they maybe even legally can't go [34:29] and focus on in research groups [34:31] nowadays. So if you go and you do that [34:34] for a demographic or a micro community, [34:37] that's something that you're empowering [34:39] not just yourself and your community, [34:40] but also for them to kind of get around [34:43] targeting you and still benefit from the [34:46] information that's helpful. So yeah, [34:50] please feel empowered to do that [34:52] >> and and I'd be happy to meet anyone [34:54] after u just to meet in the little [34:57] tables out here to give you a demo on [34:59] how you can do that directly. [35:03] Yeah, thank you. Um, I knew this would [35:05] be a good session and it was better than [35:06] I thought it would be. So, thank you [35:07] very much. Uh, my name is BJ Kitchen. [35:09] I'm a higher ed administrator. If there [35:12] are any faculty in the room, please [35:13] don't hate me from the outset. Uh, but [35:15] the reality is um everything you said [35:19] around the entrepreneurialship they you [35:21] can do it now. Even if you don't know, [35:23] maybe you can. My question is what I [35:27] would really like to see more of is the [35:29] kind of thinking and entrepreneurialship [35:32] and and ideas that you've shared here [35:34] today spread like wildfire [35:38] >> across the up and cominging students in [35:40] the world K through 12, higher ed and [35:43] and so forth. One of the reasons for [35:46] that is I've been as I've been watching, [35:48] you know, Sam Alton and the and the [35:50] other, you know, uh I forget how [35:52] Caldwell put it. I think um tech [35:54] oligarchy, bro oligarchy I think she [35:56] called it. Um this idea that privilege [36:00] doesn't invite diversity to come [36:03] challenge its privilege. [36:05] >> It looks for obedience upfront. [36:08] >> And you gave a great example of that [36:09] when you said you were invited at the [36:11] end [36:12] >> to give feedback. I think that's a good [36:13] example of that. So my question is and I [36:16] pay you for your time. would you be [36:18] willing to speak to a university [36:21] community to uh talk to them about these [36:23] very ideas that you're sharing today? [36:26] >> Uh absolutely. Uh so I've spoken at [36:28] Harvard three times. I spoke at [36:31] Georgetown University a couple weeks [36:33] ago. I've spoken uh I regularly guest [36:37] lecture at University of Washington. Uh [36:39] Norfolk State University. Like I could [36:42] list a lot. Yeah. Thank you. I I'm [36:44] signed with an agency and all they do is [36:46] run me to universities. So, I'm very [36:49] used to that crowd and it is so it's [36:51] inspiring and there's never a time when [36:53] a student doesn't come up to me and [36:55] challenge my footprint and I'm like, you [36:57] stay curious and furious. I love that, [37:00] you know, and it's it's one of the best [37:02] crowds to be in front of. Um, and just [37:05] to see them light up and understand that [37:08] they have an opportunity not just to [37:10] build but to build with each other. I [37:13] even have trained a group of 70 students [37:16] at once. Each table created a custom GPT [37:19] and I plugged the instructions in up [37:21] front and we all pressure tested and you [37:24] know red team each model and it was just [37:27] a beautiful experience. So yes, would [37:30] love to um that's my LinkedIn and that [37:33] phone number is my booking agent [37:35] actually. So I would love to thank you [37:38] so much. Now before we uh ask for your [37:42] closing words, I just want to do a a a [37:45] plug to to my my Nigerian new friend [37:49] here. Um next week I will be in Cape [37:52] Town where we are having a north south [37:55] conversation about the proposed UN [37:57] convention on AI and human rights. and [38:00] we're looking at how to incorporate a [38:03] wide variety of non-western indigenous [38:06] points of view, perspectives, ethics [38:09] within the larger AI ethics and human [38:11] rights frameworks. So, specifically [38:14] while we're there, we're going to be [38:15] looking at Ubuntu ethics and how certain [38:18] aspects of of that practice can be [38:21] incorporated into AI and reflected back [38:24] to its its users. And so, please [38:28] continue this conversation with us [38:30] because this is not something that's [38:32] just a an offshoot of what we've been [38:36] doing in our either of our works. This [38:39] is core to our our DNA. This is core to [38:41] our life's purpose. Which leads me to [38:44] ask, so here you are giving a talk at [38:49] MGM Grand in Las Vegas. [38:53] What is your closing words? What are [38:55] your closing message that you want to [38:57] give our colleagues today? [39:00] >> Well, I want to say thank you for [39:03] showing up. I was a little bit worried [39:06] that the title of the session would be a [39:10] bit scary for people uh considering the [39:13] political climate right now. I say that [39:16] respectfully. Uh, and I think you should [39:19] really give yourselves a round of [39:20] applause for just still even being here [39:22] and being curious and wanting to know [39:24] more and being a part of the discussion [39:26] and conversation about diversity and [39:30] inclusion. So, I think that we should do [39:33] that.