ATD Footer Logo

Vince Han is a senior level executive and entrepreneur with significant experience starting and scaling digital media and technology companies that rely on venture capital to attack underserved marketplaces. Vince has a proven track record in raising capital, recruiting top‐level talent, forging strategic business development partnerships, and leading teams to execute business and product development plans. Currently, Vince is President of ChangeAnything, a division of VitalSmarts, an innovator in corporate training and organizational performance. ChangeAnything is a dynamic online platform that helps people make the changes in their lives and workplace that are important to them. Prior to ChangeAnything, Vince held executive roles with a variety of leading digital media companies such as Culture11 and Adswizz. Vince was founder and CEO of Aha Media Inc, an innovative company revolutionizing the way video is programmed for online audiences. Prior to starting Aha, Vince was an “Entrepreneur‐In‐Residence” (EIR) at Battery Ventures which co‐founded Ruckus Network, the largest provider of legal, on‐demand, digital music and movies serving the college market in the United States. While at Ruckus, he served in several key executive level positions such as board member, CEO, and President. Vince is currently a board member of Reptide Media, a digital ad sales rep firm that specializes in high CPM, custom content ad campaigns. Vince also advises several early stage, venture‐backed startups, lending his experience with maneuvering successfully through the digital and social media landscape as well as assisting companies to build winning management teams. Vince holds a B.A. from Brigham Young University and an MBA from the MIT Sloan School of Management.

Vince Han

Building Enhanced Performance Capability: Tools, Techniques, and Technologies

September 2017 - TD Magazine

Are Chatbots the Future of Training?

Imagine how businesses suffer when questions like these go unasked: When these questions go unasked, meeting employees needs becomes a guessing game. As a result, employee turnover can increase, and company culture suffers.Chatbots can help field man

I Have a Question

在过去的几年中，许多组织都在不断探讨有关聊天机器人Chatbot的话题。探讨Chatbot的功能，以及如何为人才发展提供支持。这是一种引人注目的沟通媒介，因为与聊天机器人交换消息就像与朋友发短信一样容易，另外，这位虚拟成员从来不会疲惫，也不需要睡觉，对于很多人才发展官来说，这简直是他们的梦想。和任何科技创新一样，聊天机器人的实施是否能够成功取决于完善的设计、出色的项目规划和卓越的写作水平。许多人才发展团队可能没有专门的软件工程师来帮助构建支持学习课程的软件，他们凭借越来越多的聊天机器人创作平台的出现，使设计和部署变得越来越简单。相信以下内容，可以帮助大家更好的设计聊天机器人来支持学员学习。设计支持学习的聊天机器人时需要考虑的因素1. 目的明确现实中的人工智能技术（例如聊天机器人）并无法像好莱坞中的描绘（以《钢铁侠》中的贾维斯和《太空漫游》中的哈尔为例）。电影中人工智能可以与人类进行深入的情感交流。然而现实中对人工智能技术（例如聊天机器人）进行编程，主要是使其能够执行简单的任务，例如发送提醒和回答常见问题。简单并不意味着没有价值。相反，设计能够向目标用户提供清晰的情境信息的聊天机器人可以为组织带来巨大的价值。为确保用户不会对与聊天机器人之间的互动抱有不切实际的期望，一定要明确地告知用户聊天机器人的用途。说明聊天机器人会做什么，不会做什么。例如，当你的聊天机器人首次向新学习者介绍自己时，它可能会说：“嗨！我是机器人教练，我将向你发送一些关于你最近完成的有关倾听技能的课程的提醒信息。未来30天内，每周我会发送两封电子邮件。你可以随时发送‘停止’选择不再接收我的邮件！”在上例中，聊天机器人阐明了专业知识的内容，与用户互动的频率，这次体验将持续的时间以及如何选择不再接收电子邮件的方法。通过确定明确的期望，你就可以消除用户的疑惑和顾虑。2. 确定用例如果有明确的商业案例，学习者认为邮件能够为他们带来真正的好处，那么最优秀的学习聊天机器人就能够带来最大的投资回报和用户参与度。评估聊天机器人是否是合适的学习工具时，思考：• 能够为组织解决什么问题？• 学习者能够通过体验获得什么好处？能够合理平衡公司和学习者获益的用例包括新员工入职、培训强化和绩效支持。新员工入职聊天机器人互动范例3. 选择正确的渠道当你想象聊天机器人时，你可能想到的是网页上弹出的小窗口，询问是否有任何问题需要回答，并引导用户跳转至可点击菜单以获取帮助。或者你可能想象 app 中的聊天机器人，例如协助填写保险理赔单。这些都不是唯一的学习聊天机器人的形式。学习型聊天机器人最适合短消息、Microsoft Teams 或 WhatsApp 等消息传递平台，因为通过这些平台可以进行主动且持续的对话。为了充分利用聊天机器人，应该将其配置在学习者目前正在使用的渠道中。也就是说，不需要下载新的 app，也不需要记住任何网址。相反，聊天机器人就在你与同事和朋友之间的用户聊天窗口旁。你甚至还可以为不同的受众选择不同的渠道，确保每个人都拥有最佳体验。例如，想要联系你的的移动销售团队，请在短消息平台上设置一个聊天机器人。然后，对于经常使用 Microsoft Teams 的管理者，就在该软件中设置聊天机器人。4. 消息传递个性化当今的聊天机器人技术已经过大量测试，能够确保妥善地在你需要的时候将消息传达给目标对象。然而，消息是否被正确接收则取决于其质量。消息内容的个性化程度越高，对学习者来说就越有意义和价值。个性化也可以非常简单，例如加入用户的名字，也可以相对复杂，例如根据学习者的绩效和偏好设计不同的适应性学习途径。可以通过两种方法采集信息以实现学习体验的个性化：根据用户提供的信息或从外部系统提取信息。基于用户的个性化可以包括诸如选择工作目标、首选主题或预先设置一天内接收提醒消息的时间。例如以下聊天机器人对话：5.推拉平衡推拉Push and Pull这一术语来源于营销领域。推动营销是指主动向消费者宣传产品，而拉动营销则是指将消费者吸引到他们想要选择的内容中。你可以对聊天机器人进行编程，将内容推送给学习者，例如提醒、督促、挑战和新的培训材料。你还可以对聊天机器人进行编程，以便在学习者需要帮助时（例如，为了获得绩效支持）可以立即访问。拉动场景正是利用了聊天机器人从不睡觉、不会累，而且永远在线的理念。绩效支持聊天机器人互动范例为了优化与聊天机器人共同度过的学习之旅，可以对聊天机器人进行设计，以同时利用推和拉策略。例如，在入职场景中，可将按照话术的机器人对话按照 70-30 的比例分别采取推动和拉动策略。推动部分可以包括定时发送消息，以提醒用户截止日期、签到问题以及有关组织的提示。脚本的拉动部分可能包括你希望新员工提出的问题（我在哪里可以找到有关我的健康福利的更多信息？我们的带薪休假政策是怎样的？）及其相关答案。6.制定内容管理策略聊天机器人的一个明显的好处是一旦设置完成，它就能自动运行，节省时间。但是，随着时间的流逝，内容可能会过时。内容管理策略包括一个定期重新审查脚本的计划，或者可以将聊天机器人连接到原始内容存储库，这样就只需要更新一次信息。在早期，请确保确定哪种方法最适合你的组织，并将更新过程纳入正常的工作流程中。由于相比传统的软件，聊天机器人的测试和迭代都更加容易，因此一旦启动，你就能够获得有价值的数据和用户反馈，以供将来用户使用。功能设计元素，例如反馈问题、A/B 测试和通知都有助于促进迭代。你可以在整个设计中使用其中的每一个元素，指明如何以及何时改进聊天机器人脚本。计划在第一次发布后至少迭代一次脚本。7.事先确定成功指标在开始设计脚本之前确定将用来衡量成功的指标。问自己还有关键利益相关者，怎样才算成功。通过事先选择指标，你就能够更好地确定聊天机器人应该说的内容的优先顺序，并完善这些内容。因为聊天机器人脚本编写的一个主要挑战就是想要让聊天机器人脚本涵盖所有内容，因此事先选择指标就有很帮助。如果你事先知道你将如何衡量成功，则可以确定哪些功能和内容最重要，并且应该成为用户体验的重点。事先就成功指标达成协议还可以令你成为在组织内实施聊天机器人的更有效的倡导者。每个员工都需要一个内部倡导者，你的虚拟团队成员（聊天机器人）也不例外。因为对你组织来说，聊天机器人可能是一项相对较新的创新，因此最坏的情况就是会存在怀疑论者，而利益相关者不发表反对意见就已经是最好的情况。因此，始终让人们可以看到聊天机器人如何做出贡献。本文内容摘选自TD杂志最新刊Chatbots, You Had me at Hello. 作者Vince Han。全文仅供ATD学习账号成员查看。欢迎购买年度个人学习账号（900元/年），学习更多内容。

Chatbot，“初见倾心”

January 2021 - TD Magazine

Chatbots can keep learners engaged and reinforce training.

You Had Me at Hello

今年1月，OpenAI的对话平台ChatGPT-3成为互联网史上热度上升最快的应用。这只是过去几周来，人们对GPT和其他聊天机器人的兴趣愈发浓厚的例证之一。作为最早看到会话式人工智能（AI）的潜力并为其设计应用程序的学习专业人士之一，我很乐见人们如此关注这项引人入胜的技术。我在《人才开发中的AI》（AI in Talent Development）这本书中，概述了几种将聊天机器人融入强大的学习体验的方式，并以OpenAI的ChatGPT系列的第三个版本，即ChatGPT-3，演示了其中的大部分用例。当ATD邀请我撰写一篇以此为主题文章时，我发现自己几乎词穷了，想不到如何为这个热搜话题添加更多新意。所以，我决定做自己最擅长的事情：传达核心信息，并指明实际的应用方向。让我们先了解一下它的基本概念。ChatGPT-3的工作原理ChatGPT以自回归模型运行，它其实没有听起来那么复杂。Transformer是一种程序、模型或算法，可以将代码转换为其他东西——在本例中就是指人类语言。“自回归”意味着该软件会基于先前的事件，预测未来最有可能发生的事件，这也正是人类大脑的运行模式。它允许ChatGPT和类似的应用程序模拟两个人之间的对话。要制作这类应用程序，就必须掌握两项关键的技能：1. 自然语言编程自然语言处理（NLP）允许机器人通过人类语音或书写来识别输入，无需使用菜单或代码。这种功能还可以支持机器人通过文本进行回复。在NLP问世之前，大多数聊天机器人都是由菜单驱动的，机器人通过预先编写的选项数据库，选择最适合的选项，并据此作出回应。对话式AI代表了人类使用计算机的一种新方式。它是一个用户界面，与之前的UI（用户界面）相比，它需要的硬件和先验知识较少。注意！不要将“自然语言处理”与使用相同首字母缩略词的“神经语言编程”（Neural Linguistic Programming）相混淆，后者是一个被广泛否定的伪科学术语，也是一种无效的实践。2. 海量训练数据集与任何神经网络（包括你的神经网络）一样，AI模型必须使用一种叫做“深度学习”的技术进行有效训练。坦率地说，ChatGPT 令人震惊的强大功能并不是什么新鲜事。正如David Chalmers（澳大利亚哲学家和认知科学家）所言，不存在所谓的新技术。行业已经出现了许多具有类似能力的机器人，我在2020年就详细描述了与之相关的信息。但ChatGPT-3采用了创新的训练方式，并且它的效能同样令人耳目一新。首先，Open AI使用了最大的训练数据集，结合了一个包含大部分互联网的数据库、一个庞大的图书库和维基百科。然后，Open AI在内部完善该机器人之后，向世界免费推出了这款产品。在我们体验其功能的过程中，我们实际上也是在强化它的训练。我们不妨将它视为一项众包的培训工作。庞大的训练数据库规模生成了不负众望的结果——这些令人印象深刻的结果能够很好地模仿人类的表达，几乎难以分辨出是机器人所为。不过，这里仍然存在一个根本性的问题：聊天机器人的优劣取决于它的训练程度（就像我们人类一样）。如果你思考GPT培训的来源，就不难理解为何它有时候也会出错。我们如何使用ChatGPT ？因为该产品仍然有待升级，所以截止本文发布时，业内可能还会涌现关于ChatGPT 的新用例。以下是一些有助于节省时间的用法，作为一名学习专家，你可以借助它们提高工作效率和产能。撰写内容如果你让机器人写一段介绍某个主题的内容，它几乎能够立即应答，并提供言之有理且语法正确的文本。由于它的培训包含未经验证的来源，例如互联网和维基百科，所以你必须对其生成的某些内容进行修改和纠正。尽管如此，你仍然可以获得优秀的初稿，并在此基础上进行润色。文档摘要需要寻找某本书或某篇文章的摘要？将文本粘贴到ChatGPT的提示中，它就会返回一篇很好的摘要，其中还会提到绝大部分的文章要点。编写和诊断计算机代码因为ChatGPT是用于分析和生成人类语言输出的，所以它在计算机语言方面的表现也不逊色。程序员正在使用它来诊断代码。如果你要求它开发一项新功能，它甚至可以生成新的代码。翻译文档ChatGPT比其他翻译软件更有效，能够将网站或文章翻译成另一种语言。出题和阅卷你可以将内容输入GPT，并要求它生成带有答案的试卷。为它提供一个广泛的主题，比如世界历史，它仍然可以生成一份相关的试卷。将学生的考试答案输入GPT，GPT就会给他们打分，它甚至还能给论文评分！制作游戏我让GPT制作一款游戏，目的是向六年级的学生传授关于太阳系的知识，机器人返回了一款关于太空冒险的有趣游戏。提高写作水平我这篇文章当然不是由GPT代笔，但我可以要求GPT审查我的初稿并提出修改建议。执行研究我要求机器人以APA（美国心理学协会）的风格，提供一份关于在教育领域使用聊天机器人的最新同行评审研究列表，并收到了一份格式正确的列表。ChatGPT的问世为何令某些人感到担忧？虽然大多数人对GPT给予了积极的评价，但有一些教育者对这款新工具的功能深感担忧。模仿人类写作教师们对GPT的写作能力感到担忧，因为它能够像人类一样，写出言之有理的文本，并根据已知作者的文风进行写作。那么，学生就有可能要求机器人模拟他们之前的写作风格写一篇论文，而这就很难判断该GPT论文究竟算不算学生的原创。但这种模仿能力仍然不是什么鲜事。AI早在数年前就已经模仿伟大艺术家的风格进行了艺术创作。ChatGPT只不过是将相同的预测方法应用于语言领域罢了。ChatGPT也会犯错另一个令人担忧的问题是，机器人能够以极具说服力的方式提供虚假信息，以至于某些用户忘了亲自验证其回答的正确性。例如，我让机器人计算五个数字的平均值，结果很有趣。我收到了关于“均值”的统计学定义以及如何计算的详细说明，但我并没有提出这个要求。不过，这个机器人将这五个数字相加的结果算错了。公平地说，ChatGPT当然不是为解决数学问题而构建的。但它没有告诉我或者报告该结果可能有误。与任何热心的助手一样，它已经在尽力为人们提供答案。如果我是一个依赖GPT的学生，我就会被误导，并因此而犯错。大多数此类错误与让机器人聊天体验突飞猛进的原理同出一源——它的主要信息来源是互联网和维基百科。Vince Han是构建学习聊天机器人的先驱之一，我请他总结了GPT新功能带来的挑战，他的回答如下：“ChatGPT和AI质量提高的一个影响就在于，它改变了用户需要的技能。例如，大多数用户已经掌握了如何搜索的技能，即知道如何创造关键字以及如何筛选搜索结果。现在，人们将需要辨别能力，或者学习如何通过一个聊天机器人识别真相和观点，因为聊天机器人只是提供了一个看似确定的答案。因此，从绩效的角度来看，人们现在的提问的确可以立即得到答复，但该答复的准确性和相关性将取决于他们自己的判断。”作为教育工作者，我们现在必须正视一个人们热议了数十年的问题——大多数领域、人口统计和工作小组的批判性思维正急剧下降。在我看来，ChatGPT进一步扩大了这种差距，也让我们看到了弥合这种差距的紧迫性。结论微软拥有独家使用该模型的许可证，我们实际上都在为微软工作，为其发掘可能为微软的AI产品找到出路的绝佳用例。目前，这个平台推出了免费版和专业版。我们可以通过这项工具，开始了解AI的前景以及我们未来的定位。正如机器学习专家Santiago Pino所言：“AI不会取代你，能够取代你的是使用AI的人。”

ChatGPT 对人才发展行业有何影响？

In January, OpenAI’s conversational platform, ChatGPT-3, became the fastest-growing app in internet history. This is just one indication of growing curiosity about GPT (generative pre-trained transformer) and other chatbots in the past few weeks. As one of the first learning professionals to see the potential of conversational AI and one of the few designing the applications for years, I’m excited to see the recent attention paid to this engaging technology.In my book, AI in Talent Development, I outline several ways chatbots can be used as part of a robust learning experience, and ChatGPT-3, the third version in OpenAI’s ChatGPT series, demonstrates most of them. When ATD asked me to write a post on this subject, I struggled to find the value I could add to the exploding conversation. I decided to do what people say I do best: Give you the bottom line and point the way towards practical applications. Let’s start with the basics.How ChatGPT-3 WorksChatGPT operates on an autoregressive model. It’s not as complicated as it may sound. A transformer is a program, model, or algorithm that converts code into something else—in this case, human language. The autoregressive part means the software uses prior events to predict the most likely future events, which is what your brain does all the time. It allows ChatGPT and similar applications to simulate a conversation between two humans. Learn the two keys to making an application like this work:1. Natural Language ProgrammingNatural Language Processing (NLP) allows the bot to recognize input through human speech or writing, eliminating the need for menus or code to engage with it. This same ability enables the bot to reply through text. Before the development of NLP, most chatbots were menu-driven, and the bot responded by selecting the best fit from a database of pre-written options. Conversational AI represents a new way for humans to use computers; it’s a user interface requiring less hardware and prior knowledge than any previous UIs.!CAUTION! Don’t confuse Natural Language Processing with the pseudo-science term using the same acronym. Neural Linguistic Programming is widely disproved and an invalid practice.2. Massive Training DatasetsLike any neural network (including your own), AI models must be effectively trained using a technique called deep learning. Quite frankly, the stunning capabilities of ChatGPT-3 are nothing new. As David Chalmers points out, there is no new technology here. There are many bots with similar abilities, and I described them in detail in 2020. How this bot is trained and how well this approach works is innovative. First, Open AI used the largest training dataset, combining a database containing much of the internet, a huge library of books, and Wikipedia.Then, after refining the bot in-house, Open AI released it free to the world. As we all started to play with it, we were training it further for them. Think of it as a crowd-sourced training effort. The sheer size of the training database resulted in what everyone always expected: Impressive results that can mimic human expression so well it is hard to tell a machine generated it. However, there continues to be an underlying problem: Chatbots are only as good as their training (like us humans). If you consider the source of GPT’s training, you can see why it sometimes makes mistakes.What Can We Do With ChatGPT-3?Because this product is still being explored, there may be new use cases by the time this post publishes. Below are a few applications that can save time and make your work as a learning professional more efficient and productive.Write ContentIf you ask the bot to write a paragraph introducing a subject, it comes back almost instantly with plausible, grammatically correct text. Because its training included unvalidated sources like the internet and Wikipedia, some of the content it produces must be corrected. Still, you can get excellent first drafts and refine your content from there.Summarize DocumentsLooking for a “Reader’s Digest” version of a book or article? Paste the text into a prompt to ChatGPT, and it returns a nice summary that should hit most of the highlights.Write and Troubleshoot Computer CodeBecause ChatGPT was built to analyze and produce output in human language, it is also great at computer languages. Programmers are using it to troubleshoot code. It can even generate new code if you ask it to develop a new capability.Translate DocumentsFar more effective than other translators out there, ChatGPT can take a website or article and translate it into another language.Write and Grade QuizzesYou can feed your content into GPT and ask it to generate a quiz with answers. Or, give it a general subject, like world history, and it will do the same. Feed your student’s quiz responses into GPT, and it will grade them (even essays!).Create GamesI asked GPT to create a game teaching sixth-grade students about the solar system, and the bot came back with a fun adventure through space.Improve Your WritingI didn’t do this for this article, but I could have asked GPT to review my draft and suggest improvements.Conduct ResearchI asked the bot to provide a list of current peer-reviewed studies on chatbot use in education using APA style and received a properly formatted list.Why Does ChatGPT Have Some People Worried?While most of the comments about GPT are positive, you will find a few educators deeply concerned about this new tool’s capabilities.Mimicking Human WritingTeachers are worried about GPT’s ability to write plausible text reading like a human created it and write in a known author’s style. A student could ask the bot to write a paper in the same style as one of their previous works, and it would be hard to tell the GPT paper wasn’t an original student effort. But, again, this mimicking ability is nothing new. AI has generated art in the style of great artists for years. ChatGPT applies the same predictive approach to language.ChatGPT Makes MistakesAnother concern is the bot’s ability to present false information so convincingly that some users forget to validate the response with research. For example, I asked the bot to calculate the mean of five numbers, and the result was interesting. I received a detailed explanation of the statistical definition of “mean” and how to calculate it, which I did NOT request. However, the bot added the five numbers incorrectly.To be fair, ChatGPT wasn’t built to perform math problems. But it didn’t tell me that or report the result might be incorrect. Like any willing assistant, it did its best to provide an answer. If I had been a student relying on GPT’s help, I would have been wrong. The reason for most of these mistakes is the same reason that it is such a stunning improvement in the chatbot experience: It’s primary sources of information are the internet and Wikipedia.I asked Vince Han, one of the pioneers in building chatbots for learning, to sum up the challenges presented by GPT’s new capabilities, and here’s his response:“One of the effects of ChatGPT & increased quality in Artificial Intelligence in general is shifting the skills users need. For example, most users have developed a skill for how to search which is knowing how to craft keywords and how to sift through search results. Now, people will need the skill of discernment, or learning how to decipher truth and views from a chatbot that provides what reads like a definitive answer. So, from a performance support perspective, yes, one can now get an immediate response to a question but how accurate and relevant that response is will be up to them to determine. “As educators, we must now confront a problem that has been boiling for decades—the decline in critical thinking across most domains, demographics, and workgroups. In my opinion, ChatGPT makes this gap more visible and increases the urgency for closing it quickly.Closing ThoughtsBecause Microsoft owns the license for the exclusive use of the model, we’ve all been working for them, revealing exciting use cases that may find their way into AI-enabled Microsoft products. For now, the platform consists of both free and professional versions. It’s an excellent way to begin learning about the promise of AI and our place alongside it. As machine learning expert Santiago Pino says, “AI will not replace you. A person using AI will.”Where Can You Learn More?There are so many articles about this subject right now that I’ve limited myself to a few favorites:Under the Hood of Chat GPT With Myra RoldanHow ChatGPT3 Impacts the Future of L&D in an AI WorldChatGPT Has Educators Scrambling to Keep UpWhat Is ChatGPT and How to Teach With ItWhat Is ChatGPT and Why Does It Matter?What Is ChatGPT? Is It the Beginning of the End?Ger’s Learning Notes #51 - ChatGPTFor more information and resources, Check out ATD's AI Resource page.

ChatGPT-3 Has Everyone Talking

Mary Vonnegut (mvonnegut@primegenesis.com) is formerly the president of Gumps Catalog, senior vice president of marketing of Hanover Direct, vice president of marketing of Ross-Simons, and chief operating officer of Passport Internationale. She is co-author of Onboarding: How to Get Your New Employees Up to Speed in Half the Time.

Mary Vonnegut

Vern Hanzlik is senior vice president and general manager of Qumu. He has spent more than 20 years building and growing enterprise software and service companies. Previously, he served as president of EMEA at TEAM Informatics, as well as a member of its board. He also co-founded Stellent, an enterprise software company acquired by Oracle.

Search results for 'Vince Han'

Categories

Refine By

Search results for 'Vince Han'

Categories

Refine By

Role

Topic